Home

A Sensor System for Unsupervised Residential Power Usage

1. 1 U S Energy Infrmation Administration Residential energy consumption survey 2011 http www eia gov 2 A H McMakin E L Malone and R E Lundgren Motivat ing residents to conserve energy without financial incentives Environmental and Behavior Journal vol 34 2002 3 P3 International Corp P4400 Kill A Watt Operation Manual 4 Watts Up https www wattsupmeters com 5 X Jiang S Dawson Haggerty P Dutta and D Culler Design and implementation of a high fidelity ac metering network in JPSN 2009 6 G W Hart Nonintrusive appliance load monitoring Pro ceedings of IEEE vol 80 1992 7 S Patel T Robertson J Kientz M Reynolds and G Abowd At the flick of a switch Detecting and classifying unique electrical events on the residential power line in UbiComp 2007 8 S Gupta M S Reynolds and S N Patel Electrisense single point sensing using emi for electrical event detection and classification in the home in UbiComp 2010 9 Y Kim T Schmid Z Charbiwala and M Srivastava Viridiscope design and implementation of a fine grained power monitoring system for homes in UbiComp 2009 10 Y Kim T Schmid M Srivastava and Y Wang Chal lenges in resource monitoring for residential spaces in ACM BuildSys 2009 11 S Drenker and A Kader Nonintrusive monitoring of electric loads IEEE Computer Applicati
2. phase changes of the 3 speed tower fan can be correctly associated with the appliances As the differences between the powers of the tower fan in different speeds can be as low as 6 W the phase changes of the tower fan cannot be detected only based on power readings Table I shows the groundtruth measurements by KAWs and the estimation results of various approaches Both Supero and Oracle can accurately estimate the power and energy of each appliance The relative error of energy consumption averaged over all appliances is 3 6 and 3 1 for Supero and Oracle respectively As Light 1 2 and 3 have no nearby sensor Baseline uses the groundtruth states of Light 1 2 and 3 For other appliances Baseline uses the closest sensor to detect the state of an appliance As Baseline does not perform data correlation and event clustering the detections contain excessive false alarms For instance as the hair dryer is very noisy all acoustic sensors will raise detections when the hair dryer is on which introduce false alarms for all other acoustic appliances Water boiler and bath fan suffer from the same issue as well As a result Baseline yields wrong power and energy estimates for several appliances In fact it is highly difficult to deploy dedicated acoustic sensors as they can be easily triggered by any noisy appliance in the home Our results show that acoustic data from multiple sensors must be jointly processed to produce correct detections
3. problem as follows Light Cluster Appliance Association Problem Find a B and A to minimize E a 8 A subject to that Ym iv Omg 1 and V7 yy Omg 1 The constraint means that A is a one to one mapping To solve the above problem we first fix and 8 and then find A to minimize E a 6 A under the constraint Algorithm 1 Acoustic Transition Appliance Association Algo rithm Input acoustic transition set 7 non primarily monitored appliance set A Output acoustic transition appliance association 1 C 0 2 for transition k in 7 do 3 find sensor 7 with the largest absolute change of signal energy in k 4 if sensor is a primary sensor then 3 associate k with the corresponding primarily monitored appliance 6 else 7 C CUf k 8 end if 9 end for 0 cluster the transitions in C using k means algorithm based on their absolute power changes with A as the number of clusters 11 sort clusters according to their centers 12 sort appliances in A in terms of power 13 associate the sorted clusters with the appliances in A in order which is a linear assignment problem 22 We employ the Hungarian algorithm 22 with a time complexity of O M to solve this sub problem Henceforth the final solution can be found by enumerating a and in their possible ranges The association process can be further sped up by identifying the dedicated light sensors Cluster m is an dedicated cluster if Rm N Rn Yn m i e cluster m is m
4. and groundtruth appliance usage are collected and fed to the system Such a training process is often labor intensive and sometimes intrusive A promising solution is to sense the light acoustic and magnetic signals generated by appliances and then correlate them with the total household power measurement to infer per appliance energy consumption 9 However to achieve autonomous monitoring this approach would typically require sensors to be carefully installed for each appliance which may result in high installation cost 10 and reduced usability for non professional users In this work we ask the question is it possible for a residential power usage monitoring system to use only inexpensive sensing devices be easy to install and yet be capable of working based on a small amount of easily obtained prior information without resorting to supervised in situ training Such a system must automatically detect events of interest autonomously associate the events with the correct appliances and finally infer the power usage of each appliance Several key challenges must be addressed for achieving the unsupervised power usage monitoring First homes are a highly dynamic and complex environ ment Inexpensive sensors typically have limited sensing capabilities and hence likely produce false alarms or miss important events Second when the sensors are installed in an ad hoc manner it is highly difficult to associate an event detected by
