Home

tomas sedlacek a real-time positioning system of manufacturing

1. sseoeeesessssssoeeersssssssseeersssssssseerees 31 Table 4 1 List of scenarios for the production line ssssoeeessssssssseerrsssssssseeresssssssseeeees 43 Table 4 2 Characteristics Ol SCMSOLS rea ve anareaas neti ecd untae needs 47 Table 4 gt List of importance Were nts ciinei a a 60 Tables Event Ces Chi Pll Ol reenen a S 66 Jable 32 Experiment des ipt Oisen a EA TE 69 VII LIST OF ABBREVIATIONS 2D Two Dimensional 3D Three Dimensional 6LOWPAN Internet Protocol version 6 over Low power Wireless Per sonal Area Network AGV Automatic Guided Vehicle AJAX Asynchronous JavaScript and XML AOA Angle Of Arrival APIS Advanced Pallet Information System CAD Computer Aided Design CCID Community Colleges for International Development CVG Coriolis Vibratory Gyroscope DPWS Device Profile for Web Services eSONIA Diagnosis and Control Towards the Asset aware and Self recovery Factory EMV Expected Monetary Value EU Expected Utility FAST Factory Automation System Technology FIFO First In First Out GPS Global Positioning System GUI Graphical User Interface HMI Human Machine Interface HTML HyperText Markup Language HTTP HyperText Transport Protocol ID Identification number IEEE Institute of Electrical and Electronics Engineers IP Internet Protocol IPv4 Internet Protocol version 4 IPv6 Internet Protocol version 6 ISO International Organization for Standardization IT Information Technology JSON JavaScrip
2. c cc ceeseeeseeeeeeeeeeeeeeees 71 S91 Mone Carlo Localza ON aaaea S 71 5 3 2 Simple Decision making algorithm cccccceseseeeeceeeeeeeeeeeeees 12 5 4 Graphical User Intertace OUD aar ered ede eaiverncd datas eaareead ones 73 CONIC IUSION aii testing ttasivaat T tiantvn eid etait eal ata Oana eed 75 6 1 Pane WO koae a EA 76 Ree 8 Cl che a T E T T A T1 APP nd AA User Manual For GUT regre a E S 8l LIST OF FIGURES Figure 2 1 Schema of an Intelligent Agent 20 0 e 7 Figure 2 2 Working principle of probabilistic localization ccceeecccceeeceeeeeeeeeeeeees 17 Figure 2 3 Example of Monte Carlo Localization a Initialization b Ambiguity due to symmetry c successful loCaliZation ccccccccccessseeeeeeceeeeceaaaeeseeeeeeeeeeaaas 18 Figure 2 4 Global MST market by application in 2011 ee eeeeeeeeeeeeeeeeeeeeeeees 20 Figure 2 5 Detail of accelerometer structure 20um thick single crystal comb structure used for Capacitance Measurement scerna E a 21 Figure 2 6 Traditional mechanical gyroscope ccccccccesssseesseececeeeeaseesseceeeeeeeaaeeeeees 22 Figure 2 7 Working principle of Coriolis vibratory SyrOSCOPEC ccececcecceeeeeeeeeeeeeees 23 Figure 2 8 Working principle of optical gyroscope with set Of MIITOTS ccceeeeee 24 Figure 2 9 The general architecture Of a sensor system ccsssseeeeeceeeeeeeeeeeeeeaaeees 24 Figure 2 10 RFID tags used for a animals tag
3. Hyunsook Kim Eunhwa Kim Kijun Han An Energy Efficient Tracking Method Springer Berlin 2006 Iozan L I Rusu C Collin J Takala J A study of the external factors that affect the measurement data of a MEMS gyroscope sensor Towards an inertial navigation system 2010 9th International Symposium on Electronics and Telecommunications ISETC Eagle S Jones Stefano Soatto Visual Inertial Navigation Mapping and Localization A Scalable Real Time Causal Approach 2010 Internation Journal of Robotic Research Juan C Alvarez Antonio M L pez Rafael C Gonzalez Diego Alvarez Pedestrian Dead Reckoning With Waist Worn Inertial Sensors 2012 Instrumentation and Measurement Technology Conference Mouser Electronics Online http M mouser com ProductDetail Atmel ATZB 24 A2 APPENDIX A USER MANUAL FOR GUI Graphical User Interface GUI provides operator a functional interface for not only pallet monitoring but also for a control and dynamic interaction with the process of pal let localization There are available four different GUI environments or web pages for operator disposal each serving for specific purpose welcome page design page online monitoring page and offline monitoring page The GUI is called APIS and stands for Advanced Pallet Information System Welcome Page Welcome Page is a home page of developed GUI serving at the same time as the main finger post referring to all constituent subpages too Its appearanc
4. TAMPEREEN TEKNILLINEN YLIOPISTO TAMPERE UNIVERSITY OF TECHNOLOGY TOMAS SEDLACEK A REAL TIME POSITIONING SYSTEM OF MANUFACTURING CARRIERS DEPLOYING WIRELESS MEMS ACCELEROMETERS AND GYROSCOPES Master of Science thesis Examiner Professor Jose Luis Martinez Lastra Examiner and topic approved in the Automa tion Mechanical and Materials Engineering Faculty Council Meeting on 6 June 2012 ABSTRACT TAMPERE UNIVERSITY OF TECHNOLOGY Master of Science Degree Programme in Machine Automation SEDLACEK TOMAS A real time positioning system of manufacturing carriers deploying wireless MEMS accelerometers and gyroscopes Master of Science Thesis 80 pages 11 Appendix pages September 2012 Major subject Factory Automation Examiner Professor Jose L Martinez Lastra Keywords real time positioning localization pallet wireless sensor MEMS accel erometer gyroscope b5LOWPAN artificial intelligence Modern manufacturing systems face ever increasing pressure to maximize efficien cy of production processes minimize downtime due to unexpected deviations from normal operation and maintain agility in dynamic market conditions Detailed real time asset tracking is essential for achieving these goals Pallets are widely used for transporting raw materials intermediate products and fi nal products in automated assembly and manufacturing lines A sophisticated pallet monitoring system can provide possibilities for optimizing pallet r
5. matics and engineering to fulfil some goal or to optimize a process Throughout the rest of the thesis a focus will be given especially to the rational ap proach of AI since it corresponds to a subject of this thesis pallet localization repre sents a goal which a certain AI algorithm must fulfil in order to increase the possible output accuracy In following chapter a rational or intelligent agent is going to be in troduced as the central of intelligence in rational approach 2 1 2 Intelligent Agents According to 3 an agent is anything that can be viewed as perceiving its environ ment through sensors and acting upon that environment through actuators The simplest example is a human agent The human agent has eyes ears or tongue as sensors to per ceive the environment and hands as actuators to interact with it More abstract illustra tion can be shown in case of software agent in virtual environment The software agent perceives data from real environment through a set of keyboard inputs files or web streams and acts on the environment by displaying screen writing data or controlling some real processes In AI terminology a term percept refers to perceptual input and percept sequence to complete history of everything what the agent perceived If the agent is inserted to its environment it generates a sequence of actions based on percepts it received to change its state in the environment as can be seen in the Fig ure 2 1 To d
6. wards and finds out that just one block backwards was a block with input event equal to LEFT and at the same time this LEFT event was not for that block and for that agent registered yet Calculated penalty for this block is V2 1 1 9 x 1 2 2 28 which is lower than for the block in forward direction of the pallet movement and thus with high probability the examined event is just a late event Nevertheless in result the agent will not move back to that left turning it is too late pallet is actually on the right turning already but the event will be supple mented to agent event history and the agent stays at a position where it is Figure 4 13 1b If a new event LEFT theoretically comes again the algorithm would not take the block into the consideration any more since the LEFT event has been for the block already applied Scenario 2 The property of backward searching might be very useful also in scenario when agent enters junction where it can move either straight artifi cial event TIME generated or right RIGHT event If the RIGHT event will be a bit delayed the artificial time interval will elapse and TIME event will be gen erated causing the agent to move straight Figure 4 13 2a However if the RIGHT event comes the algorithm at first looks forward in order to find a block that is in this case 7 blocks distant V1 7 1 2 x 1 8 4 and then looks backwards where two possible blocks with input event of RIGHT are found
7. 16 17 REFERENCES Dhamee Yousuf Adam Smith and the Division of Labor Online 1996 http www victorianweb org economics division html Assembly Line Online Wikipedia 2012 Citation 12 April 2012 http en wikipedia org title Assembly_line Stuart Russel Peter Norvig Artificial Intelligence A modern approach 2003 Decision theory Online LessWrongWiki 2012 http wiki lesswrong com wiki Decision_theory Dung Truong My Decision Theory Online 2005 http www fit hcmus edu vn tmdung courses SoftApp Ch5_TMD_2005 pdf Hansson Sven Ove Decision Theory A Brief Introduction 1995 Decision Making Online August 2010 http decision making role blogspot fi 20 10 08 decisions under certainty risk and html Jones James Decision Theory Online Richland University 2003 http people richland edu james summer03 m160 decision html Game Theory Online Investopedia http www investopedia com terms g gametheory asp axzz24XmZTdmK Thrun Sebastian Particle Filters in Robotics 2002 Uncertainty in AI Jens Steffen Gutmann Wolfram Burgard Dieter Fox Kurt Konolige An Experimental Comparison of LocalizationMethods Online 2002 http www cs cmu edu biorobotics motion www papers sbp_papers integrated 3 guttman_compare pdf Werger Barry Brian Cooperation without deliberation A minimal behavior based approach to multi robot teams Online 1998 http www cs uml edu holly 91 549 readings werger pdf
8. While the block 3 is not examined at all since the RIGHT event was already ap plied a while ago to this block the second block has penalty equals just toV2 3 1 9 x 1 5 7 which is less than for block 1 and therefore the agent will jump to the second place and agent position history is upgraded according to the new conditions as can be seen in Figure 4 13 2b IMPLEMENTATION 62 Last evnets RIGHT S TOP GO Last events STOP GO RIGHT RFID TIME TIME Current event LEFT Current event RIGHT blocks with input event LEFT blocks with input event RIGHT V1 8 4_ 7 v3 x V2 5 7 f 4 44 oo mt a A 7 pallet rerouting 2b Figure 4 13 GUI representation of algorithm behaviour with calculated penalties for two different scenarios I delayed LEFT event a before the event is received b after as a result the event is just supplemented 2 de layed RIGHT event on RIGHT Y JUNCTION a earlier than the event is re ceived a TIME event is generated and agent continues straight b after as a result the agent path is rerouted 3 Asynchronous tracking phase Calibration Asynchronous tracking is a phase of agent life cycle when its position does not match with position of real pallet and thus they are not in mutual synchroni zation In fact there is only one way how to recognize pallet dislocation using RFID event Hence whenever a RFID event comes and the workstation from which the event
9. quires rotating element influencing negatively overall size of the sensor A work re lated to this method can be found in 60 e Distance related measurements Distance related measurements are primarily based on measurement of propaga tion time between transmitter Sensor and receiver base station or sensor and can be divided according to a way of propagation into three main groups 61 o One way propagation time measurements it measures the difference between the sending time of a signal at the transmitter and the receiving time of the signal at the receiver Biggest drawback of this method is a must for absolutely synchronized clocks in both devices which makes this method less attractive o Roundtrip propagation time measurements it measures the difference between the time when a signal is sent by a sensor and the time when the signal returned by a second sensor is received at the original sensor again Significant advantage is that there is no need for precise synchro nization since the same clocks are used for measurement however pro cessing time on second sensor must be taken into consideration and thus this delay must be subtracted Processing time could be known in ad vance by tests or sent side by side with the localization signal o Time difference of arrival measurements TDOA this method calcu lates an unknown position of transmitter as a time difference between the transmitter and set of receivers wi
10. the WS1 is responsible for raw material loading a sheet of paper and final product drawn cell phone unloading and WS7 is a buffer that serves as a temporal storage for pallets In a time this research was done and this thesis was written the pallet buffer was still under maintenance and thus for all the time out of order but for completeness it 1s going to be mentioned here as well since it was and soon will be again a part of the production line Each workstation apart from the buffer is equipped by the same 4 axis SCARA robot SONY SRX 611 while the buffer has just a pneumatic manipulator for lifting and storing a pallet from a conveyor to prepared shelfs and vice versa All workstation are further equipped by conveyor system pen feeders safety system and RFID NFC IMPLEMENTATION 42 readers Their proper function and communication with outer world is governed exclusively by smart remote terminal unit RTU called S1000 S1000 is a Web Service based controller that is manufactured by Inico Technoloigies company There are several S1000 deployed per workcell and each of them is responsible for control of different segment of workcell robot conveyor system or pen feeders Figure 4 1 FASTory line with labelled workstations 4 2 Production operations The production of drawings starts in workstation 1 where a robot equipped by a special gripper with suction cups inserts an empty sheet of paper to available pallet and covers it with
11. varying in shape in three different colours what gives 9 combinations for each element and 9 3 729 combinations in total for the final drawing Table 4 1 List of scenarios for the production line All workstations can draw all types of elements in any colour Each workstation can draw only certain element in any colour Each workstation can draw all types of elements but only in one certain colour Each workstation can draw all types of elements in any colour but only one element per pallet 4 3 Pallet A pallet is a structure of different sizes and shapes designed for product transporta tion Usually the term pallet refers to a standardized and universal wooden structure used in industry for storage or transportation of heavy and over sized material out of the factory However in production and automation industry the pallet known also as IMPLEMENTATION 44 manufacturing carrier or industrial pallet refers to smaller structures that transport products within a scope of production line Their movement is highly automated and usually provided by complex conveyor system The conveyor system can use for pallet movement either running belts spinning rollers or moving chains and others Pallet de sign is tightly coupled with the product they are supposed to hold in vast majority they are made to order and therefore a bit expensive In this research there are used brand new pallets designed by lvaro Res a Rey which were deployed
12. Criterion realism table for businessman example States gt It rains It does not WA a 0 8 Alternatives J rain As can be seen in Table 2 3 the highest average weight in businessman example is for his second alternative and thus a realistic businessman would let the um brella at home B Sequential Decisions Sequential decisions are decisions where the agent s utility 1s not known during the moment of decision as in simple decisions but it depends on sequence of decisions Therefore techniques such as planning and searching are applied to calculate whole sequence of actions that will lead to maximized utility Very nice example of se quential decision making is chess where the overall utility winning does not de pend only on one particular figure movement but complete sequence of them Three major sequential decisions are commonly used Markov Decision Process Partial ly Observable Decision Process and Game theory Due to limited space of this the sis only a very brief description will be given to each of them 3 Markov Decision Process MDP MDP is applicable to fully observable envi ronments and is based on Markov Property which says that agent s action which leads to another state depends only on the state the agent is right now and not on the preceding states MDP is characterized by a transition model of the environ ment where the initial state is known and all possible transition from each state are kno
13. Eve RIGHT RIGHT Forward 12 30 Searching size Searching size eves LEFT LEFU Backward 1 9 1 9 Forward 20 80 Others Coeficients ee i 20 20 Move time 10 10 Searching size Searching size pape ee p Others Others Submit Submit Submit Submit Figure A 8 Example of Variables panel Display panel that is shown in Figure A 9 is used for visualization control in Moni toring area With it help a user can adjust arbitrary display properties It consists of four sliders two list boxes and one label and for better lucidity their list and description is going to be again presented in Table A 5 Table A 5 Display panel Adjusting element Description of function Time Eff Switch between time mode and efficiency mode Enabling or disabling visualization of a map of blocks Enabling or disabling visualization of background image of pro BGND ee duction line Pallet number selection in case of one pallet in system only its number is listed there if more pallets are in system then besides their numbers also an option all is added to list box in order to Pallet number ee track all pallets at the same time however for this scenario time or efficiency mode is disabled and pallets are displayed in random colours Production line selection names of all maps stored in system are listed there by picking a map for the list the server map as well as Line maps the map in monitoring area is automatically re
14. Problem is that 6LoWPAN is designed for quick exchange of low amount of data so in case of bigger amount of data the message must be divided into more data packets which are sent separately In reality it means to send for the same amount of data instead of one data packet four data packets which represents 400 increase in communication traffic With TCP usage another fact arises retransmission of lost packets which again unfortunately increase a chance of network overloading IMPLEMENTATION 47 Table 4 2 Characteristics of sensors Supply voltage 1 71 V to 3 6 V 2 4 to 3 6 V Free fall detection Integrated low and high Sleep mode pass filters Other FIFO Sleep mode Temperature sensor FIFO Temperature sensor 4 4 2 RFID NFC tag detection To efficiently control the drawing process the production line needs to have a capa bility for pallet recognition Therefore each pallet 1s equipped by a passive NFC tag which stores unique identification number of the pallet The tag is placed in the bottom of middle part of the pallet as might be seen in Figure 4 3c and while the pallet moves its identification numbers is read by NFC readers sensors distributed across the pro duction line The NFC reader shown in Figure 4 6 is connected by serial link to S1000 a smart RTU that broadcasts data of the reader to all subscribers via Web Services The reader is a commercially available device characterized by fast reaction low price and sim
15. a change in capacitance between set of teeth on the vibrating structure and set of teeth on support In comparison to mechanical gyroscope the Coriolis vibratory gyroscope due to its working principle cannot provide information about absolute orientation but just a change in orientation On market there are currently available one two or even three dimensional gyro scopes 18 23 INNER FRAME SPRINGS CORIOLIS SENSE FINGERS Figure 2 7 Working principle of Coriolis vibratory gyroscope 25 Optical gyroscopes Working principle of optical gyroscopes is based on fact that propagation of light beam varies in respect to a movement of conductor which leads the beam In practice it means that if two laser beams are sent to a long optical fi bre conductor at the same time from both sides while the optical fibre is rotat ing there is going to be measured a phase shift between both beams The phase shift is equivalent to the rotation of the optical fibre This phenomenon is known as Sagnac effect and it does not mean anything else than a fact that the beam propagating against rotation have a shorter path then the second beam Sometimes there is used a set of mirrors for beam propagation instead of opti cal fibre as can be seen in Figure 2 8 25 26 27 THEORETICAL BACKGROUND 24 Light source Half silvered mirror Detector Figure 2 8 Working principle of optical gyroscope with set of mirrors 25 2 2 4 Senso
16. a metal frame to prevent paper slide Then a pallet is released and it moves via conveyor system to particular workstations The conveyor system is circular and thus allows pallet to circle around the line as long as its drawing is fully complete As soon as the pallet reaches entrance of particular cell the system must decide whether the pallet will be in the cell processed something will be drawn or bypassed to other workstation There are several scenarios listed in Table 4 1 according which the system decides what to do Generally if the robot of certain workstation is idle and capable of required operation the pallet 1s accepted to enter the workstation otherwise it is by passed to next workstation In Figure 4 2 are shown both possible directions of the pal let movement within one workstation IMPLEMENTATION 43 Bypass conveyor Main conveyor E Figure 4 2 Workstation with main conveyor and bypass conveyor marked As soon as the pallet undergoes all required operation steps and drawing is ready it moves again to workstation 1 where the robot takes the finished drawing and put an empty paper sheet back in its stead If there is no need for another drawing the empty pallet goes to buffer where is taken away from the conveyor system The final drawing which should represent a final product consists of three elements frame screen and keyboard For each of those elements there are three different ver sions
17. algorithm as its program its brain that is responsible for correct decisions In this chapter the idea will be more extend ed and actual implementation will be presented Each agent is characterized by its own life cycle consisting of three different phases initialization synchronous tracking and asynchronous tracking Since each of them represents a different life period of the agent it was necessary to develop for each of them also a special decision making algorithm In following paragraphs both phase and appropriate algorithm is going to be described IMPLEMENTATION 58 1 Agent initialization phase Agent initialization as name prompts is a first phase of agent life cycle when a new agent is introduced into the system In reality it corresponds to a time when a system receives an event log from new unknown pallet which was inserted to the conveyor system This phase lasts as long as the agent is success fully localized within its environmental model After receiving the first event log the agent position is mostly not yet known since the single event log contains only simple information about one pallet movement Nevertheless the system at least registers the new agent into own maps and prepares enough resource capac ities for tracking and storing agent s incoming activities As the pallet moves new event logs are received and stored After each of them the initialization al gorithm is applied to search and find a pred
18. available there is a possibility to hide the image to do not confuse operator APPENDIX A USER MANUAL FOR GUI 85 Pallets are represented in the monitoring area by a line of colourful circles where a head symbolizes current pallet position and following circles are creating a tail symbol izing a historical pallet track The diameter of circles is getting smaller from the head towards the track tail The length of the tail can be freely chosen from 1 only head up to 20 by intuitive slider adjustment located in Control area eSONiA i Information Efficiency 45 0 Acc ID 12 Workcell WS10 Laps 1 Position 3800 300 Lap complete 00 04 00 Lap Time 00 00 45 Total time 00 04 47 Control Variables Time Eff Path BGND rer ra eee Pallet number 12 Line maps FASTory v T Figure A 4 Example of real time monitoring page There are available two visualization modes how a pallet can be displayed in Moni toring area The basic mode is a time mode where a time spent by a pallet in each posi tion is represented by a smooth change of circle colour Hence when a pallet enters a block the circle has green colour but while the time runs and pallet does not move the colour is changing into orange and later to red colour The time mode with long tracking path is a perfect tool for analysis of pallet dynamics because an operator can easily dis tinguish where each pallet stopped get delayed
19. backward search all blocks for which the certain event was already once ap plied and registered for the particular agent are of course expelled Each of those found blocks represents a decision alternative to which a proper utility must be assigned That is a work for utility estimation mechanism However before the working principle will be described in next paragraph it is necessary to explain how the mechanism understands a term of utility Usually the higher utility a decision alternative has the better outcome for the agent it represents On contrary the developed estimation mechanism calculates utility in opposite way the lower the better Therefore the term utility will be called further in the explanation by a term penalty that describes better the actual meaning A penalty V for each block x from set of alternatives is calculated according to Equation 4 1 where d is a distance in number of blocks between a block at which an agent is currently located and the examined block x coefficient w is an importance weight of each event according to values listed in the Table 4 3 The lower number a certain event has the more significant and more reliable event it is and vice versa Furthermore the equation contains a set of adjustable coeffi cients too The orientation coefficient Cg adjusts a direction profitability because it assigns a different value for forward movement by default is equal to 1 2 and for backward movement by de
20. device router Normal position On side above con Y A Dynam c S open pallet veyor system Normal position On side above con Y Static ae open pallet veyor system Normal position On side above con Y Dynamic 2 closed pallet veyor system Normal position On side above con Stat aa atic closed pallet veyor system Normal position In the middle down Dynamic closed pallet the conveyor system Upside down In the middle down Dynamic closed pallet the conveyor system Upside down On side down the G Dynamic closed pallet conveyor system os Based on results shown in the Figure 5 5 it was proved that there is a significant difference between open red columns and close pallet non red columns which is caused by device encapsulation and subsequently by signal reflexions and absorptions However the choice between open and close pallet is in fact already given due to testbed properties since open pallet cannot hold wireless device at the same time with a paper for drawing operations and these measurements serves exclusively for compari sons The results also proved the theory that upper plate is impermeable for the signal and thus the signal propagates mostly downwards efficiency increased in experiments where edge router was placed below the conveyor line blue columns Only minor change is then in position of edge router down the conveyor line Position in the middle gives slightly better results than
21. in Figure 5 4 To increase signal receiving efficiency as much as possible there was done set of experiments to determine the best mutual position between wireless devices and edge router A benchmark of efficiency was represented by amount of received messages by the edge router per minute for each wireless device For better credibility of the experi ment there were used two wireless devices simultaneously and each of them transmitted one messages per second Thus in ideal case the edge router should receive 60 messag es per minute for each wireless device DISCUSSION ON RESULTS 68 metal plate device a metal bottom plate with holes b Figure 5 4 Signal propagation schema a normal position b upside down position In total there were tested seven different mutual positions between edge router and devices and each experiment lasted for exactly twenty minutes The overall processed results are shown in Figure 5 5 where the vertical axis stands for amount of received messages while the horizontal axis displays letters that represents each experiment of which description can be found in Table 5 2 Each experiment has two values for the two pallets that were measured Figure 5 5 Efficiency of received messages in dependency of mutual position be tween edge router and devices DISCUSSION ON RESULTS 69 Table 5 2 Experiment description Position of wire Position of edge i o Situation Schematic icon less
22. in summer of 2012 The pallet is especially designed for transpor tation of a sheet of paper of size A6 along the conveyor system of production line de scribed in Chapter 4 1 The conveyor system uses for pallet movement a complex of running belts which carry away the pallet The pallet shown in Figure 4 3a consists of three parts Figure 4 3 a Pallet b embedded wireless device c RFID tag The lower part is made from a thick steel plate and serves as an interface between conveyor system and pallet It must be both durable because of the wear of belts and heavy enough to provide the pallet necessary stability The middle part is just an empty plastic box that serves as a room for various embedded wireless devices attached to the pallet from the bottom Finally the upper part is a detachable metal frame for keeping the sheet of paper in fixed position while pallet moves or a robot draws The right ori entation of the frame in respect to pallet is ensured by central pins and mounted by set of magnets The pallet has beside those three main listed parts also a rubber bumper in the front and a lock on its side The lock serves for stopping pallet in a fixed position on the conveyor by extensible pneumatic piston while the conveyor is still running under the pallet 4 4 Devices In order to detect pallet position while it moves throughout the conveyor system certain sensors must be naturally installed to provide desired information Since the
23. in this research belts are used for pallet transportation Naturally pallets are not tightly connected to belts and thus slip appears Also when a pallet is stopped by stopper the belt is still running underneath the pallet what in both cases causes again relatively significant noise Even though START and STOP events would be very convenient to determine by X axis since both events actually are a result of acceleration in X axis but due to mentioned noise in X axis it was infeasible to obtain reliable results in the end at all Y axis is from perspective of utilizabibility in this research absolutely useless the pallet is guided from both sides by rails and therefore there is no meaningful acceleration in this direction Z axis does not suffer by so many disturbances as X axis It is a convenient data source for event EXIT since taking a pallet out of the conveyor line means a continuous acceleration and deceleration in Z direction Moreover it was dis covered that Z axis can serve for sensing START and STOP events as well with surprisingly better results than for X axis When a pallet stops the noise level in Z direction significantly decreases in contrary to X axis when due to vibrations on stopper the noise level stays more or less the same which can be easily de tected When a pallet is released by stopper to move again the noise in Z axis rapidly increases The schema how the patterns look like for all the three events is illustra
24. initialization for each agent is visually displayed to the user in screen in form of percentage of completeness inside a waiting ring as is shown in Figure 4 12 IMPLEMENTATION 59 me 25 Fy Initializing Pallet lost Pallet EX TI 4 4 Figure 4 12 Visualization of all three agent s abnormal states Initiali zation phase lost synchronization and pallet exited line all in 25 of completeness 2 Synchronous tracking Synchronous tracking is a phase when an agent is continuously monitored and its estimated position corresponds to real position of the pallet so they are in mutual synchronization The only mechanism how to check it is based on RFID events and it is going to be described more in next phase So far in this phase it is supposed they are in the synchronization The synchronous tracking decision process goes in simply words as follows As soon as the event log is registered it is assigned to appropriate agent the agent that represents the pallet from which the event log came and algorithm determines utilities for all possible agent alternatives states The alternative with the best utility value in the end is chosen as the final one and the agent changes its state accordingly Thus the biggest challenge of this algorithm was to develop a mechanism for utility evaluation However before the mechanism can be described more in details it is necessary for better understanding to men tion a si
25. mutually differ mostly in different terminology and their comparison is shown in Figure 2 14 Since the second version 2 0 is the latest one the further explanations will refer to this version primari ly Each WSDL file term for a WSDL document describing certain Web Service in order to fully specify given Web Service must contain exact set of specifications There fore according to W3C definition a WSDL file defines service as collection of network THEORETICAL BACKGROUND 31 endpoints In WSDL the abstract definition of endpoints is separated from their con crete network deployment or data format bindings This allows the reuse of abstract definitions such as interface which is abstract collection of operations The concrete protocol and data format specifications for a particular endpoint establish a reusable binding Specifications summary is listed and more described in Table 2 5 42 48 49 WSDL 1 7 WSDL 2 0 description definitions o eS 23 Ze service Eg Cx Figure 2 14 Comparison of WSDL structures 48 Table 2 5 Description of WSDL 2 0 specifications WSDL 2 0 Term oe The root element of WSDL file that defines service name and Description namespace Collection of related endpoints Defines address or connection point to Web Service Often HTTP Endpoint l URL string Bindi Specifies a SOAP binding style RPC Document transport proto indin 5 col for the Web Service and endpoint Defines Web Se
26. occupants Naturally there are many environments where agents can operate but fortunately according to 3 it is possible to categorize them into several groups Each kind of environment then determines the appropriate agent design and techniques for its implementation Fully observable vs partially observable if agent s sensors have constant access to complete state of the environment so the agent has all aspects that are relevant to the choice of action the environment is called fully observable e g chess board a map for navigation agent On the other hand the environment is partially observable and that is the most common case when the agent does not have enough information for making an action choice for example due to missing sensors or noise in sensors e g automated guided vehicle in city traffic passengers construction sites or the automatic vacuum cleaner unknown distribution of dust closed doors etc THEORETICAL BACKGROUND 8 Deterministic vs stochastic if agent s next state of the environment is completely determined only by current agent state and agent action then the environment is said to be deterministic otherwise it is stochastic The environment can be partially observable and deterministic at the same time For instance battleship players don t see oppo nent s ship location but as soon as they select a target to fire they either hit or miss Episodic vs sequential in episodic envir
27. of localization too Those five movements are START STOP LEFT RIGHT and EXIT and their meaning is pre sented in Table 5 1 Along with the meaning the table also contains information how the data pattern looks like and which data source is scanned in order to discover it The DISCUSSION ON RESULTS 65 process how those two last properties were discovered is based on observations which were done by studying each of sensor axes and which are listed in following paragraph Gyroscope All gyroscope axes provide relatively reliable and noiseless data Gyroscope X axis roll and gyroscope Y axis pitch are for this research mean ingless since the pallet moves and turns only in horizontal plane of the conveyor system Gyroscope Z axis yaw is suitable for detection of pallet turnings in XY plane such as LEFT and RIGHT where the signal creates relatively well detectable s1 nusoidal curve as can be seen in schema in Figure 5 2a Noise level however does not allow to determine exact angle of turning unless special data mining algorithm would be implemented which is out of scope of technical possibilities of given wireless devices Accelerometer All accelerometer axes suffer by significant noise of which source is unambigu ously a way how the pallet is transported by conveyor system o Noise in Z axis is caused by relatively rugged conveyor system on which pallet jumps o Noise in X axis is caused primarily by the type of conveyor system
28. of many additional modules for various services For purposes of this research Spring MVC Model View Controller module was selected as a design pattern for in IMPLEMENTATION 50 ternal logic and also because this framework provides hooks for extension and customi zation for web applications and RESTful Web Services According to this pattern the overall software function is allocated into three main areas of responsibility Model View and Controller e Model model stores all data structures objects algorithms and represents whole computational power In this research the model keeps whole process model 1 e conveyor maps pallet information and locations history of event logs It also stores and applies all localization and other algorithms Finally the model is also responsible for keeping track of list of clients and amount of data that were sent to those clients e View view is a main input output interface through which the user can interact with the process A view both generates the graphical output for the user to de scribe states of various objects and at the same time it can accept user com mands too In this research the view or graphical user interface GUI is repre sented by a web page in web browser on client side For development of the web page a Google Web Toolkit was used as a perfect tool for transformation of pure Java code into JavaScript that is implemented in HTML The developed GUI for pallet monitorin
29. pal let is still in motion there were only two reasonable ways of its detection There could IMPLEMENTATION 45 be either machine vision applied to track the pallet continuously or each pallet can be equipped directly by small wireless sensors While first solution would probably pro vide very nice results it is very expensive because it requires installation of many cam eras compare to the latter solution which is significantly cheaper since the prices of MEMS sensors are very low nowadays and does not require any special hardware in stallation 4 4 1 Wireless embedded device with MEMS sensors In this research wireless embedded device produced by INICO Technologies was used to determine the pallet position The overall view of the device is shown in Figure 4 4 while in the Figure 4 5 1s the device open and there are shown its particular compo nents The main control unit of the device is a powerful ATMEL ZibBit chip with dual antenna and wireless IEEE 802 15 4 ZigBee OEM module capable of broadcasting in 2 4GHz frequency band 69 The chip is governed by Contiki OS an open source op eration system for low power embedded devices which is providing necessary capabili ties for wireless communication and signal processing The device is equipped by a 3 axis accelerometer and a 3 axis gyroscope which provide complete movement infor mation of the device in a space Additional sensor characteristics could be found in Ta ble 4 2 The de
30. pallet stopped and no event was accidentally received the agent still con tinues moving Because of ability of self discovery it was able to localize itself very soon again but in the end most of the time the agent spent just by self localization and thus most of the time the resulting probabilities of agent location were too small to pro vide sufficient determination of a place where the pallet actually was Generally speak ing the environment in which was agent controlled by MCL algorithm implemented was not suitable and any attempts to apply MCL for pallet localization on given testbed was immensely challenging or even unfeasible Final conclusions and observations of MCL Method strengths are e Versatility method can be easily applied to any kind of map e Simplicity and scalability the algorithm is very simple the amount of particles determines directly accuracy and computing difficulty and can be scaled flexibly according to computation power of processing device e Fast reaction algorithm gives in most cases very fast results while the agent is localized only in few steps e Self recovery problem of kidnapped agent is under certain conditions solved very quickly if at the place are available particles e Adyustability all model probabilities are easily tuned according localization re sults DISCUSSION ON RESULTS a2 Method weaknesses are e Incapability to react on delayed events algorithm is built to
31. parameters of each block by simple clicking on particular block and set ting appropriate properties from option panel For each block three possible properties can be set Stop RFID and WS of which description can be found in Table A 2 Any time if necessary there is a possibility to go back to Structure design and modify the map of which blocks are kept in building area However if the save button in map Structure design is pressed all Parameter design properties will be lost Table A 2 Additional block properties Property description Defines a block where pallet is supposed to stop mostly by a stopper Algo rithm uses this information to distinguish better between pallet stop caused by a stopper or by accident obstacle conveyor failure etc Defines a block that is equipped by RFID reader which provides helps pallet localization after receiving an RFID event Defines block membership to certain workstation which is part of Having this property allow algorithm to assign pallet occupying certain block to appropriate workstation APPENDIX A USER MANUAL FOR GUI 84 In this manner each block by block is parameterized and immediately saved by saved button The change is confirmed by visual representation as might be seen in Fig ure A 3 where a group of blocks were assigned to Workstation are thus those blocks are embodied into purple rectangular to show borders of the workstation Finally even though the map is
32. reader too Some event logs might be missing for some reason not generated lost etc Event log can come later contributing to a change in event order No event log duplication for the same event occurs Event RIGHT and LEFT are highly reliable and significant events if one of these events is generated it is very high chance the pallet really turned O W Ot o Events GO and STOP are not reliable and should have only support meaning 4 8 1 Monte Carlo Localization First attempt to process event logs and extract pallet location out of them was done using AI localization technique called Monte Carlo Localization MCL of which mathematical notation was presented in Chapter 2 1 6 For successful algorithm devel IMPLEMENTATION 55 opment and implementation a set of initial assumptions and clarification must be at first introduced e Because the pallet is pulled by the conveyor system under normal circumstances always forward the control signal for the algorithm is given by time interval For instance every 700ms there is a control signal generated to move a pallet one state forward e The necessary time interval for the pallet to move between two states in process map was measured on real production line and lately fine tuned e Probability of movement is always given by a vector of probabilities that deter mines with what probability the pallet stays after movement again in the same state with what probability it moves succe
33. stop and others In contrary RFID sensors distribut ed in the conveyor system inform the agent when its pallet the pallet which the agent represents passed certain location in the conveyor system Agent s action as a response to perception sequence is an adequate change of agent state visualized in the virtual map The agent must be fully utility based agent of which goal is going to be a Self localization within the conveyor system in the most efficient way The efficiency as a performance measure will in this research represent accuracy with which the agent copies actual pallet position 3 2 2 Environment description The research environment is represented by an arbitrary conveyor system in which each agent needs to localize its position The agent does not know its position in ad vance and once the position is known and agent is moving within the system there is no way how the location can be proofed While pallet moves there are many factors influ encing its movement stoppers other pallets operators defects etc but agent can per ceive the movement only by discrete event logs that may or may not come Even though the agent s current state is a result of previous agent s actions the next agent action is RESEARCH METHODS AND MATERIALS 39 depended only on current state and the action and thus has episodic character Therefore in summary the research environment is partially observable stochastic discrete and episod
34. was also effort to standardize a communication among them so many communication protocols were developed since that time such a Bluetooth ZigBee or LOWPAN 32 For this research 6LOWPAN was chosen because it was the first protocol on market which supported direct IPv6 addressing bLoWPAN actually brings IP Internet Proto col directly down to small low cost sensor devices The IPv6 utilization is key factor for creation of so called Internet of Things which allows every single thing in uni verse to have its own unique address Moreover the use of IP based protocols is very convenient because they are very well known technologies that do not require any mod ifications in the existing network infrastructure Devices can be directly connected to other IP networks for example Internet without translation gateways or proxies which can be seen in alternative communication systems using e g ZigBee 33 Machine to Machine M2M 34 or Bluetooth 35 This significant downside restricts their appli cation without additional gateways for use only in small scale local ad hoc networks However there is a strong trend in recent years among abovementioned protocols which converge also towards Internet based approach The working principle of 6LOWPAN will be more described in next paragraph 32 36 2 4 1 6LOWPAN architecture 6LOWPAN architecture as shown in Figure 2 11 relies on islands of IPv6 stub net works of low power wireless are
35. where are shown experiments related to localization of mobile robot and pedes THEORETICAL BACKGROUND 36 trian using inertial sensors Nevertheless there is a lack of real applications especially for industrial purposes in general 2 6 3 State of the art vs the topic of this thesis Based on the research survey it is evident that a topic related to localization is be tween researchers still very attractive Their main interest is split into two sections While the first one is related to localization based on wireless sensors the latter aims to improve localization algorithms based on signal from inertial sensors However in case of the topic of this thesis an attempt to categorization is not unambiguous at all Alt hough according to the data source it inclines to the localization based on inertia sen sors there are a few significant reasons that make this topic unique Firstly vast majority of the current research works related to inertial sensors use for localization algorithms direct signal from accelerometers or gyroscopes The continuous sensor signal is easily processed quantities such a speed or acceleration are evaluated and appropriate localization is determined and calibrated However in case of this re search there is introduces brand new approach for object localization based only on a discrete amount of event logs describing object movement In this research the raw sen sor data are preprocessed already within embedded
36. 72 is displayed there and then by pressing arbitrary events If a user wants to use RFID event a neces sary name of workstation must be inserted into text box on the right of RFID button in format WSx where x is a number of workstation Furthermore besides event button there are also two special buttons The button Reset calls a reset function in server that deletes all stored values of model and restores its initial state There is also a button Cal ibration that initialize a manual calibration process recalculation of all process coeffi cients based on calculated overall efficiency Information Infomation Efficiency 79 0 Control ID 9 7 Workstation WS5 RFID Wo EXIT Laps 1 Position 2300 1000 Lap complete 00 01 25 Lap time 00 00 05 LEFT 172 RIGHT Total time 00 01 30 START STOP Control Reset Cal brate Vanables Variables Figure A 7 Example of Information panel and Control panel APPENDIX A USER MANUAL FOR GUI 88 The purpose of Variables panel is to provide general overview of all process model coefficients The user can freely change arbitrary coefficient in order to improve the process For each coefficient there 1s available a text box where a user can write down a desired value Right next to it there are two numbers the first black number repre sents current value of the coefficient stored in model while the grey number in square brackets informs a user about initial value of the coefficient bef