5. can be efficiently computed The sensor remains in the fast sampling model when the acoustic signal energy is above a We set a low threshold na such that the acoustic sensors will not miss the sound triggered by an appliance of interest Note that different from a light event that refers to the switching on off of a light an acoustic event refers to the sound heard by a sensor Therefore the sensor will continuously report acoustic events while the sound persists We refer to the switching or phase change of an acoustic appliance as an acoustic transition Owing to intrinsic complexity of the acoustic modality acoustic transition detection is achieved by advanced pattern recognition algorithms on the base station This is due to the fact that simple algorithms cannot well handle the dynamic acoustic signals while complex pattern recognition algorithms pose significant computation overhead on sensors For instance the EDF based on signal energy can easily miss important events especially when the sounds from non power events e g shower and acoustic appliances e g bath fan overlap with each other C Power Event Detection Various commercial off the shelf smart meters e g TED 17 can deliver real time power consumption readings As the total power consumption is critical for identifying appliance events and estimating the power of each appliance the real time power readings are transmitted to base station for storage The base
6. monitor 6 lights including in candescent bulbs and standard compact fluorescent lamps Different colors of TelosB motes in Fig 11 represent differ ent placements which are also labeled with the initials of the color names i e R G B Y and BK The red and green placements follow the conservative deployment strategy discussed in Section VIII C The blue and yellow placements follow the incremental strategy to reduce the number of sensors from 6 to 4 As there is no sensor in the living room the black placement does not follow any Table IV Rim AND CLUSTERING ASSOCIATION RESULTS Dining light Kitchen light Doorway light Living light 1 Living light 2 deployment strategy All placements ensure the coverage requirement The distances between sensors and lights were determined by visual estimation We conducted a controlled experiment to evaluate each placement Table IV shows the set of sensors that can detect the same light 1 e Rm defined in Section VI A The clustering and association results of the red to yellow placements are correct In the black placement of 3 sensors although all events can be detected they cannot be correctly clustered For instance although sensor 6 can detect the near dining light 13 W and the farther living light 2 150 W the changes of light intensity from the two lights measured by sensor 6 are similar leading to incorrect clustering We then deployed 11 Ir
7. sensor s intensity at unit distance from the light source to its power The range of 8 can be easily obtained in offline lab experiments Based on the ranges of a and 6 Supero automatically learns the values of a and in a specific deployment such that the association minimizes the discrepancy between the measurements and the decay model This is desirable because otherwise determining their exact values through in situ calibration would be labor intensive and infeasible for non professionals The association between clusters and appliances is for mally represented by a square matrix A am j axm If cluster m is associated with appliance j am j 1 Otherwise am 0 Let um denote the average of the feature vectors in cluster m Hence the i component of Hm denoted by Hm i is the average change of light intensity measured by sensor when the corresponding appliance is turned on and off By denoting Rm as the set of sensors that makes positive decisions in cluster m we define the error caused by associating cluster m with appliance j as Emi Deis 8 Pm diy lim il where Pm is the power of the appliance that generates the events in cluster m We estimate Pm as the median value of the absolute power changes i e Ag of the events in cluster m For certain a 8 and association A the total error is defined as E a 6 A X vmvi aAm j m j Based on this error metric we formulate the light cluster appliance association
8. station detects interesting power events based on changes in power readings As the characteristic of power readings is similar to light readings we also apply the EDF to detect the rapid increases and drops in power The settings of the EDF will be discussed in Section VIII D Multi modal Data Correlation Due to limited sensing capability and complex home environment sensors can easily raise false alarms and miss important on off events of appliances For instance a light sensor may report events when nearby window blinds are opened and closed and an acoustic sensor can be trig gered by human conversations To deal with these sensing errors we present a two tier fusion approach to correlate the light power events and acoustic transitions reported by different sensors The first tier uses a short moving window to correlate the events transitions from multiple sensors of the same modality The events transitions falling into the same window are regarded to be generated by the same source This is equivalent to an OR rule decision fusion that can largely reduce the overall miss rate The second tier correlates the results of the first tier with the readings from the smart meter to remove false alarms Specifically the base station first calculates the power change of an event transition from the first tier If the power change is smaller than the minimum wattage among the monitored appliances the event transition will be discarded In S
9. the bathroom In this section we present the results of a controlled experiment in which we intentionally turned on and off the appliances The controlled experiment allows us to understand the micro scale performance of Supero 2 Energy Estimation Accuracy Fig 9 shows the groundtruth information power readings event detection and clustering results for the controlled experiment During this experiment all two light false alarms are identified by the multi modal event correlation There is no light miss detection Moreover all light events are correctly clustered and associated with lights For acoustic modality the non power sounds such as toilet flush and tap water can be identified by the multi modal data correlation From the third chart in Fig 9 Supero fails to detect the off event of fridge and total four events of water boiler In the experiment the fridge turned itself off about 3 seconds after the tower fan finished its transition from the first speed to the second speed As the window size of the window based acoustic event detection algorithm is larger than 3 seconds the off event of fridge is missed The miss detections of water boiler are caused by the delay of sound However as discussed in Section VI D by jointly treating fridge and water boiler as acoustic and unattended appliances these misses can be successfully recovered by the events detected from power readings Other detected acoustic transitions including the
10. we omit the details of the clustering algorithm Our experience shows that the clusters with a small number of feature vectors often affect the accuracy of clustering results To improve the robustness of clustering we detect outliers as follows If the size of a cluster is smaller than a small threshold its member feature vectors are regarded to be outliers which are discarded and then the clustering algorithm is re executed Outliers are produced by unidentified false alarms and rarely used appliances and hence removing them has little impact on the accuracy of overall energy consumption estimation The setting of the outlier detection threshold will be discussed in Section VIII B Acoustic Event Clustering and Transition Detection A challenge of acoustic event clustering is that many appliances such as multi speed fan have multiple opera tion phases Unfortunately the number of phases of many appliances cannot be easily determined by the user For instance refrigerators have different phases depending on the brand model and when they were made Moreover the number of actually used phases of an appliance such as multi speed fan highly depends on the habit of the user and hence is unpredictable In addition the overlaps between the sounds from different appliances and noises e g shower water flush can also result in unpredictable number of acoustic patterns As a result it is infeasible to assume a known and fixed number of clust
11. 200 3 4 0 0206 0 5 9 1481 0 0481 1 8 0 0289 41 0 3 23 28 35 0 0028 9 7 0 0045 45 1 A 507 0 0168 3 1 0 0163 0 0 ll 459 0 0150 5 1 0 0018 88 6 3 0 0795 1 4 0 0848 8 2 A 0 0020 N A 0 0048 N A 4 0 0154 4 8 0 0154 4 8 ll 0 4840 3 4 0 0472 0 9 O O Error is the relative error of energy in percentage with respect to the KAW measurements t Fridge s rated power is not available However its power events can be correctly associated when a rated power of 80 W to 400 W is given to Supero Table II ENERGY BREAKDOWN DURING 7 DAYS IN APARTMENT Table III ENERGY BREAKDOWN IN HOUSE 1 ppliance Baseline ppliance upero Name Error Name ITOT Light I X J I 0 9 ntry light 008 3 Light 2 0 8 4 92 08 Hall light 0109 19 Light 3 1 7 6 25 1 7 Kitchen light 0056 5 8 i Dining light 0113 24 6 Light 4 0 1 1 48 1 7 rae a Living light 0040 3 1 Light 5 0 7 0 41 55 ea Master bed light 0061 6 0 Water boiler 1 6 0 100 Master bath lish 0052 3 6 Tower fan 17 9 0 24 66 2 rae corre 5 Master bath fan 0068 2 3 Rice cooker 1 2 1 01 0 8 Guest bed light 0056 21 2 Hair dryer 0 4 0 02 73 2 Guest bath light 0070 0 6 Fridge 3 2 11 8 3 2 Guest bath fan 0097 0 0 Bath fan N A 0 N A Stove burner 4675 1 6 Router 3 04 433 Water dispenser 0518 N A Average error 1 0 Average error 6 Power P and W h Error is the relative error of energy in percentage with respect to the KAW measurements between the base station and sensors can affect the perf