37. Ashley W Stroupe Kevin Sikorski Tucker Balch Constraint Based Landmark Localization Online 2002 http pdf aminer org 000 560 920 constraint_based_landmark_localization pdf Wolfram Burgard Sebastian Thrun Markov Localization for Mobile Robots in Dynamic Environment Online 1999 http www cs cmu edu ats cs project jair pub volume 1 fox99a html jair localize html Thrun Sebastian Bayesian Landmark Learning for Mobile Robot Localization Online 1997 http robots stanford edu papers thrun landmark pdf Dieter Fox Wolfram Burgard Frank Dellaert Sebastian Thrun Monte Carlo Localization Efficient Position Estimation for Mobile Robots Online 1999 https www aaai org Papers AAAI 1999 AA AI99 050 pdf Global Digital Signal Processors DSP Market by Intellectual Property IP Design Architecture amp Applications 2011 2016 Online Digital Signal REFERENECES 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 78 Processors Market February 2012 http www marketsandmarkets com Market Reports global digital signal processing market 157 html Gottschalk Mark A Sensor make cars smarter Online DesignNews 10 6 1997 http www designnews com document asp doc_id 225450 amp dfpPParams ind_ 184 industry_auto aid_225450 amp dfpLayout article Kempe Worker Inertial MEMS Principles and Practice m sto nezn m C
38. CU aie sescecasnacegtncacecamanaesad NN 38 322 Environment CES CH POM esen A 38 32 EN VANONIMECIL MOU hean sadeuieneinas oaeuia ras aiaaes 39 3 2 4 Appropriate algorithm SelectiONn ceeeeccccccceeeesseeeeeeeeeeeeeeeeeeees 39 4 7 EaI test ile miele lla AEA Mee TRC nen re eer Kn Oe en re rer anh ne wre tener ne rere Terr rete 4 4 1 Product OM TING 43255 cobodecveetbaveieuvtaakedcatacus beteduanetareltawbanitecdieessbatedgsned wiaieoetens 4 Aled DCS CRIPUION cag ceaceomevsaaraenascce aE ERT 4 AD PROCUCHOMOPCTALMOMS copicistae Zostse acetate a e ete A eee 42 4 3 PN ramen nee rm ate mmr mo Ie ane REN mee ISN eS 43 4 4 DEVICES ccidi cite T E 44 4 4 1 Wireless embedded device with MEMS sensors scncccccceecscens 45 4d2 REDD UNPC tac Geteci Om 2 acerceivienceainweeactetolereeaiuaccunncen 47 4 5 Data Pranst er architecture mersani E A OTEA 48 4 6 Overall software architect le ocinrenissisrine naiiai A E E 49 4 7 Environmen mode benenspiuen ni A N A A SN 53 AS Alborthm implementa on sieneen a eee enceeis 54 4 8 1 Monte Carlo Localization cc cc ccccccesecccesecccessecceeseceeenecceeeeecens 54 4 8 2 Simple decision making algorithm cc ceeeceeecceeeeeeeeeeeeeeeeeees 57 PDISCUSSIOM OL TO SUNUG aap a 64 5 1 Signal processing and event detection essssssssooeerssssssseceersssssssseeeressss 64 D2 sBGCeWOMEE POSIL OM srei E E ues 67 5 3 Results of used artificial intelligent methods
39. IEEE 802 15 4 MAC ak IPv4 IPv6 IP Multicast IEEE 802 15 4 PHY Figure 2 16 Comparison of DPWS and 6LOWPAN protocol stack 32 54 2 6 State of the art in real time pallet localization Although the production engineering was in industry always in the centre of interest and both companies and academic researchers from all over the world invest many re sources to its development it is astonishing that only a few of the research studies pub lished in recent years were actually related to localization of pallets within a production line Not that there would not be an interest in issues related to localization problems but the vast majority of research prefers to localize more interesting objects such as human beings or mobile robots Nevertheless since a contribution of their technologies might be easily applicable for industrial environment these works were included into a literature survey as well According to the survey in way to find the present edge of industrial research in this field it was discovered that there are currently two areas of main research stream inter ests localization using wireless sensor networks and localization based on inertia sen sors In next two paragraphs both techniques are going to be described in more details 2 6 1 Localization based on wireless sensor networks WSN During the last years localization based on various wireless communicators and sen sors networks has enjoyed raising pop
40. O Figure 5 6 Results of Experiment B Experiment D E Proximal wee 123 45 6 7 8 9 101112131415 1617 18 19 20 Minutes Number of messages Figure 5 7 Results of Experiment D DISCUSSION ON RESULTS 71 5 3 Results of used artificial intelligent methods 5 3 1 Monte Carlo Localization Monte Carlo Localization has excellent results for localization of mobile robots and thus it was expected to be a promising solution for pallet localization as well However this algorithm did not live up to given expectations In fact the major obstacle was in a way of which measurement data were gathered In case of mobile robots the robot at certain time moves with some probability to certain place and makes a measurement of its environment around Then it gathers data evaluates them and continues in move ment So the measurement data are tightly related to the particular movement space and time interval In contrast to the case of mobile robots the pallet movement is continuous uncon trolled Gust pulled by conveyor belt and at the same time independent on measurement Therefore a measurement results do not correspond to any particular movement have artificial timestamp which is created not at time the event log is created but when it is received and due to their transmission technology they might be even significantly de layed or lost In reality it was very difficult to synchronize pallet movement with agent movement If
41. T E32 fao retk s 13E lt PRIR ado APPRIT 113524 Soh egy KES 1 oR j behets leh A e OL i 4 ee 398 gong fae pit Vio Sa he ae i f E PNT h iA TE 456 sS H ilf H 3 X o OO wepeco ape 92 6 1 48 fi iii STOP START GO 1824 EXIT p008 boao 980 Figure 5 3 Example of event generation from data signal of Z axis accelerometer 5 significant areas a pallet moves b pallets stopped by stopper c pallet moves again d pallet stops e pallet is taken by operator away from conveyor system 5 2 Edge router position The transmitted signal strength from used wireless devices is usually very low due to their simple low powered antennas The situation gets even worse because both the chip with antennas is packed into plastic cover and the wireless device itself 1s encapsulated into the pallet structure The design of the pallet is done in such a way that the pallet body in which the device is hidden influences negatively the properties of signal transmission Based on observations it was discovered that the top pallet metal plate shields the device signal very significantly whereas the pallet plastic body is more sig nal permeable but on the other hand its dome structure causes with high probability signal reflections Therefore there left only one clear propagation way for the signal and that is down towards ground through holes in pallet bottom The signal propagation schema is shown
42. TATION 54 Of course the line design interfaces is versatile and allows user to create any other line in the same way as for the production line used in this research Moreover there are options to define each block to be part of certain workstation to have stopper or RFID reader Finally the map is exported in form of text line to main server where is stored translated into appropriate machine language and immediately prepared for usage 4 8 Algorithm implementation To ensure smooth functionality of the two algorithms which are going to be devel oped and implemented at first a list of assumptions needs to be created to summarize all significant characteristics of given testbed with respect to pallet localization Assumptions e Conveyor system o Map or schema of conveyor system is known in advance o Conveyor system is flat no lifters o Pallet movement is strongly restricted by a conveyor system pallet never leaves the conveyor system except a case when a pallet is taken manually away from the line o If the pallet leaves conveyor the pallet is no anymore monitored and it must be initialized again o Conveyor drags a pallet always only in one direction forward o Conveyor can stop o The conveyor system can accommodate many pallets at one time o Pallet can start from arbitrary place e Events o Motion events are generated from sensors attached to pallets Besides motion events there 1s available information from RFID
43. WPANSs which are connected to other IP network through one LoWPAN Edge router e Extended LoWPANSs use multiple edge routers within a single network con nected by a backbone link e g Ethernet e Ad hoc LoWPANS are not connected to any IP network e g Internet they work without any infrastructure one of its routers takes on the network respon sibilities and serves as a simplified edge router Instead of making and managing complex route tables for network communication e g in ZigBee 6LOWPAN uses so called Neighbour Discovery ND ND is a basic mechanism of IPv6 responsible for Router Discovery Address Resolution Duplicate Address Detection Redirection and Prefix and Parameter Discovery All these functions simplify registration of node to an edge router of any arbitrary LOWPAN and hence allow free movement of nodes throughout the LOWPAN between edge routers or even between whole LoWPANS 32 36 An ideal application of the 6OLOWPAN in industrial environment would be a network of sensors distributed around the factory collecting data not only in the raw form of digital or analogue inputs but also having capability to process data to give more mean THEORETICAL BACKGROUND 28 ingful information such a temperature humidity light intensity power consumption etc Wireless technology provides a flexible monitoring solution allowing the sensors to be attached to mobile machines such as AGYs fork lifters and obtaining information n
44. Y sicceess sien aetna aesaawenhe a 9 PN Te CalWZatlOniceiacerssateesiecdiietaenea dia a a aa 15 22 ADEVICES ANG TECHNOLO ie Senrerieia a 20 2 2 1 Micro system technology MST ccccccccccccceessseeseeeeeeeeeeeeeeeees 20 De dOCEVET OIC CIS e E ie iuekeleedieitel 21 Dede NGYTOSCODES eera aan a a Muti 22 224A SCNSOMCOMPOMEMIS erena A uueatiaestacugtedeeedaes 24 293 Radio Freguency IdennficatioM esien eana N 25 2e OLOV EAN corone a a A A I le aeeestoreeet 26 ZAM GLOWPAN architecture ceria e a 26 2 5 Service Oriented Architecture and Web Services cccseeeseseceeeeeeeeeeeeees 28 Boul Web Service archite Chine cerida aa E AN 29 ye DONE r nent er na er Tern Teme er ee nee Tene er a eT 29 DDD WDD eeceiaietntunasda di needs au cerersiacedau a O 30 2a OPW Sissies heceaciys gnacthiea Ea uston entete testes anes ees 32 2 6 State of the art in real time pallet localization cc eeeeeeeeeeeeeeeeeeeeeees 33 2 6 1 Localization based on wireless sensor networks WSN 33 2 6 2 Localization based on data of inertial sensors ccceeeeeeeeeeees 35 2 6 3 State of the art vs the topic of this thesis nnsesssseseeesssssseee 36 Researchime thods and materials s r oasis tacuuiasnslrawta taste n a Manweh 37 3 1 Signal preprocessing and event detection sseoeeesssssssssceersssssssseerrrssssss 37 3 2 Localization algorithm using Artificial Intelligence ee eeeeeeeees 38 J2 NB
45. a networks LOWPANSs Note The terms 6LoWPAN and LoWPAN are often used interchangeably In this paper however the 6LOWPAN will refer to general technology and LOWPAN to a specific network or node A stub network is a network which is capable of receiving or sending IP packets but does not act as transit to other networks Each LOWPAN network might consist of router nodes and host nodes These nodes differ to each other only in way whether a proximal node takes a responsibility of message redirection between distant nodes and edge router which one would be otherwise for the distal node out of reach Edge router is a device that routes traffic in and out of the LOWPAN while handling 6LoWPAN compression and Neighbour Discovery for the LOWPAN If a LoWPAN is connected to IPv4 net work it 1s an edge router which handles smooth interconnectivity In case of Ad Hoc THEORETICAL BACKGROUND a network the role of edge router is given to arbitrary node which fulfils at least the most essential functions for smooth network running 32 37 EEE eo ow _ Y E N Remote server Internet f gn Edge router R R W H f H Edge router z Edge router N Se PEN R R R H Simple LOWPAN Extended LoWPAN H Host R Router Ad hoc LoWPAN Figure 2 11 6LOWPAN architecture overview 37 Three major types of LOWPANs are distinguishable based on their architecture 32 e Simple Lo
46. aking and depends both on decision making environment and used decision theory model Generally there are distinguishable three types of decision making environments for which is defined several decision theory models 3 6 7 1 Decision making under Certainty complete and accurate knowledge of util ity for each alternative there is only one utility value for each alternative 2 Decision making under Risk multiple possible utilities of each alternative and each state of nature can be identified and probability of occurrence can be attach to each state of nature 3 Decision making under Uncertainty multiple possible utilities of each alter native and each state of nature can be identified but there is no knowledge of the probability of any state of nature Decision under Certainty if all utilities are certain and known the decision itself is reduced to simple choice of alternative with optimal utility For instance there are two products A and B Product B is 15 cheaper than product A while quality of both is the same Then the decision is clear and product B is chosen Decision under Risk instead of optimizing given utilities as in case of decision under certainty the decision under risk optimizes expected utility while there is only a given probability for each utility that it might happen The most common method is called THEORETICAL BACKGROUND 12 Expected Utility EU or in connection with money decision Expec
47. already saved in server there is also possibility to export the map into single URL string for storage purposes Thus if server crushes the map can be restored again by just inserting the URL string into a web browser com mand line Only in case of a new designed map FASTory line is pre coded and loaded always automatically with server start Now the model map is ready to be used eSONIA4 Figure A 3 The Parameter Design tab of Design Page a active tab b property setup area c working area d example of workstation membership representation Real time monitoring Page This web page serves as a main interface for real time monitoring and control of pal let localization system and can be seen in Figure A 4 The page consists of two main distinguishable areas a Monitoring area and Control area Monitoring area monitoring area truly dominates the page It is the main visuali zation element where the map of conveyor system is displayed as well as the locations of all the pallets To improve human readability and overall visualization credibility there is also a background image of CAD model of the real production line displayed behind the map that matches in dimensions with the designed map In a case that a dif ferent production line is designed and monitored than the FASTory line of which image is
48. ambridge University Press 2011 Vigna Benedetto Physical Sensors Drive MEMS Consumerization Wave STMicroelectronics Accelerometer Online Wikipedia 2012 http en wikipedia org wiki Accelerometer Bernstein Jonathan An Overview of MEMS Inertial Sensing Technology ModernMedicine Gyroscope Online Wikipedia 2012 http en wikipedia org wiki Gyroscope Vibrating structure gyroscope Online Wikipedia 2012 http en wikipedia org wiki Vibrating_structure_gyroscope Gyroscope Online SensorWik1 org http sensorwiki org doku php sensors gyroscope Bo Zhang MTE Kahn Overview and Improving Fiber Optic Gyroscope Based on MEMS NEMS Fabrication editor Institute of Physics Publishing Journal of Physics 2006 International MEMS Conference Vol 34 Fibre Optic Gyroscope Online Wikipedia June 2012 http en wikipedia org wiki Fibre_optic_gyroscope Difference Between RFID and NEC DifferenceBetween net Online 2012 http www differencebetween net technology difference between rfid and nfc Radio frequency identification Online Wikipedia 2012 http en wikipedia org wiki Radio frequency_identification A New RFID Privacy Framework Has Been Taken Up by Europe RF Dcanada Online November 2011 http www rtidworld ca a new rfid privacy framework has been taken up by europe 333 RFID Tag Direct Industry Online 2012 http www directindustry com prod premo corporation rfid tags 50493 387062 html Zach Shelby Carst
49. and which section of conveyor part was smoothly ridden Information about pallet delays can also serve to determine production line bottlenecks but such a tool would be out of scope of this thesis Example of time mode is illustrated in Figure A 5 Efficiency mode is another visualization mode that graphically displays the effi ciency of event generation As was already said every block of map is defined by cer tain number of input events the events that would be received in ideal case when a pallet enters the block The difference between ideal case and real one is represented by colour change as follows No event missing means the circle is displayed in green col our One missing event in orange and two and more missing events in red colour The efficiency mode gives an overview how well the events are generated in which section of line a sensor has difficulties of event generation and in fact also to graphically recog nize whether the pallet tracking is actually effective or not APPENDIX A USER MANUAL FOR GUI 86 Figure A 5 Example of main monitoring area with time mode Inseparable part of monitoring area is also an Event area Event area shows a line of received events represented by appropriate icons for one chosen pallet in order they were received Their graphical representation in Event area is shown in Figure A 6 STOP RFID LEFT EXIT START RIGHT Figure A 6 Example of Event area with labelled events Control ar
50. bruary 2006 Web Services architecture overview Online DeveloperWorks 2000 http www ibm com developerworks webservices library w ovr SOAP over UDP Version 1 1 Online OASIS 2009 http docs oasis open org ws dd soapoverudp 1 1 os wsdd soapoverudp 1 1 spec os html SOAP Online Wikipedia 2012 http en wikipedia org wiki SOAP Overview of SOAP Online Oracle 2002 http java sun com developer technicalArticles xml webservices WSDL Online Wikipedia 2012 http en wikipedia org wiki Web_Services_Description_Language Web Services Description Language Online W3C 2001 http www w3 org TR wsdl Devices Profile for Web Services Online OASIS 2009 http docs oasis open org ws dd dpws 1 1 os wsdd dpws 1 1 spec os pdf G Moritz DPWS for bLoOWPAN Online 2010 http tools ietf org html draft moritz 6lowapp dpws enhancements 01 Device Profile for Web Services Online Wikipedia 2012 http en wikipedia org wiki Devices_Profile_for_Web_Services Understanding Devices Profile for Web Services WS Discovery and SOAP over UDP Online MSDN 2012 http msdn microsoft com en us library dd179231 aspx Devices Profile for Web Services Online Doxygen Trac 2007 http trac e technik uni rostock de projects ws4d soap doxygen stable html dpws_page html Anurag Kumar D Manjunath Joy Kuri Wireless Networking 2008 Guoqiang Mao Baris Fidan and Brian D O Anderson Wireless Sensor Network Localization Techniques Onlin
51. came from does not correspond to a workstation in which the agent is located the agent is said to be lost visualization is shown in Figure 4 12 The algorithm that would localize the agent again is almost identical to the one presented in the initial phase The algorithm again waits until initial se quence of events is received and if the initial pattern is found the agent state or location is again recovered and agent is again synchronously tracked Besides RFID event also EXIT event can get the agent into the asynchronous tracking phase and again the initialization sequence of events is needed in order to track the pallet again IMPLEMENTATION 63 Based on previous paragraphs where fundamentals of the algorithm were explained one can notice that there are many coefficients in use Even though they are key compo nents for smooth algorithm operation they were chosen based only on empirical experi ence and they do not have to be optimal for given process Therefore a revalidation mechanism was implemented in order to verify given coefficients and propose new co efficients that would lead to better efficiency i e better accuracy of agent positioning Firstly however a benchmark according which the tracking efficiency is measured must be introduced As the benchmark an overall localization efficiency of all agents will serve and which is calculated as an average value of all particular agent movement efficiencies The movement efficie
52. development was a part of this research too Thus the prime goal of the algorithm was to extract information of significant pallet movements from a stream of raw digital data coming from the sensor The choice of appropriate data mining method which would be able to extract de sired information was significantly influenced by low processing capabilities of their chips Therefore as a most convenient solution a method called pattern matching was chosen Pattern matching is a simple algorithm that checks a sequence of data for presence of constituent of some pattern The pattern is mostly known in advance and describes such a signal behaviour that corresponds to known event a change in amplitude a cer tain shape a phase shift Therefore whenever a sensor signal matches with certain pat tern the corresponding event is generated The methodology of implementation of pat tern matching goes as follows 1 Record raw data of all sensor axes while pallet moves throughout the con veyor system and note down all significant pallet movements 2 Study thoroughly the records and try to distinguish a special signal pattern for each pallet movement 3 Design an easy algorithm which would be able to recognize the patterns from raw signal and simulate it on recorded data 4 Incase of success implement it to chips of wireless devices RESEARCH METHODS AND MATERIALS 38 3 2 Localization algorithm using Artificial Intelligence Artificial Inte
53. device and only if the sensor signifi cantly moves over some threshold value an event log carrying movement information 1s generated and received by localization algorithm Of course this development would never be feasible if very important assumption is not declared The object is an indus trial pallet which moves in predefined and known conveyor system Nevertheless based on the survey it seems that such a research scenario is still unexplored which makes its development even more challenging 3 RESEARCH METHODS AND MATERIALS 3 1 Signal preprocessing and event detection Nowadays the wireless embedded devices equipped by modern MEMS sensors do not have to bother with analogue to digital conversion of the sensor data any more since the sensors include in their packages already an embedded A D convertor Never theless another type of conversion is desired In recent years there has been a signifi cant effort to shift also a low level data preprocessing directly into embedded devices As aresult such smart devices are able to convert a stream of raw digital sensor data into a set of discrete event logs containing rich and meaningful data instead This ap proach brings a significant advantage to all consumers since they are not required to have any knowledge how the data were actually processed and generated The wireless devices used in this research are however not equipped by any prepro cessing mechanism and so the algorithm
54. e brain 1s called agent program and its design is the main job of AI The goal of the agent program is to take the current per cept as input from the sensors and return an action to the actuators Based on 3 there are four basic kinds of agent program distinguishable which encompass the principles underlying almost all intelligent systems Simple reflex agents the simplest kind of agent where decision is made exclu sively on current percept according condition action rule while ignoring the rest of per cept history Condition action rule is set of conditions if something happen and appro priate actions do this For instance if there is a green in traffic light continue driving if there is a red colour stop the car Simple reflex agents are very easy to develop but their intelligence is very limited THEORETICAL BACKGROUND 9 Model based reflex agents this kind of agent maintains some sort of internal model of how the world works The model is either already in built or built from per cept history that agent collects during his existence The presence of model is for agent in partially observable environment very profitable since it helps agent to keep track of the part of the world it cannot see right now Goal based agents compare to the two abovementioned agents where decisions were made either according specific situation e g every right junction turn a car left or stored knowledge respectively e g
55. e 2006 http sydney edu au engineering electrical people guogiang mao UserFiles File publications Mao06 Wireless pdf REFERENECES 57 58 59 60 61 62 63 64 65 66 67 68 69 80 Localization Algorithms in Wireless Sensor Networks Pal Amitangshu 1 2010 Network Protocols and Algorithms Vol 2 Jennifer Yick Biswanath Mukherjee Dipak Ghosal Wireless sensor network survey Elsevier 2008 Computer Networks Piedrahita Giovanny Andr s Guayacundo Diana Marcela Evaluation of Accelerometers as Inertial Navigation System for Mobile Robots 2006 Robotic Symposium Rong Peng Mihail L Sichitiu Angle of Arrival Localization for Wireless Sensor Networks Online 2006 http www4 ncsu edu mlsichit Research Publications aoaLocalizationSecon06 pdf Neal Patwari Alfred O Hero III Jose A Costa Learning Sensor Location from Signal Strength 2007 Advances in Information Security Eddie C L Cahn George Baciu S C Mak Using Wi Fi Signal Strength to Localize in Wireless Sensor Networks 2009 International Conference on Communications and Mobile Computing Sarah Spieker Christof Rohrig Localization of Pallets in Warehouses using Wireless Sensor Networks 2008 16th Mediterranean Conference on Control and Automation Christof R hrig Sarah Spieker Tracking of Transport Vehicles for Warehouse Management Nice 2008 International Conference on Intelligent Robots and Systems
56. e briefly described in next paragraphs In addition a special chapter dedicated only to MLC will follow later on 13 Triangulation depends on the range and bearing to set of landmarks and us es geometrical functions to compute a single point that is most consistent with the current location Triangulation is probably the oldest localization techniques known from ancient times and surprisingly even nowadays many robots systems use for their navigation right this method Triangulation may be done with low weight given to previous object positions and probably nowadays the most known application is GPS 13 Markov localization addresses the problem of state estimation based on sen sor data It is a probabilistic algorithm that maintains a probability distribution THEORETICAL BACKGROUND 17 over the space of all hypothetical places beliefs where a robot might be The probabilistic representation allows weighing each hypothesis in a mathematical ly sound way Markov localization is characterized by updating all state beliefs in form of new probabilistic belief distribution Bel after every sensor meas urement This method is very suitable for kidnapped robot problem since it al ways maintains also other hypothesis However even though Markov localiza tion is quite efficient it requires some extra computational time to compute all coefficients and covariance which can be critical for some real time localization systems 14 The F
57. e installed around the production line As an edge router a popular S1000 controller produced by Inico Technologies was used However this controller is in comparison to S1000 con trollers used in control of production line supplemented by a W1 Z add on wireless module as can be seen in Figure 4 8 While the wireless module is responsible for all 6LoWPAN duties such network initialization communication maintenance or security the S1000 itself serves only as a gateway to IPv6 network without having any pro cessing responsibilities Both edge routers are connected to the same backbone IPv6 local area network and to internet respectively Figure 4 8 Edge router S1000 controller with wireless module Sensor nodes are directly accessible by their unique IPv6 which allows easy access and quick configuration Each sensor node consists of a list of subscribers which will receive further the sensor data To minimize the network traffic only one subscriber is preferably used The prime subscriber in this research is a main server called InfoStore which is responsible for all data collection and processing SCADA or other visualizing device can display either raw sensor data by direct sensor accessing or processed data from the main server 4 6 Overall software architecture As a core platform for server software architecture an open source application framework called Spring was chosen due to its scalability interoperability and availabil ity
58. e is captured in Figure A 1 where labelled circles represent buttons to the individual subpages The user can choose by clicking on appropriate bubble from three different modes e Design new line designing brand new map of conveyor system e Real time monitoring real time tracking of pallets Offline monitoring offline tracking of pallets based on event stored in da tabase facet APIS eSONIA 4 Design new line Offline monitor Real time monitoring Figure A 1 Welcome Page of APIS with three bubbles that creates hyperlinks to page of Real time monitoring Offline monitoring and Design of new line APPENDIX A USER MANUAL FOR GUI 82 Design Page Design page is an interface for development of own process model represented by a map of the conveyor system The process of map design consists of two phases a map structure design and a parameter design located in the page in two separate tab panels Map structure design it is a first phase of map modelling where user builds a raw structure of a conveyor system as a combination of well defined blocks similarly to LEGO principle as shown in Figure A 1 The building blocks are available in block panel and each type represents certain section of conveyor system such as right turn left turn straight and others Summary of all available blocks is listed Table A 1 Block by block is placed into arbitrary position in building area by simple drag and drop opera tion I
59. e particle distribution is given to all blocks in this case there are two particles Thereupon a RIGHT event is measured IMPLEMENTATION 56 c d e Now for each block i a simple calculation is done All particles from each box are multiplied by an appropriate probability whether the particle is on is not at certain place while some measurement is done and their probabilities are summed into non normalized importance weight W For example for a first left upper block i 1 there are two particles multiplied by a probability of 0 8 because there are placed in right turn block and right turn event was detected Thus the value of non normalized importance weigh is Wnn 2 0 8 1 6 The same operation is done for each block and their results are summed into normalizing factor 7 n n Onn 5 16 3 04 9 2 i 1 Next step is a normalization of all non normalized weights Therefore the non normalized importance weight of each block i is divided by the calculated normalizing factor y to get normalized importance weight w Wnn Wn r So for all right turn blocks the normalized importance weight is 0 1739 and for other blocks only 0 043 Algorithm is now almost complete In last step it is necessary just to recalculate number of particles for each block by following formula where Np is total number of particles NV 8 2 16 In this way we get for all right turn blocks 2 7 particles and for rest of block j
60. ea in control area of the page there are located all adjusting elements that an operator needs to interact with localization algorithm They are arranged in dedi cated panels called Information panel Control panel Variables panel and Display pan el and next paragraphs will bring their description The Information panel which is shown in Figure A 7 displays information of all pallet properties that serve for pallet movement analysis List of included properties is presented in Table A 3 APPENDIX A USER MANUAL FOR GUI 87 Table A 3 Description of elements shown in information panel Overall efficiency is calculated as a percentage ratio between total amount of input events and real amount of events Tt Pallet identification number OC lt Number of laps the pallet did since was introduced to the system Position Actual pallet position in millimetres measured from upper left corner of production line WS8 Lap complete A time of last complete lap A time of actual lap Total time of pallet movement The Control panel shown in Figure A 7 has the same function as a remote control for TV With its help an artificial pallet can be introduced to the system as well as it is pos sible to generate artificial events This panel serves primarily for testing purposes and during normal monitoring phase it should not be used The artificial pallet is created by writing its ID into text box in the centre of navigation star a number 1
61. ecifies a set of in build services that each device must implement such as e Discovery services self advertising and discovering other services e Metadata exchange access to metadata of other hosted services e Eventing and Subscription services DEVICE 1 Hosting SERVICE HOSTED MESSAGEs CLIENT 1 SERVICEs DEVICE m CLIENT n Hosting SERVICE HOSTED SERVICEs MESSAGEs include discovery description control and eventing Figure 2 15 Architecture layout of DPWS clients devices and services 53 In the Figure 2 16 there 1s shown common DPWS stack with all transport protocols and all DPWS specifications such as WSDL SOAP WS Policy WS Addressing WS Discovery WS Eventing or WS Transfer However an explanation of these specifica tions is out of the scope of this thesis and can be found for instance here 50 Neverthe less the Figure 2 16 presents also a stack of 6LOWPAN protocol to demonstrate their THEORETICAL BACKGROUND 33 mutual similarity From the figure is clear that both protocols use the same network and transport protocols i e Pv6 protocol and UDP TCP transport protocol and therefore it is very easy to utilize Web Services in the 6LOWPAN communication DPWS stack 6LOWPAN stack Application Specific Protocols Application Protocols i c 7 toes WS Eventing bes Discovery UDP WS Policy WS Addressing TCP wali HTTP Transfer SOAP WSDL XML Schema IPv6 Y HTTP LOWPAN UDP TCP
62. eeeeeeeeeeees 53 Figure 4 11 MCL algorithm for pallet localization ce eeseeeseeseeeeeeecceeeeeeeeeeeaeaeaees 57 Figure 4 12 Visualization of all three agent s abnormal states Initialization phase lost synchronization and pallet exited line all in 25 of completeness 59 Figure 4 13 GUI representation of algorithm behaviour with calculated penalties for two dilere ING SCC IAI lOS raene a a A O 62 Figure 5 1 Accelerometer and gyroscope AXES cccccccccssseessseeccceeeeaeeeseseeeceeeeeaeeesess 64 Figure 5 2 Pattern matching a Z axis of gyroscope b Z axis of accelerometer 66 Figure 5 3 Example of event generation from data signal of Z axis accelerometer 67 Figure 5 4 Signal propagation schema a normal position b upside down position 68 Figure 5 5 Efficiency of received messages in dependency of mutual position between Cdee TOULCT ANG GE VICES sari kcussteutiacrtin n a 68 Figure 5 6 Results of experiment B Figure 5 7 Results of experiment D VII LIST OF TABLES Table 2 1 Decision matrix with expected utilities for case of businessman 12 Table 2 2 Opportunistic loss table for businessman example ccceeeeeeseceeeeeeeeeeeeees 13 Table 2 3 Criterion realism table for businessman example ccccseeseseeceeeeeeeeeeeees 14 Table 2 4 ACCelerOmeter Properti csiecaiswiasstansiaras bn t dasvsndtara a R 22 Table 2 5 Description of WSDL 2 0 specifications
63. efined initial event pattern The ini tial pattern is artificial set of events that must be received from a pallet to suc cessfully accomplish initial phase of the agent The initial pattern can be freely chosen by a user and consists of combination of any three known events such as for instance RFID LEFT RIGHT or RIGHT LEFT RIGHT etc The unique pattern consisting of three events received right after each other should be sufficient for vast majority of conveyor systems If conditions allow the user can choose also an event UNDEFINED to skip a re quired event and reduce amount of events in initial pattern By default used in this research the first event must be a RFID event RFID event provides infor mation about exact pallet position within the conveyor line and thus precisely localizes agent Therefore there is no need for any other events more and thus by default the two rest initialization events are chosen to be UNDEFINED In case of FASTory line used in this research as a testbed the RFID event must be always the first initial event because of symmetry of conveyor system As soon as the initialization algorithm receives the right combination of events it launches a scan matching a technique that tries to match the given ini tial pattern with a priori model of the conveyor system If such a pattern is found the initialization phase is over agent is successfully localized and it switches to synchronous tracking phase Process of
64. en Bormann 6Lo WPAN Wiley 2009 ZigBee Alliance Online http www zigbee org M2M Comunity for machine to machine embedded wireless and connected consumer devices development Online www m2m com Bluetooth technology Online www bluetooth com Hunn Nick Essentials of Short Range Wireless Cambridge University Press 2010 The Wireless Embedded Internet Online http 6lowpan net REFERENECES 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 79 Jaac n Martinez Jos L M Lastra Application of 6LoWPAN for the Real Time Positioning of Manufacturing Assets Tampere Online 2011 http www iab org wp content IAB uploads 201 1 03 Martinez pdf Ricardo Silva Jorge S Silva Claudio Geyer Valderi Leithardt Fernando Boavida Use of GPS and 6LOWPAN in Mobile Multi Sink Wireless Sensor Networks Issues and Feasibility 2011 Hugo Haas Allen Brown Web Services Glossary Online W3C http www w3 org TR ws gloss Thuraisingham Bhavani Secure Semantic Service Oriented Systems Auerbach Publications 2010 ISBN 978 1 4200 733 1 7 Minor Johannes Bridging OPC UA and DPWS for Industrial SOA Tampere Tampere University of Technology 2012 Jose L Martinez Lastra Ivan M Delamer Semantic Web Services in Factory Automation Fundamental Insights and Research Roadmap IEEE Transactions on Industrial Informatics Vol 2 No 1 Fe
65. en technology and right this feasibility was by this research finally also successfully proved as was described in the conclusions This the sis work somehow opened doors to new promising area of wireless sensor application and give a potential for next growth and experiments Therefore there 1s going to be proposed in next paragraph a possible direction in which a future work might go One of the possible ways how to continue in this research would be better utiliza tion of resulting data from this research Since the developed localization algorithm provides complete information about all pallet movements and locations throughout the process it would be very convenient to use this information not only for passive moni toring as is done in this research but as valuable data source for active process control too Nowadays process mining seems to be one of the most convenient tools for active management and control The process mining has potential to build brand new process model just based on those data it can provides conformance checking in which the real process is compared with model outputs or process mining can also actively modify and redesign process model based on data from this work to increase overall process effi ciency Therefore the utilization of process mining is proposed to be next and most con venient step in this research 7 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
66. ent logs stored in data base serve later on for further analysis or for offline monitoring As soon as the event log is processed by the process model and pallet location is updated it sends result data IMPLEMENTATION 51 pallet location back to the controller together with a list of clients that are going to be informed about this update in pallet location Sensor Edge router smart RTU Assembly Line Ehternet DPWS bLOWPAN J User interface Web y browser WS endpoints REST controller o Database 3rd paries t applications Process model Pi Spring amp is MVC java MySQL Figure 4 9 Overall software architecture Finally there is also important to describe a communication between server and cli ent or any other cine party interested in the process Their mutual communication is al ways initiated by a client which sends an AJAX request to the server with appropriate input data wrapped in get parameters of the request Based on the way how the server handles the requests there are distinguishable two methods called short polling and long polling of which meaning is going to be explained in next paragraphs Server response is always in a form of a JavaScript Object Notation with padding JSONP which is a data format that provides a possibility to be retrieved also from a server that is located in a different web domain than the requesting client is This feature 1s known also as a by pass of the same ori
67. entiful applications the history of MST is not that long concerning from beginning only sensors and starts in 1960s when accurate hydraulic pressure sensors for aircrafts were needed However the biggest boom of MST has started in 80s side by side with enormous growth of information technologies and automotive industry Now adays as it could be seen from chart in Figure 2 4 the IT peripherals ink jet print heads RW laser heads cover most of all MST applications followed by automotive and con sumer electronics industry which has registered enormous growth in last 5 years Led by smartphones tablets notebooks and other devices equipped by microcameras posi tion light and pressure sensors the consumer electronics has become the most develop ing market areas for MST todays Computers BICT a Consumer Electronics Automotive Industrial Military Defence amp Aerospace Medical RF and Others Figure 2 4 Global MST market by application in 2011 10 billions 17 THEORETICAL BACKGROUND 21 Also the times when a car was equipped only by a few sensors is already passed by Modern car nowadays consists of multitude of sensors gathering information so essen tial that the car would never run without them The sensors continuously monitor hun dreds of vital signs such as ignition and valve timing exhaust quality throttle position intake air temperature and pressure wheel slippage tire inflation steering wheel pos
68. er 2 Deploy the wireless embedded devices to carriers and establish their wireless communication 3 Propose develop and implement algorithms that utilize artificial intelligence for real time continuous carrier monitoring and its localization within produc tion line by using event logs generated from wireless embedded devices 4 Design and develop a graphical user interface that allow user to o Design own arbitrary map of production line o Real time monitor all pallet positions o Monitor pallet position based on historical event logs stored in database INTRODUCTION 4 1 2 2 Methodology The approaches how to accomplish thesis objectives are 1 2 3 4 5 6 7 8 9 Establish wireless communication using given wireless technology deploy 6LOWPAN edge routers and set correctly the communication protocols Execute a sequence of measurements to collect raw accelerometer and gyro scope data while carrier moves in conveyor system Design and simulate event log generating algorithm based on measurement data and deploy it to wireless devices Make a literature review to find out the optimal solution for algorithm develop ment while taking into account an event based character of the process Design the key elements of overall software architecture and implement it run server create web service endpoints develop GUI for monitoring and testing purposes and develop their mutual communication Make as soon a
69. eriod from 2012 Aor 5 12 00 00 to 2012 Aor 6 12 00 00 Speed 1x TimeLine Load STOF Reset Information Vanables Figure A 11 Control Panel of Offline Monitoring Page
70. escription Language that explicitly specifies a set of available operations with described input and output parameters Mutual communication between machines and Web Services is ensured primarily by exchange of SOAP mes sages However before there will be given more detailed description of the SOAP pro tocol and WSDL there will be outlined the basic Web Service architecture in next sub chapter as a basic structure how to reach the mutual communication The theory of Web Services 1s concluded by introduction of DPWS Devices Profile for Web Services protocol as one of the most common standard tools for enabling Universal Plug and Play UPnP philosophy in world of Web Services 41 42 It might be appropriate to mention that due to the scope of this thesis the following subchapters will provide only essentials of Web Services and SOA but for those who would be interested in current state of the art of Web Services especially in factory au tomation there is a nice work related to Semantic Web Services written by professor Jose L M Lastra available here 43 THEORETICAL BACKGROUND 29 2 5 1 Web Service architecture Existence of the Web Service alone would be pointless if it would not be found by any requester It would be similar to a situation when a manufacturer produces some product only to storage it in warehouse As well as the manufacturer needs a dealer to distribute goods to customers Web Services need own Web Service broker to br
71. etermine whether the state was from overall perspective desired and the agent performed well there must be implemented some objective performance meas ure Usually the performance measure is set by an agent designer because agents alone have tendency for deluding themselves i e in case of failure they attempt to divert the reason to a fact that it was not actually their goal With all this terminology it is now possible to provide a definition a rational agent as follows THEORETICAL BACKGROUND 7 For each possible percept sequence a rational agent should select an action that is expected to maximize its performance measure given the evidence provided by the per cept sequence and whatever built in knowledge the agent has 3 In addition to above mentioned properties the agent is also characterized by an en vironment in which the agent operates All together it gives to every agent a unique set of properties called PEAS which stands for Performance Environment Actuators and Sensors Perception Figure 2 1 Schema of an Intelligent Agent 2 1 3 Agent environment In general the reason of building an intelligent agent is that the agent actually sym bolizes a solution for certain problem and the problem is task environment For instance in case of chess agent the environment will be chess board and chess opponent or in case of automatic vacuum cleaner agent the environment would be a house where the agent cleans a dust and house
72. even a control signal that 1s necessary for bias setting or for operational conditioning Actuator can serve for instance for feedback control as well as for test signals In reality these building blocks do not have to have exact bor ders and their functions can overlap or be merged 19 Input h y Trans signal Y Sensor ducer Actuator Figure 2 9 The general architecture of a sensor system 19 Electronics Output signal Input stage THEORETICAL BACKGROUND 23 2 3 Radio Frequency Identification Radio Frequency Dentification RFID is a tagging technology that is gaining widespread attention by using radio waves for information propagation and having great number of advantages compare to conventional tagging methods like for instance bar codes The conventional tagging systems usually require direct line of sight which is their most significant drawback In contrast using RFID tags allows a tag to be hidden and its information is still easily accessible RF transmission such a radio waves can transmit information across long distances and RFID is not an exception especially when they are powered In fact the way of powering divides RFID technology into 3 groups active passive or their combination Active RFID tags extend their reachability by using own source of energy and they are used for instance in pet tagging when is desired to find it in range of kilometres The passive tags rely on energy rece
73. f main conveyor and bypass and one workstation 1 that is located in very right only one main conveyor and no by pass The map was created in special web interface developed especially for this purposes and which is described in Appendix A In that interface there is a list of all available types of blocks from which an operator simply drags and drops the chosen block into a grid until like a LEGO building kit the whole map of production line is created There 1s also possibility to rotate blocks to fit the scenario 1f needed The resolution of the grid is 100 mm in both directions and the size of each block in both directions is always whole multiple of 100 mm A design of all these building blocks was done in such a manner that if a block that occupies 2 grid cells and thus it is long 200 mm it also rep resents 200mm long section of conveyor system in reality This synchronization gives the model perfect property to determine easily the pallet position if a block at which the pallet stands is known The size of 200 mm for vast majority of blocks is not a coinci dence but it was chosen deliberately The reason is that the length is exactly equal to length of pallet means that each block represents also a position which pallet can adopted if it is stopped by s stopper or by queuing line of pallets stopped in front of the pallet ee ee o Ll Figure 4 10 Example of a map of FASTory production line IMPLEMEN
74. f necessary there is also a possibility to rotate all blocks to desired orientation by clicking Rotate button If a map is completed its identification name must be inserted and the map is saved by pressing save button As soon as the map is saved to server the map designer can jump into second phase Parameter Design fact APIS eSONIN a Figure A 2 The Structure design tab of Design page a active tab b building ar ea C building blocks panel APPENDIX A USER MANUAL FOR GUI 83 Table A 1 List of building blocks Conveyor representation description STRAIGHT Straight section of conveyor RIGHT TURN Right turn of angle 90 gt d LEFT TURN Left turn of angle 90 iny B RIGHT Y Junction where pallet can continue either straight or JUNCTION turn 30 to the right j RIGHT Y Connection which a pallet can reach either by straight ty CONNECTION movement or by right turn form side E Diagonal section of conveyor section that follows N RIGHT 50 TURN BIGHT Y JUNCTION Right 90 turn differs to RIGHT TURN in size and RIGHT LONG represents longer section of turning The reason is bet TURN ter map alignment to real conveyor system Parameter Design it is the second phase of map design that is not mandatory for building of sufficient working map but it is a phase that specifies the map a bit more to allow localization algorithm to work better During Parameter Design an operator can freely change
75. fault set to be 1 9 In this way the alternatives in a direction of pallet movement forward are made more favourable The history coefficient Cp 1s coefficient which takes into consideration a fact whether an agent already visited the examined block Cp 1 2 or not cp 1 Visited means that the agent already stood at that block but it entered the block by dif ferent event than by the event that is currently examined As soon as the mechanism assigns penalties to all alternatives the algorithm chooses the block with the lowest penalty and the block become a new state of the agent Ha c eG Eq 4 1 Table 4 3 List of importance weights Importance weight IMPLEMENTATION 61 The backward search is a way how the algorithm can cope with late events In order to understand a motivation and function of this powerful mechanism two following examples are going to be shown and described Scenario 1 There are given three blocks straight block followed by left turning and immediately by right turning The agent stands on the straight block when suddenly a RIGHT event is registered According to the abovementioned formula the agent jumps across left turning directly to right turning as is shown in Figure 4 13 la However right after that a LEFT event comes The algo rithm searches at first for block in forward direction and finds a nearest block that is 4 block distant with penalty V1 4 x 1 2 x 1 4 8 Then it looks back
76. ficient revalidation for environment change e Relatively simple and extensible algorithm e All coefficients can be changed arbitrary according to given environment Method weaknesses e Dependency on coefficients from the global point of view e Dependency of RFID events as a source of initial location in case of symmet rical maps 5 4 Graphical User Interface GUI One of the goals of this thesis was also to develop graphical user interface that would allow user to monitor comfortably pallet positions to obtain necessary important data from the conveyor line and to control and interact with the process of localization itself Therefore there were developed three interfaces in total Design view Real time monitoring and Offline monitoring of which properties and functionalities will be brief ly described in next paragraphs Nevertheless in case of interest more detailed infor mation how the GUI looks like works and should be used can be found in Appendix A of this thesis e Design view the interface allows user to build own brand new conveyor map by dragging and dropping predesigned blocks The design process is very easy intuitive and fast The interface allows user also to set additional properties to various sections of the designed conveyor line such as position of RFID sensors position of stoppers or belonging to certain workstation e Real time monitoring as name prompts this interface provides real time in formatio
77. freshed updated but all new configuration data that were saved for the particular map are lost By default there is available a line called FASTory Tail slider Intuitive slider for adjusting a length of pallet tail vee The label servers as a room for displaying all important notifica Notification label tions confirmations errors etc APPENDIX A USER MANUAL FOR GUI 90 Time Eff _ Path BOND A em T Pallet number F Line maps FASTory LEFT Figure A 9 Display panel Offline Monitoring Page The Offline monitoring page illustrated in Figure A 10 is an interface that serves for retrospective purposes when a user is for some reason interested in a special event that occurred in a past For instance in a case of lost pallet the user can check at exactly which place and at what time the pallet got lost what should contribute with high proba bility to its discovery Another example can be a case of a defective product when a pallet with the product can be traced back into the history in order to check whether the pallet accomplished all necessary workstations or not The appearance of the web page is Shown in Figure A 10 The control of Offline Monitoring Page is very simple and since the monitoring area and vast majority of control area of this page is similar to Real time Monitoring page the emphasis is going to be put mainly on those adjusting elements that are specif ic for this page eS ONIA Contr
78. g as one of the most important outputs of this thesis 1s present in Appendix A e Controller controller is responsible for information flow between the view and the model It actually transforms incoming REST requests from view into model operations and sends back appropriate results Among other server re sponsibilities belong communication with database and endpoint data forward ing In addition to main Spring MVC structures there are also a two Web Service end points established in the server The endpoint serves as a main gateway for filtering all incoming Web Service SOAP messages Only desired messages are parsed and for warded in meaningful data to REST controller First endpoint is dedicated for main messages coming from wireless devices while the latter one is responsible for messages coming from line RFID data Whole software architecture is presented in Figure 4 9 In summary the overall architecture of information flow from sensor to client looks as follows Every event log that is generated by a wireless embedded device is encoded into SOAP message sent via LOWPAN network to edge router and then using Ethernet link to appropriate server endpoint The WS endpoint parses the SOAP message accord ing WSDL specification and sent decoded event log to REST controller The controller forwards the event log to a process model for further evaluation and subsequently it sends also a copy of the event log to database for storing The ev
79. g polling was deployed Short poling remained to be used solely for one shot re quests for instance to initialize a web page to check whether the server works or not or for sending user control data to server and retrieving control confirmation back On the other hand long polling was applied for all regular updates of the web page 1 e pallet actual positions and related information list of available maps as well as list of all initi ated pallets and coefficients By this change it was ensured that all web pages are al ways fresh and they do not overload a client processor IMPLEMENTATION 53 4 7 Environment model As was mentioned earlier because of limited perception capabilities of agents used in this research the model of conveyor system plays very important role in agent navi gation and presents fundamental knowledge about the surrounding environment There fore there was given a significant effort to match the reality as close as possible Since the main objective of this research is pallet localization it is not surprising that the model is related to a map The map presented in Figure 4 10 consists of 154 blocks where each block represents certain section of the conveyor system of the production line There are blocks denoting straight sections various turnings junctions and connec tions as well as places where an RFID readers or pallet stoppers are placed Concretely there can be seen ten drawing workstations each consists o
80. gin policy Server as was already mentioned can handle client s requests in two different ways and web page must be for both ways properly defined In this research a combination of both ways were implemented since each of them has its own advantages which are described in following paragraphs e Short polling it is a conventional way of gathering data by most current inter net web pages During short polling the web page regularly once per few minutes sends a request to server to get a new data and the server immediately IMPLEMENTATION a2 replies back with appropriate response The response can either contain new data that were requested by the web page or it can be also empty if no new data are currently available e Long polling it is relatively new technology which reduces amount of AJAX requests to necessary minimum In fact long polling suppresses empty respons es to be sent The mechanism is actually quite simple The client web page sends an AJAX request to a server and server either replies with desired data if they are available or waits until there will be something to be send During that time the connection is open As soon as the server obtains a new data a JSONP with data is created and sent as a response to the client In case there are more clients the response is sent to all of them List of all clients or references to their connections are stored in the server Naturally the client request does not la
81. ging b product tagging in shops c industrial tagging robust and mounting holes cs sseseeeseseeeeeeeeeeeeeees 25 Fieure 2 1LOLOWPAN arcniteCture OVELVICW meerenoniise ni e nest 21 Fieure 2 12 Service Onented Arcmitectire osongina A T 29 Pioure 2 13 SOAP Message structure corine daateaceudinansaietunsbendiaadbaantwasseudiaansasecwabenss 30 Figure 2 14 Comparison of WSDL Structures 200 0 cccccccsneeeeeeseseseeeeeeecceeeeeeeeeeeeaaaaeas 31 Figure 2 15 Architecture layout of DPWS clients devices and services 32 Figure 2 16 Comparison of DPWS and 6LOWPAN protocol stack eens 33 Figure 4 1 FASTory line with labelled workstations ccccccceseeeeeceeeeeeeeeeeeneeeees 42 Figure 4 2 Workstation with main conveyor and bypass conveyor marked 43 Figure 4 3 a Pallet b embedded wireless device c RFID tag cceeeeeeeees 44 Figure 4 4 Wireless embedded device equipped by accelerometer and gyroscope 45 Figure 4 5 Components of wireless embedded device ccccccesseeeeceeeeeeeeeeeeeeneeees 46 Figure 4 6 NFC Reader ACR 120 S B and schema of its information flow 48 Fieure 4 Overall data ransier arc ile cute eir a tad caswerieatseecceonectens 48 Figure 4 8 Edge router S1000 controller with wireless module cccceeeeeeeees 49 Fieure 4 9 Overall software architecture casni E E 5I Figure 4 10 Example of a map of FASTory production line 2 0 0 0 eececcceee
82. gnificant role of input events Every block in map is characterized besides other parameters by set of input events The input event of a block is an event that must be received in order to move to the certain block Thus for instance a right turn block has as input event an event RIGHT However it is not constraint and therefore as soon as the mod el map is loaded to system special engine is launched to discover for each block of the map its neighbours and its ancestors in order to determine what kind of events are input event for that particular block Besides already known events generated by wireless sensors or RFID sensors there is also one artificial event TIME generated by the system model itself The event TIME represents pallet free movement in the conveyor system It is generated in a case when a pallet is located in a block of which neighbour does not expect any special event and pal let just passes by Such a block has as an input event just TIME and it might be for instance a straight section of conveyor when the pallet just moves from one block to other without generating any of inertial event IMPLEMENTATION 60 When a certain event is registered called further as examined event to the algorithm the algorithm at first looks certain amount of blocks forward by de fault 80 and certain amount of blocks backwards by default 20 to look up in the map for all blocks that have input event equal to the event that is examined In
83. h In his thesis IEEE 802 15 4 protocol was used for a wireless transmission with 6LOWPAN technology on top of it Another work related to pallet localization are inter esting works of Sarah Spieker and Christof Rohrig 63 64 who aim to improve ware house management workflow by continuous position monitoring of euro pallets wood en standardized pallet used in warehouses or for goods transportation within the ware house using strength of Wi Fi signal In another work 65 that is not directly related exactly to industrial automation authors aim to develop an interesting solution for quasi continuous localization of wireless sensors with respect to long battery life In fact they propose a technique for sensor tracking while the sensor is in sleep mode to safe battery by forecasting its trajectory 2 6 2 Localization based on data of inertial sensors Localization using data from inertial sensors such as accelerometers or gyroscopes 1s nowadays almost mature discipline used primarily in navigation and localization of mo bile robots There are plenty of research works proposing different localization tech niques suitable for certain environments movement conditions sensor types or pro cessing capabilities of computational units Therefore most of the works are related just to development of new mathematical models for noise reduction and accuracy im provement Some works related to position tracking can be found for instance in 66 67 68
84. h capabilities of self learning The second sec tion attempts to get a reader familiar with theory related to used devices and communi cation technologies Thus a theory of micro system technology is presented of which sensors are used for pallet movement detection followed by data transfer technology called 6LoWPAN that is used for wireless communication between the sensors and out er world and theory of Web Services as way of data transfer and presentation In the last the third section a current state of the art is outlined to show the link between the thesis work and the present edge of research in this field supplemented by their compar isons 2 1 Artificial Intelligence Al In this chapter Artificial Intelligence is going to be introduced as a multidisciplinary research field that attempts to develop autonomous tools for solving various tasks At first a brief introduction is going to be given to get reader familiar with concept of AI The following subchapters will be dedicated to a description of intelligent agents as a core element of any AI solution together with description of environment as a place or problem the agent is dealing with Finally the chapter is concluded by two subchapters that present a minimum theoretical background of two AI disciplines that are going to be used in the further research The first is related to a decision theory as a theory de scribing a methodology for an agent how to decide well in partic
85. hat I know what happened if I had only picked this other action instead of the one I actually did I could have done better So to make their decision before the event occurs they cre ate an opportunistic loss or regret table In the table for each state of nature is utility by utility of all alternatives subtracted from the largest utility for that state of nature The loss table would look like for businessman as is shown in Table 2 2 Finally for each alternative the highest loss is taken and picked up the smallest loss out of them Hence an opportunistic businessman would rather take the umbrella to suitcase loss is 10 since he does not want to risk a bigger possible loss 15 8 Table 2 2 Opportunistic loss table for businessman example States gt It rains It does not Alternatives J rain Ombreta 0 0 Nowmbrella 5 0 Criterion realism Realist it is a compromise between optimistic and pessimistic decision Criterion realism uses for determining the right alternative by choosing the highest value of weight average WA while using coefficient of realism a The coefficient a is arbitrary estimated Each utility is weighted as shown in Eq 2 2 Maximum utility of each alternative is weighted by a mini mum utility of the alternative by 7 a and any utility value for other state of na ture is omitted 5 THEORETICAL BACKGROUND 14 WA a max Utility 1 a x min Utility Eq 2 2 Table 2 3
86. i tion and many others As one can see there are many MST sensors available but despite this research con cerns exclusively with accelerometers and gyroscopes the following chapters will be dedicated just to them and inertia sensor generally 18 19 20 2 2 2 Accelerometers Accelerometer is a device that measures acceleration proper of a test mass of well defined weight It is commonly used for motion detection as well as for space orienta tion All accelerometers work in similar way They measure a displacement as a ca pacitance change between two rows of teeth of which detail is shown in Figure 2 5 of their test mass that is mounted freely by springs and which moves according to a result ant of all forces acting on the test mass Therefore even if an accelerometer is let in gravitational field of Earth in steady position the accelerometer will never measure zero acceleration but instead the acceleration that is equal to gravitational acceleration of Earth Hence it is possible to determine sensor orientation in space In contrast there is one phenomenon if an accelerometer is in gravitational free fall and even though its speed is continuously increasing the accelerometer will measure only zero acceleration because it is at rest in a frame of reference in which object is weightless Thus accel erometers are insensitive to acceleration of gravity per se In addition in Table 2 4 there are listed the most significant acceleromete
87. ic Moreover the environment is dynamic since the agent moves and new event logs are perceived while the agent tries to distinguish the next most probable state 3 2 3 Environment model Since the environment is only partially observable the agent must make noticeably bigger effort to keep track on its movement In case of absolutely unknown environ ment the agent must discover its world alone from scratch make own map of the world and then more or less rely on it This approach requires agent to have relatively good sensors and sufficient capability to distinguish surrounding environment However from agent s designer point of view it is highly desirable to make agent s life as easier as possible Such approach contributes among others not only to computational time reduction but mainly to more accurate results One of the common solutions might be to equip agent already from beginning by the map or model of world if it is known and available This approach will be used also in this research since it 1s supposed that the layout of used conveyor system is fixed and unchangeable throughout the monitoring Because of the discrete environment in this research the developed map can be fully grid based i e the world is divided in cells of fixed size In grid map a resolution and a precision of estimated state are fixed before hand Knowing and developing of a map of the conveyor system in advance is in this re search actually a key elemen
88. igure 2 2 illustrates a working principle of probabilistic localization In itially the robot does not know where it is and thus the Bel C is uniformly dis tributed Figure 2 2a Then the robot observes a door next to it and changes its belief accordingly Figure 2 2b Subsequently robot moves a certain distance forward and as a result the belief is shifted too and flattened Figure 2 2c Fi nally the repeated observation of a door presence prompts the robot to modify its belief as is shown in Figure 2 2d The belief now corresponds to a real loca tion a b c d Figure 2 2 Working principle of probabilistic localization 15 THEORETICAL BACKGROUND 18 Monte Carlo Localization is another probabilistic localization algorithm that has been successfully used for many years in mobile robot navigation In fact MCL is very similar to Markov Localization MCL in comparison to Mar kov localization represents the probabilistic distribution as a series of samples The density of samples then represents the probability that a robot is at certain position MCL as well as Markov localization is able to solve both global local ization problem and kidnapped robot problem in very robust and efficient way In MCL large number of hypothetical states is initially randomly distributed around the space in form of particles With every object move subsequent sen sor measurement is taken and the probability of hypothetical state is upda
89. ional representation uses for value pattern description a set of relations such as better than not worse than the same as and many others Finally according individ ual preferences the best alternative is chosen Unfortunately relation representation in THEORETICAL BACKGROUND 11 not all the time unambiguous and so called cyclical preferences might appear when relations close a loop and there is no best alternative In case of businessman the repre sentation might look like as follows e Itis better to have an umbrella when is raining than not do have it e tis better to do not have an umbrella when is not raining than to have it e To have an umbrella when is not raining is not worse than to do not have an umbrella when is raining In contrary the numerical representation attempts to assign numerical values to each of alternatives The numerical values are in decision theory more appreciated since they are easy to calculate with However it is not easy to always transform every thing to exact values In case of businessman example the wet suit can be washed for certain amount of money which represents a monetary value But what kind of value has the alternative to have unnecessarily heavy suitcase with umbrella when is not raining Thus so called utility is presented to reduce all values into one single entity describing goodness in term of numbers The way how the utility will be handled falls into fifth step of decision m
90. irect way to its initial position 11 12 Dense sensor matching the algorithm attempts to match dense sensor scans against a surface map of the environment without extracting any landmark features In case of complex maps where it is very difficult to determine proper landmarks is this feature a significant advantage One of the most known dense sensor matching tech niques is scan matching The scan matching is a process of translation and rotation of a range scan ultrasonic scan laser scan in such a way to achieve the highest possible overlap between sensor scan map and a priori map of environment The result of a scan matching is a single Gaussian distribution over the environment map that can be calcu lated for instance by Kalman filters 11 Landmark based approaches these approaches use landmarks stable recog nizable features distributed throughout the environment to make estimates of position that are combined with a model of the robot s motion The strength of landmark based approaches is applicability in wide range of environments indoor or outdoor using natural or artificial landmarks On the other hand it is a bit challenging to use these ap proaches to dynamic environments with lack of landmarks and while the sensor meas urements are generally noisy There are three major and most promising approaches triangulation Markov localization Kalman filters and Monte Carlo Localization MLC which are going to b
91. ived from interrogating devices which they transform for transmission of their own information However sometimes long distance reachability is not desirable In case of sensitive data stored for instance in passports personal identification cards or credit cards a dif ferent technology called Near Field Technology NFC is used NFC is a subset of RFID technology that reduces a reach of a tag up to few centimetres The tags are usual ly passive since they do not need long reach and even shielded to insecure that only appropriate device can read its data 28 Figure 2 10 RFID tags used for a animals tagging 29 b product tagging in shops 30 c industrial tagging robust and mounting holes 31 THEORETICAL BACKGROUND 26 2 4 6LOWPAN The abbreviation 6LOWPAN refers to a set of standards for using IPv6 over low power processing limited embedded devices over low bandwidth wireless networks Although this definition might sound for somebody unclear and confusing it stands for a relatively young and simple technology for wireless communication between small electronic devices Probably the biggest impulse for development of such technology was recent increase in industrial system complexity which leads to a growing need for more process data and causing higher interest in various sensors Flexibility needs have guided a shift from hard wiring towards wireless networks of distributed sensors Side by side with new technologies there
92. lity values for all alternatives and states of nature that are going to be used for method presentation are listed in Table 2 1 in simple decision matrix 8 Table 2 1 Decision matrix with expected utili ties for case of businessman States gt It rains It does not Alternatives J rain Ombreta 5 s Nowmbrella 0 6 THEORETICAL BACKGROUND 13 Maximax Optimistic maximizes the maximum utility for every alternative it looks for the best what can happen for each alternative and then chooses the alternative with biggest utility best from the best Very similar decision as lot to player that just see the winning price and he or she is not interested in any probabilities Therefore an optimistic businessman would decide not to take the umbrella since no rain no umbrella 18 gt 15 rain umbrella 8 Maximin Pessimistic maximin agent always looks for the worst utility what could happen for each alternative and then from this set chooses the best one Hence a pessimistic businessman would take always umbrella since the worst utility for having umbrella 8 is higher than the worst utility for missing um brella 0 8 Minimax Opportunistic this kind of decision is based on opportunity loss Minimax finds always alternative that minimizes the maximum opportunity loss with each alternative It is kind of decision when if an agent looks back after the state of nature has actually occurred and says Now t
93. lligence disposes nowadays of many suitable localization decision making algorithms to determine and track position of intelligent agents This topic was more deeply described already in Chapter 2 1 from which was clearly evident that there is no universal localization algorithm that would cover all localization scenarios and the choice of the proper algorithm is matter exclusively of agent capabilities environment and overall task utility Hence to reduce the possible amount of algorithms for this re search in order to choose the most convenient one a definition of the research state in terms of AI must be done at first 1 e to ask and answer questions e how is an agent represented in this research e what is his environment e what is his overall task 3 2 1 Agent An agent in this research is going to be a virtual entity representing a real pallet in the conveyor system One agent represents always one pallet even though there might many pallets in the system introduced and therefore their number is going to be always equal to number of pallets Agent state is visualized in virtual map and in fact it at tempts to represents the real pallet location Agent s perception is entirely discrete and ensured by receiving event logs from two sources inertial sensor and RFID sensor The inertial sensor is embedded in wire less device attached to the pallet and it is going to provide information about pallet movement such as left right start
94. mply provide desired results There is actually no information about pallet movement or position in between those sensors If certain pallet stacks somewhere falls down from conveyor or it is in the worst case even taken out by irresponsible operator the conventional monitoring system would never be capable to detect it and react ap propriately Therefore a problem of continuous real time monitoring of pallets became a main motivation for this research and for writing this thesis as well Nowadays GPS Global Positioning System has become a main leader in localiza tion systems for all outdoor activities since it is a simple and reliable technology which provides relatively high accuracy However as everybody knows it is not applicable in places out of sight of satellites and therefore there is currently an extensive research for developing enough reliable and easy applicable system for indoor premises too There are many proposed ways among which two most significant approaches are distinguish INTRODUCTION 3 able First one is an orientation towards localization using wireless sensor networks signal strength time of signal flight signal angle etc while the latter one and of which also this thesis is part of is based on application of inertial sensors There is more space dedicated to both approaches in Chapter 2 6 where their advantages and disad vantages are depicted 1 1 2 Justification of the Work The research described in thi
95. n about positions of all pallets in given conveyor system It consists of two sections The prime and largest section is monitoring area where are both map of conveyor line and pallet positions graphically displayed while the second section is dedicated to localization process control The control panel contains all necessary parameters and coefficients used for localization algorithm and us er can freely adjust them to test different scenarios The graphical monitoring ar ea provides user several display modes to highlight desired information DISCUSSION ON RESULTS 74 Offline monitoring offline monitoring has similar appearance as the real time monitoring but in comparison it is capable to monitor pallet locations based on historical event logs stored in database The user can freely choose a time period in which he or she is interested and then monitor pallet locations in replay in the given time period Since it is offline monitoring the time can be speeded up or slowed down arbitrary 6 CONCLUSIONS Although the topic of object localization in arbitrary form is investigated by research groups worldwide still very intensively there is generally a lack of industrial implemen tations Therefore a goal of this thesis was to utilize current technological and methodo logical possibilities of real time localization for tracking of manufacturing carriers pal lets within real production line as a part of factory asset aware management sy
96. n iteration until n steps is reached Eq 2 1 S S U lt x w gt Eq 2 7 N N F Wnn Eq 2 8 Finally there is done normalizing of each importance weight by a division of sum of all weights Eq 2 9 and Eq 2 10 for i 1 n Eq 2 9 1 Wn y Pnni Eq 2 10 THEORETICAL BACKGROUND 20 2 2 Devices and technologies 2 2 1 Micro system technology MST Micro system technology MST or an American name Micro electromechanical systems MEMS are three dimensional 3D structures manufactured through a set of specific micromaching techniques and reaching dimensions from micrometre to nano metre scale Such dimensions allow the MST to be embedded into various electronic devices while keeping high reliability and low power consumption Due to their size the MST is usually produced in large batches which significantly contributes to their extremely low price Therefore the MST enjoys high popularity nowadays and it has been extending to all areas of human activities The application ranges from mechanical microelements needed for building microstructures and microdevices micro gears lev ers capsules etc through various microsensors that can be due to their dimensions placed almost everywhere accelerometers gyroscopes pressure sensors light sensors chemical sensors etc and end up with electrical circuits and numerous actuators need ed for micromachining and transportation of micro particles Despite pl
97. ncy of a single agent is calculated than as follows The path that certain agent drove over is stored in model maps and consists of set of blocks Each block has a certain amount of input events that the agent uses for entering the block When the agent moves the input events are registered and stored During proper syn chronous tracking for each block the agent visits all necessary input events should be received If it is not so and certain input event is missing the information is going to be stored for agent movement efficiency calculation The agent movement efficiency mechanism simply calculates for all blocks the agent visited a ratio between sum of all input events which were in reality received and sum of all events that were supposed to be received This ration is agent movement efficiency By averaging one get the overall localization efficiency that serves as etalon for fur ther calculation In order to increase localization efficiency the engine changes artifi cially coefficients and recalculate whole agent track again The new coefficients give new overall efficiency which is compared to the etalon In case the new overall efficien cy is better the actual coefficients are changed accordingly This revalidation can be done either automatically after each agent round in production line or manually by operator command For each revalidation there are examined all combinations of both coefficients Co Cp that can gain these value
98. nment There are going to be as many alternatives as possible agent s states whereas there will be only single state of nature the agent s observed perception of movement in form of event log If the event log was created then there 1s no uncer tainty or any risk that it might have happened it just happened and thus the decision is made under certainty Nevertheless the problem of the decision algorithm will be to compute appropriate utility for each alternative with given state of nature and current agent s state In summary Monte Carlo Localization and Simple Decision making algorithm will be applied for pallet localization The way how both algorithms were implemented and realized is described in next chapter while a discussion about their results can be found in Chapter 5 4 IMPLEMENTATION All research experimental tests were realized in modern laboratory called FASTlab in premises of faculty of Production Engineering of Tampere University of Technology in Finland The FASTIlab is a relatively new laboratory which can boast a newish pro duction line called FASTory The production line provides excellent conditions for var ious experiments including this research In next subchapters the production line will be introduced more in detail as well as pallets that are used for product transportation from station to station and of which localization is the objective of this research Moreover there will be introduced also embedded wireles
99. nt the algorithm demonstrated its unique robustness that was later on even more improved The men tioned robustness is based primarily on capability of discovering and dealing with late or missing events The utility estimation mechanism that works with several coefficients describing willingness of an agent to listen to certain events to move forward or back wards appeared to be very effective tool for alternative evaluation With its help the agent chooses its next state always with respect to maximization of final utility or to minimize final penalty In addition the robustness is supported by capability of self calibration that is ap plied whenever an agent loses the track of pallet to find its new correct position again Generally besides the robustness the algorithm is very flexible applicable for any kind of map of conveyor system and extensible for other features It automatically de termines the conveyor layout recognizes specific conveyor sections and discovers de pendencies of the sections In certain boundaries it can also react to changing environ ment conditions by revalidating all model coefficients and their automatic change in order to increase the localization capabilities DISCUSSION ON RESULTS 73 The observed strengths and weaknesses of the algorithm are summarized into the next two paragraphs Method strengths e Robustness dealing with late extra or missing events e Self calibration if necessary e Coef
100. nt way of transport and it could be found in many areas of industry for example automotive industry food industry chemical industry or electronics as well as in the testbed used in this research During recent years there has been increased demand for mass production of high customized products like cell phones tablets or PCB circuits generally Their manufac turing has become highly complicated task requiring sophisticated web of interconnect ed assembly lines which are capable to guide automatically a product throughout nu INTRODUCTION 2 merous workstations to produce desired final product in the most efficient way Very common solution for transportation of these products is a manufacturing carrier called an industrial pallet or simply a pallet that is carried by a belt conveyor system The pallet serves as a product carrier which provides smooth movement of product during transportation throughout whole conveyor system as well as fixed product posi tion 1f necessary The pallet is usually made from a metal frame equipped by a conven ient interface to hold product itself As soon as the product is completed it is taken away from the pallet and stored The empty pallet is subsequently sent back for new material and next production cycle starts Under normal circumstances the pallet should never leave current conveyor system Obviously the more complex conveyor system 1s developed the more difficult is his monitoring Conventional sys
101. oach was based on a technique called Monte Carlo Localization that is widely used in mobile robot localization However as it turned out the technique failed in case of pallet localization due to unsuitable environment conditions where the pallet move ment is uncontrolled and independent on measurement In contrary the second developed algorithm used for localization was a technique called Decision theory concretely simple decisions under certainty in which an agent as an intelligent virtual representation of pallet decides after receiving each event log what is the next most probable pallet position based on complex set of coefficients This technique proved to be relatively simple but on the other hand very robust The algo rithm robustness was achieved by two ways Firstly the algorithm has self calibration means that if the pallet position is lost the algorithm recognizes it and attempts to find it again and secondly a capability to cope with late missing or redundant event logs Dur CONCLUSIONS 76 ing the tests the algorithm was able to track pallets for vast majority of the time very well without any unexpected errors 6 1 Future work The research work presents one of the first attempts to localize and monitor manu facture carriers within the production line using wireless embedded devices equipped by inertial sensors Therefore the main objective of this research was to focus primarily for a feasibility of this project with giv
102. oadcast their possibilities A way how the Web Service broker is implemented is presented in typical Service Oriented Architecture in Figure 2 12 provider Figure 2 12 Service Oriented Architecture 44 If a Web Service wants to provide certain service it publishes the service to Web Service broker which serves as a manager of all published services The way how the service is published and later found is described within Web Server broker using Uni versal Description Discovery and Integration UDDI UDDI is a registry that provides searching algorithm for desired service in database of published Web Services accord ing the requirements which were sent by Web Service requester Therefore the Web Service self description is so valuable property If Web Service requester searches for a service it negotiates at first the conditions with Web Service broker who finds for him the optimal service provider and its description sends back Subsequently the requester binds with the found provider and uses its service Generally a Web Service requester can be at the same time a multiple Web Service provider 41 44 2 5 2 SOAP Abbreviation SOAP defined originally as Simple Object Access Protocol is a pro tocol that specifies an exchange of structured information in the implementation of Web Services SOAP provides basic messaging framework upon which Web Services can be build The reason is that SOAP relies always on top of other applicati
103. of Technology as part of eSONIA a European Commission cofunded research project on using service enabled embedded devices for realizing an asset aware self recovering plant The proposed solution demonstrates a novel ap proach for continuous real time pallet location tracking based on wireless sensors Il PREFACE There is end to everything and my studies are not an exception After all those long years of studying I have finally reached the point when I must stand up on my own legs and pay off the effort of those people who supported me in challenging quest of high education Right to them I would like to dedicate this page to say a word of thanks be cause without their love help and support I would never achieve to be where I am right now Firstly I would like to express my endless gratitude to my parents who worked dili gently every day to give me an opportunity to study unconcernedly who have always encouraged me towards higher goals and who have always turned every my failure into a new motivation I want to thank also you Ivusko my beloved girlfriend for your infinite patience love and comprehension during last long three years while you were waiting for me But especially I want to give you may thanks my dear because you have never lost a faith in us whenever the things didn t go right I love you I would like to thank also to Professor Jos L Martinez Lastra for his guidance dur ing developing this thesis and fo
104. ol Time period from 2012 Aor 5 12 00 00 to 2012 Aor 6 12 00 00 Speed 1x TimeLine Load STOP Reset Information Variables Time Eff Path BGND E Ton for Pallet number an Line maps FASTory r 05 April 2012 12 00 27 Figure A 10 Offline Monitoring Page APPENDIX A USER MANUAL FOR GUI 91 The entry point for this page is a Control panel shown in Figure A 11 where all functions for retrospective monitoring are located and which is shown in In a top part of the panel there are two text fields where a user chooses freely a desired time period in a past By pressing Load button the request is sent to the server which extracts from a database all stored events for the time period and makes retrospective calculation of pallet locations As a result a single JSONP object with all pallet locations and their timestamps is created and sent back to the web page Internal algorithm of the web page parses it and displays it to the user Both adjustable speed of monitoring and position in the time period are functions that should help a user to find desired target in longer time period As soon as the button START is pressed the retrospective monitoring is start ed and its virtual time is displayed in a black rectangular in a bottom of the page A user can whenever stop the monitoring by pressing STOP button Repeatable request to server for new data is allowed Control Time p
105. on layers which ensure necessary operations such a registry transport or security and therefore SOAP can be THEORETICAL BACKGROUND 30 truly lightweight compare to other protocols The most common application layer the SOAP uses is SOAP over HTTP using HTTP as an application layer but SOAP over UDP is not an exception too 45 SOAP is formatted using XML eXtensible Markup Language as a wide spread language characterized by extensibility platform independ ency self description and machine readability In the Figure 2 13 1s shown a basic structure of a SOAP message consisting of three main elements the envelope the header and the body The SOAP Envelope is a root element in every SOAP message and contains two child elements SOAP Header and SOAP Body SOAP Header is an optional sub element of the SOAP Envelope and it is used to pass application related information while SOAP body is mandatory and con tains intended information for recipient of the message 42 46 47 SOAP Envelope SOAP Header Header block Header block SOAP Body Body subelement Body subelement Figure 2 13 SOAP message structure 2 5 3 WSDL The Web Services Description Language WSDL is XML based language devel oped by W3C that fully describes interface and functionality of a Web Service in order to be used by other machines without having any preliminary information about the service Currently there are released two versions of WSDL that
106. on this junction turn a car right since on the right there is no such traffic in this hour these agents attempt to reach a specific goal for making any decision Thus all agents activities head towards the goal accomplishment while utilizing some sort of planning or search engines e g goal based agent might be a taxi that needs to utilize map navigation to find the path to some exact location Utility based agents however goal alone might not to be enough to accomplish a task with high quality output Sometimes it is necessary to add to a decision process factors such as time consumption energy consumption reliability to optimize their fi nal output That is what exactly utility based agents do they try to increase the output the utility of their work Their utility function maps a state onto a real number which describes the associated degree of high quality output 2 1 5 Decision theory Decision theory is a study of principles and algorithms for making correct decisions in order to allow arbitrary agent to achieve better outcome with respect to its goals 4 Decision theory does not attempt to be a fortune teller but rather a procedure that pro duces based on all available information the best possible logical decision and minimiz es a possibility of unfavourable outcome Thus the theory may be regarded as formaliza tion of simple common sense using mathematics as unambiguous language in which decision problem can be rep
107. onment the next agent s action is not in fluenced by its previous action Each agent perceiving and then performing is divided into single action e g defect search discovery of a new defect is not related to previ ous discoveries In contrary sequential environment is characterized by sequence of actions Current agent s state 1s a result of sequence of history actions e g chess Static vs dynamic if the environment can change in the middle of agent s next action decision then the environment is dynamic driving a car otherwise it is static battleship There is also a term semidynamic when the environment doesn t change but agent measure performance change with the time of decision like for instance in chess Discrete vs continuous to determine whether the environment is discrete or con tinuous a closer look must be done how the time is handled and how an agent perceives and acts For instance battleship is a discrete environment since it has discrete amount of states and agent s percepts and actions are discrete while in case of driving a car the perception is continuous angle of steering wheel and speed make agent s action also continuous thus the environment is continuous 2 1 4 Agent program So far agents were described in a way how they interact with their environment but this chapter introduces a way how do they behave and how do they decide based on sequence of percepts what to do The responsibl
108. ore any change was done In a Table A 4 a list of all adjustable coefficients is presented together with their description and initial values In addition in Figure A 8 there is shown a graphical rep resentation of Variables panel Table A 4 List of all adjustable process coefficients First initializing event necessary to Eve 1 RFID bring pallet to 50 of initialization process Second initializing event necessary to Events UNDEFINED bring pallet to 75 of initialization process initially skipped Third initializing event necessary to Eve 3 UNDEFINED bring pallet to 100 of initialization process initially skipped A value of forward coefficient of Co A value of backward coefficient of Backward 1 9 Coefficients Co A value of coefficient Ch represent Used Way 1 2 l ing used path or way How many block forward should Forward f Searching algorithm search size POW any block backwards should algorithm search Time in milliseconds for generating of TIME event in other words how Move time 730 long it takes to pallet to get from one Others block to another block pallet speed How many blocks backward in direc History check 5 tion of pallet should be checked for missing events APPENDIX A USER MANUAL FOR GUI 89 Information Information Information Information Control Control Control Control Vanables Vanables Vanables Vanables Events Events Events Events Eve1 RFID RFID Coeficients Coeficients Coeficients
109. ot only related to the sensor actual measured value but for instance according to a strength of wireless signal also to its position 38 39 2 5 Service Oriented Architecture and Web Services Service Oriented Architecture SOA is a set of principles and methodologies for building autonomous agile and interoperable systems In other words SOA is a frame work that allows business systems to expose their functionalities as a set of granular loosely coupled services with well defined standard compliant interfaces that can be utilized across various business platforms SOA is not new technology and in the past there were many technologies utilized to build loosely coupled architectures such as COBRA or DCOM Nowadays however the easiest and most widespread approach is usage of Web services Web Service is defined according to World Wide Web Consortium W3C as a soft ware system designed to support interoperable machine to machine interaction over a network 40 In simple words the Web Service is a service that a machine Web Service provider can provide to other machines Web Service requesters regardless of all im plementation details The service itself can represent arbitrary activity from weather forecast account statement to algorithm calculations The service interoperability is completely ensured by Web Service interface Each Web Service interface is described in a XML based machine processable for mat called WSDL Web Services D
110. outing in real time enable dynamic scheduling changes and historical traceability required for error diag nosis and repair Traditionally pallets are monitored by networks of sensors such as RFID readers or proximity sensors to collect location data These sensor networks are rarely dense enough to provide precise continuous data about pallet location Real time pallet tracking data is thus limited to recording timestamps at static checkpoints This thesis presents an asset aware management tool for continuous pallet location monitoring based on event logs obtained from intelligent wireless devices embedded in each pallet Each wireless device equipped with a 3 axis accelerometer and a 3 axis gyroscope provides accurate information about pallet movement The raw sensor data is pre processed into an event stream which is sent to a server over a 6LOWPAN net work The software developed in this research implements an algorithm for processing event logs to determine exact pallet location using artificial intelligence techniques Calculated pallet position can be provided to high level enterprise systems and to man ufacturing execution systems for use in scheduling routing and visualization of the production line Designing the SCADA system was also part of this thesis The solution was successfully deployed in the FASTory a 12 cell light assembly line in the Factory Automation Systems and Technologies Laboratory FAST lab at Tampere University
111. plays in their development a signif icant role In fact a robot is an excellent example of a physical agent that perceives en vironment with its sensors and manipulates the physical world by its effectors Localization is in robotics very intensively investigated problem since knowledge about where things are is at the core of any successful physical interaction of any robot Generally the problem of localization is in AI highly reduced to problems related to robotics Since a localization of a robot as a physical agent in physical environment has much in common with pallet localization as a topic of this thesis there is going to be given more space to it in the next subchapters for its further explanation Sebastian Thrun distinguishes in his work 10 generally three major robot localiza tion problems The simplest problem is a position tracking or sometimes called as dead reckoning where an initial robot position is known and the problem of localization 1s reducing only into error compensation of robot odometry Bit more difficult localization problem is global localization problem in which a robot does not know its initial position in advance and must find it from a scratch As soon as it is found one or more hypothetic positions are then followed Nevertheless the most challenging problem is a kidnapped robot problem in which an already well localized robot is teleported to some different place without be ing informed Hence the robot m
112. ple interface that allows obtaining data in binary format In following list more read er parameters can be found e Serial interface Baud Rate either by default 9 600 bps or 115 200 bps e USB interface for power supply e CCID like binary frame format e Read Write speed up to 424 kbps e Build in antenna for contactless tag access with card reading distance up to 50 mm e Supports NFC ISO IEC18092 tags reader card emulation peer to peer modes e Extended support for various Mifare cards e Built in anti collision system only one tag is accessed at one time IMPLEMENTATION 48 f Web Services prti NFC Figure 4 6 NFC Reader ACR 120 S B and schema of its information flow 4 5 Data transfer architecture The ability of sensors to measure pallet movement would be pointless 1f its data does not reach the appropriate interested listeners Thus it was necessary both to build a sys tem for mutual communication between wireless sensors and their listeners and to en sure their data will be secured against tapping too The overall data transfer architecture of experiment testbed is shown 1n the Figure 4 7 SCADA lA Edge P gt router Internet LAN IPv6 Backbone link i Figure 4 7 Overall data transfer architecture IMPLEMENTATION 49 Firstly there was an Extended LoWPAN network established in laboratory to ensure wireless communication with sensors nodes Thus two edge routers wer
113. process real time measurement data not delayed ones If all events had the same time delay the algorithm would work fine with just that time delay But in case of floating time delays the algorithm loses significantly an ability to track pallet at all For in stance if RIFD event which is regarded as significant and governing event comes too late the agent is pushed back close to RFID reader while in reality is the pallet somewhere else e Event to movement connection algorithm simply needs to have events related to certain movement otherwise if for instance pallet stops and event is not gener ated the algorithm will simply push agent further and lose its location e Ambiguity in case of symmetric maps there is always as many possible final results as number of axis of symmetry Thus a measurement must be supple mented by another information source such as RFID tag The Monte Carlo Localization algorithm might work well for mobile robots and similar application but in case that a movement is independent of measurement and data are obtained by event logs with fluctuating time delay this method simply fails Therefore the development based on MCL for pallet localization was in this research stopped and second approach using decision theory was used 5 3 2 Simple Decision making algorithm Simple decision making algorithm that was applied in this research showed up as a very promising solution Already in the beginning of its developme
114. que and robot is successfully located as is presented in Figure 2 3c c Figure 2 3 Example of Monte Carlo Localization a Initialization b Ambiguity due to symmetry c successful localization 16 THEORETICAL BACKGROUND 19 Monte Carlo Localization works based on particle filter algorithm and looks as follows There is a given set of particles S control vector U and measurement vector Z Eq 2 2 The set of particles consists of 1 n particles which are associated with important weights w Particle filter algorithm S U Z Eq 2 2 Now a new set of particles S is created according following steps Eq 2 3 S n 0 Eq 2 3 First of all there is iteration from i 1 n started Eq 2 4 during which a particle S from old set of particles S is picked up randomly based directly on its important weight wj for i 1 n Eq 2 4 This means that particles with higher importance are chosen more frequently Second step of the loop is a calculation of a new success state x of chosen parti cle as a state transition probability using control U and that specific particle S Eq 2 5 x p x U S Eq 2 5 Onn p zZ x Eq 2 6 Then within the same loop a new non normalized importance weight wnn 1S calculated as a measurement probability for that specific particle S Eq 2 6 A new particle with a new state x and importance weight wpn is placed into a new set of particles S and a loop continue i
115. r components Today a MST based accelerometer or gyroscope is understood as a complete prod uct that is packaged calibrated tested and delivered to the customer For customer it is required to take only minimal effort for successful integration of the sensor into his or her application Therefore the input output interface of sensor must be clearly speci fied Nowadays there are plenty sensors available on market that can meet all kind of environmental conditions and it is only up to customer which one to choose Neverthe less most of the sensors have common architecture which is shown in the Figure 2 9 and described in next paragraph The system consists of not only a sensor itself but also additional components such a transducer or electronics The role of the inner sensor is to convert input signal for instance in form of acceleration or rotation into meaningful physical objective which is transformed in transducer into an electric signal Both the sensor and the transducer of ten interact with a package since they might share the same structure and thus exchange Stress and heat too Hence one must be aware that inconvenient external conditions such a humidity or temperature might significantly affect measurement capabilities of a sensor Before the signal from transducer is processed in electronic part it is brought at first to input stage where the signal is amplified to manageable signal Sometimes the electronics is able to produce
116. r giving me an opportunity to become a member of his FAST research team while working on project eSONIA that founded my work Special thanks belong also to Corina Postelnicu and Johannes for her advices and corrections during writing this thesis Finally my thanks belong to all my friends who made my second life here in Finland more pleasant happier and in fact unforgettable It was amazing time Thank you Navid Hector Luis Payman Ale and many other Czech friends who I met here My studies are over but educating is never ending process and I will never stop Av ak ze v eho nejd v chci uz kone n dom Tampere September 2012 Tom Sedl ek 1 MI CONTENTS TA OIO O ae E E ah Macauniensenden tues l 1 1 Back o round a a E O l ELA Proben SCAU IG In Ue steeper ec oneie ba opie bere teliewue te need iodeate 2 1 1 2 Jost hncation OF THE WorKocsserisnearini irina a 3 L2 Work DescHpUO isca E EE 3 Eor ODEIO a a ke ea ne INR eee te PE 3 22 NVC OC ONG Cy ie trices aie erred ec ccd gee ees eee nee 4 ES Thess OUNDO irssatentieeicatidiredeatie E N 4 Theoretical Dacko round esenea E miebsannceinsewiaueaaatsananst 5 2 1 Aruncial Ite MIS CN COCA inna E E TA 5 2 1 1 Definition and introduction to AL ccccccccccceeesseseeeeeeeeeeeeeeeeeees 5 242 Tnteligent Abeni sei coctiaveascterntcivdinstannctedicertiivnaceursiniebwnntadasutadials 6 23 Apcnteny rone Nies a S 7 Dkk ACEM POC 2 eee RO A re On RO eR rT SR a 8 DMO AVCCISION ANCOL
117. r properties that characterize every p lt accelerometer 21 22 Figure 2 5 Detail of accelerometer structure 20um thick single crystal comb struc ture used for capacitance measurement 22 THEORETICAL BACKGROUND 22 Table 2 4 Accelerometer properties Measurement range of each accelerometer is given in multiples of g Range so for instance a 2g accelerometer is capable to measure acceleration in range of 2 g 20m s to 2 g 20m s Ability to measure acceleration in certain direction the simplest are Number of one axis 1D accelerometers having ability to detect either X Y or Z axes axis then two axis 2D accelerometers or the newest three axis 3D accelerometers that are able to detect all three axes at the same time Sensitivity A minimal change in acceleration which is detected by accelerometer 2 2 3 Gyroscopes Gyroscope is a sensor or a device that is capable to detect angular acceleration and therefore is widely used for detection of a change in sensor orientation in space or rarely to determine its absolute orientation in space There are many kinds of gyroscopes dis tinguished by their working principle Although the chapter refers solely to MST gyro scopes it might be appropriate to explain a working mechanism of a typical mechanical gyroscope too due to its long evolution and easy construction which describes the working principle of gyroscope in the best way e Traditional mechanical g
118. resented Decision theory is very wide spread discipline that contains many methods and algo rithms but its full description is unfortunately far out of the scope of this thesis In this chapter there are going to be presented two major decision theory branches Simple Decisions and more advanced Sequential Decisions Where the main focus will be giv en primarily to Simple decisions since there are going to be later on used in algorithm of this thesis THEORETICAL BACKGROUND 10 A Simple Decisions Sometimes there are known as a Single shot decisions or a One shot decision which expresses the actual meaning even more precisely Simple Decisions are used in episod ic environment where the focus is given solely to utility of one single action and the utility is known Therefore an agent decides always in such a manner that it does the best what is possible in that moment regardless if it is good also in long term perspec tive According to 5 there are six major steps which should be followed in order to make a successful simple decision and of which meaning will be describe in next para graphs 6 1 Clearly define the problem 2 List all possible alternatives 3 Identify the possible outcomes 4 List expected utilities for all combinations of alternatives vs states of nature 5 Select one of the mathematical decision theory models 6 Apply the model and MAKE THE DECISION The easiest way how to describe working principle of
119. rvice and all operations that its endpoints provide Interface together with necessary messages Defines SOAP actions and message encoding as well as the input Operation and output parameters for the operation Describes the data in form of XML Schema THEORETICAL BACKGROUND 32 2 5 5 DPWS Device Profile for Web Services DPWS is a protocol introduced by OASIS 50 that defines a minimal set of implementation constraints to enable secure WS messag ing easy and dynamic discovery service description and eventing on resource constrained devices DPWS makes use of existing WS protocols and adds several exten sions or constrains that are called DPWS specifications to enable Web Service based communication in embedded devices Application of DPWS utilizes Universal Plug and Play philosophy UPnP where DPWS compliant devices are after introducing to a net work automatically discovered described and immediately ready to use 42 51 52 As can be seen in Figure 2 15 from a SOA perspective the DPWS compliant device is an entity that hosts other services The service is a software system that exposes its functionalities to outer world by interchange of messages The messages utilize always SOAP protocol for transmission concretely either SOAP over HTTP or SOAP over UDP The device plays primarily a role of resource of metadata about hosted services The hosted services are externally visible to clients and on top of that the DPWS proto col sp
120. s Co 0 2 0 1 0 0 1 0 2 co Ch 0 2 1 0 0 1 0 2 Ch This in total gives 25 possible combinations from which the best one is chosen In this way the algorithm can dynamically update its own parameters to adapt for chang ing environment or simply to optimize localization process 5 DISCUSSION OF RESULTS The outcome of this thesis was to develop a working prototype of real time monitor ing system for localization of manufacturing carriers using data from inertial MEMS sensors while utilizing techniques of artificial intelligence The aim of this chapter is to present obtained results of all implemented methods supplemented by conclusions and discussions related to their implementation suitability and reliability 5 1 Signal processing and event detection As was mentioned already in research methodologies the first step of implementa tion of signal pattern matching was to study thoroughly the raw sensor data for determi nation of patterns corresponding to significant pallet movements Before further expla nation how the pattern matching was done it 1s necessary to introduce a reader the co ordinate system at first according which sensor axes are labelled and which is shown in Figure 5 1 Z axis Y axis Figure 5 1 Accelerometer and gyroscope axes In total there were distinguished five significant pallet movements that are both fea sible to be detected by sensors and useful for purposes
121. s devices which are equipped by accel erometers and gyroscopes and which provide the main source of data for this research 4 1 Production line 4 1 1 Description The abovementioned production line which is shown in Figure 4 1 is currently locat ed in FASTIlab in Tampere University of Technology but a just few years ago it served as a final assembly line for electronic products in one Finnish company When their production stopped they transported the line to FASTlab and it underwent some mod ernization The most significant change was unambiguously a line transformation for research activities Therefore electronic assembling was replaced by simpler activity of cell phone drawing and so the line is not called assembly line anymore but just produc tion line The reasons for transformation are obvious there is no need to bother with buying raw material as well as with storing assembled electronic devices Currently the production line provides excellent environment for laboratory experiments where the newest automation technologies can be applied and immediately tested At first glance the production line seems to consist of twelve almost identical work stations WSs mutually connected by a conveyor system However in detail look one can distinguish that the first and seventh workstation significantly differs from others While the majority of workstations WS2 WS3 WS4 WS5 WS6 WS8 WS9 WS10 WS11 and WS12 serve for drawing operations
122. s possible first algorithm prototype and verify its properties in order to keep agile process Based on the first results either continue or revalidate used algorithm repeat this test validation approach as many times as possible If necessary apply dif ferent algorithm Verify the results Polish the graphical user interface 1 3 Thesis Outline The content of this thesis is arranged as follows Chapter 2 1s dedicated to provide reader necessary theoretical background to understand the fundamentals of methodolo gies and technologies that are used throughout this thesis Chapter 3 describes chosen methodology that is going to be used for developing the final solution and outlines the process of its implementation The implementation proper is going to be explained in Chapter 4 while results and their discussion can be found in Chapter 5 The thesis 1s concluded by Chapter 6 that is dedicated to final conclusions of the thesis 2 THEORETICAL BACKGROUND The aim of this chapter is to present a necessary theoretical background and litera ture review for reader to understand the content of this thesis and used technology Be cause of large diversity of used technologies the chapter is attempted to be arranged in three sections First section of this chapter is dedicated solely to artificial intelligence as a research field that utilizes methods for autonomous solution of relatively difficult tasks without human interaction and wit
123. s therefore workers are being replaced by faster and more precise robots a single modern production machine is now capable to provide many operations at one grip and overall control has become highly automated however the Ford s main concept of assembly line has remained until nowadays almost the same Whether it was breakthrough Ford s manual line semi automatic mass pro duction line in 80 ties of last century or modern automatic high customized lines all of them have one thing in common Logistics The raw material various components and product itself must be simply transported somehow from one machine to other machine from one workstation to another from factory to customer and that all happen in exact time place and volume The more so phisticated product the more components is needed more operations must be done and more complex transport system has to be developed In fact there are two possibilities either the product is localized due to its size or specific features at one place and materi al is brought to that place for example aircraft industry or ship manufacturing or the machines components and other resources are localized at certain places and the prod uct itself is transported from one station to another In this latter scenario the products usually move for instance on rails rollers belts they hang from ceiling transport system or they move in tubes or on air cushion Due to their size it is an easy and efficie
124. s thesis was written for a Department of Production En gineering in Tampere University of Technology TUT as a part of European project under the umbrella of ARTEMIS JU called Embedded Service Oriented Monitoring Diagnostics and Control Towards the Asset aware and Self recovery Factory eSON IA From the project name is already evident that the main objectives of eSONIA pro ject will head towards development of tools for asset awareness and self recovery in production plants using embedded devices Therefore also this research aims to develop expedient tools for real time positioning and monitoring of production assets in form of manufacturing carriers transporting products by implementation of low cost and low powered wireless embedded devices Wide interoperability of the final solution is ensured by using 6LOWPAN as a wireless communication technology that supports direct network connection by IPv6 without a need for any gateways and versatile web services as a way of data transportation To gether with other research activities within TUT eSONIA research group there is going to be build a prototype for Asset aware Management Tool for complete monitoring of a production line 1 2 Work Description 1 2 1 Objectives The primary objectives of this thesis are as follows 1 Design and implement to wireless embedded devices an algorithm which will generate appropriate event logs based on MEMS data describing movement of manufacturing carri
125. ssfully to next state or even moves two states more e Ifa STOP event is detected no forward move control signal is going to be generated anymore and program waits until event GO or any other event comes e Probabilities of various measurements incoming events are set according to a reliability of the events to be correctly generated It means that for instance turn ing events copy with high reliability the actual turning of the pallet They are rarely missed or be redundant In contrast to events like STOP and GO which are hard to detect or be misplaced and their measurement does not have to nec essarily mean that the pallet actually did the action e All probabilities were firstly manually chosen to fit into given model and pre pared to be easily automatically tuned to increase overall localization efficiency The implementation of MCL algorithm for pallet localization is explained in simpli fied form in Figure 4 11 The probabilities used for this example are following Probability that a pallet is on right turn block and a right turn was detected is 0 8 Prr Rs Rm 0 8 Probability that the pallet is not on right turn block and a right turn was detected is 0 2 Pyr NRs Rm 0 2 Probabilities related to right turn are equal to left turn as well a In first step a map is given with characteristic path profile The map consists of 8 blocks five RIGHT turns one LEFT turn and three STRAIGHT blocks b Initially the sam
126. st forever and there is always some timeout usually it varies from 10s to mi nute It means that if no response comes no new data are available on server side till the timeout the connection is closed and new request is send to the server again This approach has two significant advantages Firstly there is no delay on client side for receiving new data since long polling is fully event driv en mechanism Thus there is no need for client to regularly beg for a new data anymore as in short polling but the new data are immediately sent when they are simply generated Secondly all clients receive the same data at the same moment what is much appreciated featured in some kind of branches e g stock market information Finally there is no server overloading anymore Short polling as a mature way of data retrieval was also in the beginning applied in this research with only one significant change higher request frequency 20 requests per second The frequency was obviously chosen very high to let the monitoring to be truly real time Although it worked very well in the beginning very soon a significant drawback of short polling was discovered The problem was that amount of requests was directly proportional to amount of active web pages which led on client side to increased processor usage if more web pages were open at the same time and on serv er side it leads to overloading by processing of high amount of requests Therefore a lon
127. stem In summary each pallet in the production line was equipped by wireless embedded device with accelerometer and gyroscope providing fundamental data about pallet movement Usually the sensor data are sent straight ahead for further processing How ever in this research according to the latest trends in automation the raw sensor data were at first preprocessed right in the wireless device in order to identify only the mean ingful pallet movements If the meaningful movement occurs e g pallet turns then only a single event log carrying information of the movement is published via 6LOWPAN network to subscribers Such event log can be used by any client without having knowledge how the movement was determined The development of the algo rithm that provides signal preprocessing was also a part of the thesis and it was based on technique called pattern matching since every movement has its unique signal pattern The generated events from wireless devices supplemented by events from RFID readers represent the central data source for pallet localization algorithm The algorithm is running on the server and has web interface through which a user can monitor pallet positions adjust important algorithm settings or event to build from scratch a new con veyor map In this research two different approaches were utilized to develop the pallet localiza tion algorithm and both were implementing fundamentals of artificial intelligence The first appr
128. t Ojbect Notation M2M Machine to Machine MCL Monte Carlo Localization MDP Markov Decision Process MEMS Micro Electro Mechanical Systems MEU Maximizing Expected Utility MST Micro System Technology MVC Model View Controller ND Neighbour Discovery NFC Near Field Communication OEM PCB REST RFID RPC RSS RSSI RTU RW S 1000 SCADA SCARA SLAM SOA SOAP TCP TDOA TUT UDDI UDP UPnP URL USB W3C WE net WLAN WS WSDL WSN XML XSL IX Original Equipment Manufacturer Print Circuit Board Representational State Transfer Radio Frequency Identification Remote Procedure Call Received Signal Strength Received Signal Strength Indicator Remote Terminal Unit Read Write Product name of used Remote Terminal Unit Supervisory Control and Data Acquisition Selective Compliant Assembly Robot Arm Simultaneous Localization and Mapping Service Oriented Architecture Simple Object Access Protocol Transmission Control Protocol Time Difference Of Arrival Tampere University of Technology Universal Description Discovery and Integration User Datagram Protocol Universal Plug and Play Uniform Resource Locator Universal Serial Bus World Wide Web Consorcium WorkFlow net Wireless Local Area Network Web Service Web Service Definition Language Wireless Sensor Network eXtensible Markup Language eXtensible Stylesheet Language TABLE OF SYMBOLS AND UNITS PyR NR IR PRR Rs Rm Nos SS Se as Ta Orienta
129. t because agent suffers by a lack of convenient sensors With discrete event logs containing solely information of a change in agent movement in terms of left or right are absolutely insufficient for localization in unknown environ ment 3 2 4 Appropriate algorithm selection Based on abovementioned assumptions there were chosen finally two most promis ing AI algorithms that would be capable to localize pallet within the conveyor system and that would fit to research conditions too Next paragraphs are going to attempt to explain the major reasons why each of particular algorithms was used The first algorithm was chosen from localization algorithms In AI as is described in Chapter 2 1 6 there are available three localization approaches behaviour based dense sensor matching and landmark based approach While behaviour based approach is absolutely unsuitable for conditions of this research since an agent can never under stand the world around him based solely on the events used in this research the scan matching could be theoretically used In this case an operator would need to teach every agent which kind of event sequence corresponds to each section of conveyor system and the agent would then tries to match the sequence with learned scans However since the conveyor system can consist from several identical conveyor sections e g workstations RESEARCH METHODS AND MATERIALS 40 the algorithm would be confused Drawback is also
130. ted The higher density of particles in certain state space the higher probability the state represents the actual robot state Because the total number of particles must be always constant and most probable states consume the biggest amount of particles it logically means that the improbable states have low or no particles at all Although this property gives MCL very high robustness it is also its main drawback If there are no particles left in certain state there is no chance at all for the algorithm to revalidate this state for potential object occurrence This might be problem for instance in case of solution of kidnapping robot problem One of the most common reasons is a usage of very precise sensors that do not give any uncertainty to robot movement and thus spread more the probability distribution The simplest solutions might be to add to a sensor signal an artifi cial noise 10 The Figure 2 3 shows very nice graphical visualization of working principle of MCL for global localization problem In Figure 2 3a an initial phase of MCL is shown and particle distribution is equally spread around the space As soon as the robot makes first measurement the particle distribution is immediately up dated and its result can be seen in Figure 2 3b It is clearly visible that particles gathered mostly in two positions due to ambiguity in geometric properties of ro bot environment However as robot proceeds the environment starts to be uni
131. ted Monetary Value EMV Expected utility is also sometimes called more precisely as a probability weighted utility theory that assigns to each alternative i a weighted average of its utility values under different states of nature j weighted by probabilities of those states p j as is shown in Eq 2 1 j n EU utility pQ Eq 2 1 jar The final decision is done based on the principle of maximizing expected utili ty MEU that says that a rational agent should always choose an action that maximizes the agents expected utility and thus it is chosen alternative with the highest expected utility The principle of EU is shown in next paragraph on the case of businessman If for instance it 1s supposed that a probability of a rain is 0 1 and not rain 0 9 then the expected utility of bringing the umbrella is 0 1 x 15 0 9 x 15 15 and for not bringing the umbrella 0 1 0 0 9 x 15 16 2 Therefore rational businessman would let his umbrella at home and undergo the risk to be wet Decision under Uncertainty Decision under uncertainty has available utilities for each combination alternative vs state of nature but there are no probabilities of state occurrence and thus optimization criteria cannot be applied Therefore there is used one of these four decision making models Maximax Maximin Minimax or Criterion Realism The working principle of all of them is going to be presented again on businessman example Uti
132. ted in Figure 5 2b DISCUSSION ON RESULTS 66 deg s2 LEFT RIGHT aj t mis GO STOP EXIT b t Figure 5 2 Pattern matching a Z axis of gyroscope b Z axis of accelerometer Table 5 1 Event description Sinusoidal wave RIGHT Pallet turns right Gyroscope z axis High opposite phase Accelerometer z Significant noise STOP Pallet stops Medium axis reduction Pallet start mov Accelerometer z Noise peak in ing again axis creased Pallet leaves Accelerometer z l EXIT Sinusoidal wave High conveyor axis The algorithm was uploaded to chips of wireless sensors that scan continuously the sensor data output for presence of any of these patterns As soon as the event is detect ed the event log is created and sent to all subscribers In summary implemented pattern matching algorithm was proved to be a sufficient way for event detection with respect to given limited resources and for purposes of pal let localization However its reliability strongly depends on the environment in which it is used and must be tuned for each particular application The example of real data out put from Z axis accelerometer with described pallet behaviour and marked points where an event was registered is shown in Figure 5 3 DISCUSSION ON RESULTS 67 2040 2036 a 1840 2024 2040 2040 i iio Z axis accelerometer a 612 ikea 1496 if Ain POO 6p 62 nee 5 jai jii 55246681548 Sah Pii i3 iti i PHM HO
133. tems use for monitor ing usually a set of sensors capacitive inductive or RFID sensors distributed around the line to detect pallet presence and thus provide basic information about current pallet position From the process control point of view is this system relatively sufficient till the pallet control operations changing directions entering workstations stopping etc take place exclusively at sensor places However every product that is often held by even more valuable pallet presents for a company a significant asset that needs to be monitored continuously throughout the process to prevent potential losses or damages Such a monitoring tool is called asset aware management tool and development of its techniques is currently under intensive research Some works related to this topic are listed in Chapter 2 6 which is dedicated to the state of the art of localization and moni toring of various assets 1 1 1 Problem Statement Mainly during the last years in the time when economic crises came the vendors tended to optimize their processes and production lines to found all possible faults in congruities and ineffective procedures like never before Having numerous pallets with products on top of them represents for each vendor considerable assets and thus their continuous monitoring to prevent damages or losses become one of their main objec tives Unfortunately traditional monitoring based on a network of distributed sensor cannot si
134. th known positions THEORETICAL BACKGROUND 35 e Received signal strength RSS measurements RSS measurements estimate the distance between neighbouring sensors from the signal strength measurements This technique is bases on standard feature found in majority of modern wireless sensors a received signal strength indicator RSSI This method is very attractive since it does not require any special hardware thus the sensor can be smaller lighter consumes less energy and is more cost effective Drawback of this method is increased sensitivity to disturbances such as reflection diffraction and multi path effects This method is mainly used in location estimation in WLANs but in recent years it has found a way also in localization of wireless sensor networks Related works can be found in 61 62 In summary there are many available techniques how to determine position within wireless sensor network and some of them were presented in this chapter In reality the localization algorithms based on AOA and propagation time measurements have better results and provide higher accuracy than algorithms based on RSS However this higher accuracy is at expenses of higher investment costs Probably one of the most extensive works related to localization in industrial envi ronment is a thesis work of Jaacan N M Valdez 38 where the author tries to imple ment real time localization of production pallets populated by wireless sensor by signal strengt
135. the decision theory concretely of simple decisions is to show it on a simple example concretely on case of busi nessman dilemma which is as follows There is a businessman who is going in the morning to the work and wearing a per fect suit However he has a dilemma whether to take an umbrella with him or not If he takes the umbrella and it will not rain then the umbrella will represent whole day bur den for him On the other hand if he let the umbrella at home and it will be raining his suit gets wet and it needs to be cleaned what costs him money and time In this example there are two decision alternatives having the umbrella not hav ing umbrella that a businessman as a decision maker can perform and two states of nature raining not raining that represent uncontrollable future events Generally alternatives can be either open if the range of choices can possibly expand business man example he can take also a car for instance or closed when there is limited list of available choices e g yes no questions A decision is naturally attempted to be done in rational way thus in such a way that the resulting output is always maximized e g to do not unnecessarily carry heavy suit case or pay for washing The question however arises how to express a value pattern for each alternative i e how to assign a goodness to its future output There are two most common ways relational and numerical representation Relat
136. the need of a process of learning what makes this approach in summary less preferable More promising is landmark based approach especially Markov Localization ML and Monte Carlo Localization MCL In principle both are quite similar only ML uses Gaussian distribution whereas MCL discrete distribution of particles Because the environment is grid based the MCL is preferred and used as the first final algorithm for this thesis As the second algorithm was chosen the algorithm that implements fundamentals from the decision theory The main idea is that agent that is representing a pallet will decide about its next new state a position right after it receives an event log Generally the decision theory offers two possibilities simple decision and sequence decision Sequence decision is not suitable for pallet localization since the environment is episod ic and there is no final goal to which a concrete sequence of actions would lead On the contrary the environment requires each decision to be one shot decision which attempts to increase momentary agent s utility after each incoming event log To develop proper algorithm it is necessary to determine whether the decisions are made with certainty with risk or under uncertainty what is represented by alternatives and state of nature and finally to build the appropriate model accordingly The alternative for each decision will be an agent s new state an exact position in grid based enviro
137. the position on side of workstation but on the other hand middle position is very impractical place and thus the edge router was finally placed down the conveyor line and on side of one workstation From the results showed in Figure 5 5 is clear that the amount of messages for both devices were during all experiments almost equal which proved that the measurement was correct but by closer look to individual experiments Figure 5 6 and Figure 5 7 one can notice very interesting thing The amount of received messages for both devices varies for all the experiments over the time absolutely in the same way regardless their positions Next important discovery came out from static tests Experiments B and D in Figure 5 6 and 5 7 where the pallets are not running in conveyor system but standing in Static positions one in close distance above edge router and the second one in the most DISCUSSION ON RESULTS 70 distant place of the conveyor system It was found out that amount of received messag es is for proximal device almost the same as for the distant one and therefore if the edge router is placed in the middle of the line the position of pallets does not play any signif icant role Experiment B E Proximal f E Distal N eb ete 40 N I E 30 O D Q 2 m e O _ E E A ae SS En ee ee San ae ee a ee a as Ea ee ae E EE 123 45 67 8 9 10111213 14 15 16 17 18 19 20 Minutes
138. tion coefficient History coefficient Distance in blocks Total number of all particles Normalizing factor Probability that the pallet is not on right turn block and a right turn was detected Probability that a pallet is on right turn block and a right turn was detected Old set of old particles New set of particles Non normalized importance weight Normalized importance weight Control vector Penalty New state Measurement vector 1 INTRODUCTION 1 1 Background When a Scottish philosopher Adam Smith more than 200 years ago published his famous work The Wealth of Nations about division of labour and resources he would probably never believe that his idea will completely change the face of modern industry However it took more than one century until the idea of resource division was for a first time fully transformed into reality by Henry Ford In 1913 he managed to develop a first assembly line for a mass production of legendary car model Ford T Both labour and resource division significantly improved the efficiency and volume of production Most of the machines turn to be used in their full capacity and moreover there was no need for workers to know how to build a complete car anymore They just executed a few operations related to his or her skills and shift the product further in process chain 1 2 Many things has changed since that time new technologies were discovered there was a revolution in electronics and informatic
139. ular situation and localization as a technique for finding agent s position both in known and unknown environment and which is used mostly in branch of mobile robots 2 1 1 Definition and introduction to Al From time immemorial the human mankind have tried to understand a process of how humans think i e how people can perceive understand predict and manipulate things surrounding them in order to define the key element which distinguishes people from animals However the attempt of Artificial Intelligence goes even further it does THEORETICAL BACKGROUND 6 not only try to understand the process itself but also it tries to build new intelligent enti ties to copy human intelligence 3 The history of AI is relatively new and goes back to 50 s of last century along with a birth of information technology Since that time the AI has matured a lot and nowadays encompasses many research disciplines However the more branches exist the more perspectives on AI objectives arose and the more difficult is to determine unified proper AI definition now Nevertheless as is mentioned in 3 all AI definitions generally vary along two main dimensions e Thought process and reasoning it is a human centred approach that focus to measure success in terms of fidelity to human performance by implement ing empirical science 1 e hypothesis and experimental confirmation e Rationality rationalist approach which involves a combination of mathe
140. ularity not only in academic circles Small dimen sions low costs long battery life simple communication protocols and other character istics of modern wireless devices directly predestine their wide application range and deployment in great amounts However with rising sensor volume there is also a rising need for their localization and management Currently there are available many pro THEORETICAL BACKGROUND 34 posed algorithms of sensor localization based on different technologies but their de tailed description is far out of the scope of this thesis Therefore in next paragraphs only a brief introduction with useful references to most known techniques is going to be listed to outline possible opportunities More extensive summarization of available WSN localization methods can be found for instance in 55 56 57 58 59 e Angle of arrival AOA measurements AOA is characterized by propagating signal angle measurement The receiver antenna usually of very small dimensions is rotated electronically or mechanically while it receives the sensor signal The signal strength is evaluated either by phase shift or by measuring signal amplitude The angle from which the strongest signal is detected corresponds directly to the angle the sensor is located towards receiver Naturally there are more receivers installed around and the final location is calculat ed based on their relative measurements It is an easy method that nevertheless re
141. ust 0 69 parti cles Ni On Np In next step a forward movement is applied blue arrows and left turn is detect ed For simplicity of this example a probability of movement is set to be equal to 1 This means that all particles will move with 100 probability to the very next block The algorithm for recalculation of particle amount goes in the same way as in the step c Since there is only one left turn block in whole map it is clearly seen that algo rithm in just two steps easily and unambiguously distinguished pallet location with quite reasonable probability a block where an arrow is point at IMPLEMENTATION 51 RIGHT TURN Figure 4 11 MCL algorithm for pallet localization In addition in real experiment the map is more complex and includes blocks such as junctions where pallet moves either straight or right or sinks where two conveyors are joining together The algorithm was thus supplemented by appropriate probabilities solving all possible scenarios and by ways how to deal with particle splitting and merg ing on these blocks Also uncertainty of movement was added 4 8 2 Simple decision making algorithm As was already discussed in the Chapter 3 the second algorithm that is going to be developed for pallet localization is going to utilize decision theory concretely funda mentals of simple decision making under certainty The chapter also defines an agent as virtual representation of a pallet and the developed
142. ust not blindly rely on its hypothetically found place at the time of kidnapping but be still able to appropriately react and check other location possibilities too This kind of problem was intensively investigated in last decade and many new methodologies were elaborated such as Monte Carlo Localization SLAM Simultaneous Localization and Mapping or Markov Localization known as Kalman Filters In next subchapters will be some of them presented in more details According to 11 localization techniques fall into three basic categories e Behaviour based approaches e Dense sensor matching e Landmark based approaches THEORETICAL BACKGROUND 16 Behaviour based approaches behavioural approaches rely on the physical inter action of robot actions with the environment to navigate Behaviour based localization does not rely on any internal map as a central representation of world but the world is a map itself The computational path from sensor to actuator 1s minimized to maintain high interactivity with the physical world The robot s activities produce behaviours that are layered on top of each other in such a way that robot localization capabilities are increasing Behaviour based approach is very similar to behaviour of insects and in fact is commonly used for localization of small robots with very limited computational ca pabilities indeed For instance behaviour based robot can replay a sequence of its movements to find out the d
143. vice is powered by two AA batteries and can work for long time due to its low power consumption The chip properties allow using a 6LOWPAN protocol for a wireless communication with outer world that is described more in Chapter 2 3 Figure 4 4 Wireless embedded device equipped by accelerometer and gyroscope IMPLEMENTATION 46 Batteries Reset button Config Run slider Chip aa Control i D2 i LEDs Me E eee aS Antennas AMET Nf Fec JTAG programming interface 3 axis accelerometer 3 axis gyroscope Figure 4 5 Components of wireless embedded device Due to native characteristics of 6LOoWPAN where an emphasis is on low power and low processing operations of executing device there is used exclusively UDP User Datagram Protocol for a transportation layer of communication However in this re search there was an attempt to provide sensor information as a web service using Device Profile for Web Services DPWS Therefore data are packed into messages according to the following standards WSDL Web Service Descriptive Language for describing functionality of web service SOAP Simple Object Access Protocol for specification of exchanged information and XSL schema as a set of rules according which a message is coded A message carrying certain information is then in a form of web service sent using HTTP and TCP transport protocols Nevertheless the change in transportation protocol caused significant increase in da ta overheads
144. wn as well Each state is determined also by a reward function describing how good the state is to reach the final desired state Desired state has naturally the highest reward Transition model or reward functions are usually invariable but the point is to find the optimal policy as a function that tells agent how to act on every state to yield the highest expected utility 3 Partially Observable Markov Decision Process POMDP while MDP was applied to fully observable environment POMDP are used if there is an uncertainty whether the agent actually is on certain state and if the policy for supposed state will be correct Thus POMDP works instead of agent s actual state with agent s be lief state as a state that with the highest probability represents agent state 3 THEORETICAL BACKGROUND 15 Game theory A multiagent decision model of optimality takes into consideration not only maximized utility but also the interaction between agents who are playing against each other Game theory attempts to look at the relationships between agents in a particular model and predict their optimal decisions 9 2 1 6 Localization Everybody knows that robots have become in recent years inseparable part of indus trial production and human lives almost too Step by step they increase their capabili ties become faster smarter efficient and thus slowly take over human daily routine work Therefore nobody would be surprised that AI
145. yroscope The principle of mechanical gyroscope is known for a long time although it did not serve for measurement but for amusement and religious activities The gyroscope consists of flywheel spinning in set of gimbals which can move around independently in respect to the flywheel as can be seen in Figure 2 6 If the flywheel is spun the angular momentum and precession movement causes the flywheel to be highly resistant against any forces that act to change its axis orientation In simple terms if outer gimbal is twisted the flywheel still keeps the same orientation Therefore mechanical gyroscopes have been used for a long time in marine and aerial navigation 19 23 Figure 2 6 Traditional mechanical gyroscope 23 THEORETICAL BACKGROUND 23 Coriolis vibratory gyroscope CVG or Vibrating structure gyroscope This gyroscope has become the most common type of MST gyroscope nowadays mainly due to easy construction and very low price It does not im plement any spinning flywheel as in traditional mechanical gyroscope but on the other hand it uses a vibrating structure a mass flexibly mounted at least in two directions by springs or anchors The structure oscillates in one direction as is illustrated in Figure 2 7 When the gyroscope is rotated by outer force Coriolis force starts acting on the vibrating structure in such a way that it shifts the structure perpendicularly to axis of vibration This radial displacement is then measured as

tomas sedlacek a real-time positioning system of manufacturing

Contents

Download Pdf Manuals

Related Search

Related Contents