12. Supero A Sensor System for Unsupervised Residential Power Usage Monitoring Dennis E Phillips Rui Tan Mohammad Mahdi Moazzami Guoliang Xing Jinzhu Chen David K Y Yau Department of Computer Science and Engineering Michigan State University USA Advanced Digital Sciences Center Illinois at Singapore Abstract Research has shown that providing users fine grained information concerning their power usage in the home fosters conservation Several existing systems achieve this goal by exploiting appliances power usage signatures identified in labor intensive in situ training processes Recent work shows that autonomous power usage monitoring can be achieved by supplementing a household power meter with distributed sensors that detect the working states of appliances However sensors must be carefully installed for each appliance resulting in high installation cost This paper presents Supero an ad hoc sensor system that can monitor appliance power usage without supervised training By exploiting multi sensor fusion and unsupervised machine learning algorithms Supero can classify the appliance events of interest and autonomously associate the power usage with respective appliances Our extensive evaluation in five homes shows that Supero estimates the energy consumption with errors less than 7 5 Moreover the users can deploy Supero with considerable flexibility and in short time I INTRODUCTION Since 1978 the p
13. The rationale of jointly considering the number of acoustic transitions in the minimization objective is as follows The misclassification rate typically decreases with the window size Therefore only minimizing the sum of misclassification rates will mostly result in an unreasonable small window size Fig 5 shows a case study using an acoustic sensor to identify the number of phases of a 3 speed fan and detect the phase changes We can see that the number of phases can be correctly identified as 3 Moreover the phase changes can be accurately detected C Unattended Power Event Clustering For the unattended power events i e detected power changes without simultaneous light acoustic events Supero adopts the Euclidean distance between the power changes as the dissimilarity metric and applies the k means algorithm to cluster the events into My clusters To simplify the discussion in this paper we assume that the unattended appliances are not multi phase However by extending the approach developed for acoustic modality Supero can be extended to address multi phase unattended appliances VI AUTONOMOUS APPLIANCE ASSOCIATION The event clustering does not answer which appliance triggers the events in a cluster To accurately estimate the energy consumption of different appliances the events must be correctly associated with the appliance that generates them In this section we address this issue by exploiting the correlation of even
14. We note that the performance of regression can be potentially improved if magnetic sensors are employed When attached to the power cord of the appliance a magnetic sensor may detect the working state of appliance more reliably than acoustic sensors However this significantly increases the installation cost Especially it is very difficult to install the magnetic sensors for the permanently installed appliances without exposed power cords 3 Impact of Distance Errors As the association algo rithm presented in Section VI A requires the distances be tween sensors and lights we now evaluate the robustness of the association algorithm with respect to the distance errors As Light 4 and Light 5 can only be detected by dedicated sensors cf Fig 8 b and Fig 9 Supero autonomously prunes their clusters to speed up the association algorithm as discussed in Section VI A Hence we only focus on Light 1 2 and 3 which are monitored by Node 1 and 2 The distances between the lights and nodes are within to 3 meters The distances given to Supero are distorted as follows First we proportionally increase all the distances As the association algorithm can find a best fit scaling factor 8 the association remains correct even if we multiply the distances by 10 router failures reset router 0 seg 2 8 30 8 31 9 1 9 2 9 3 9 4 95 9 6 9 7 9 8 99 Date from 1AM Aug 30 2011 to 11AM Sep 8 2011 PRR a PRR of a KAW Power kW
15. association approach for acoustic modality Sensor 7 is defined as the primary sensor of appliance j if the absolute change of signal energy of sensor i is always the largest when appliance j changes its state and must not be the largest when any other appliance changes state The appliance 7 is defined as a primarily monitored appliance The complement set of primarily monitored appliances comprises non primarily monitored appliances Different from dedicated sensor that can only sense one appliance a primary sensor can sense multiple appliances The primary sensors can be identified according to user s intuition based on the sensor and appliance locations When a sensor cannot be accurately classified as a primary sensor it can be conservatively excluded from the set of primary sensors The pseudo code of the acoustic event appliance association is in Algorithm 1 In Line 12 of the algorithm the non primarily monitored appliances are sorted according to power Hence the required extra prior information is the order of non primarily monitored appliances in terms of power C Unattended Appliance Association As the unattended appliances are not sensed by any sensor more accurate prior information about them will be required Similar to light modality the power of the appliance that generates the unattended power events in cluster m denoted by Pm can be estimated as the median value of the absolute power changes of the events in cluste
16. ategy is to jointly monitor most acoustic appliances and input their rated powers Supero is expected to become more robust to misses if more rated powers are provided gt 25 T T T T T J Z ask mee a a a a i SORTA b a aaa w an a AR aA y WD envelope 5 1 HOUMA AAAS AN Ong ynni a MG J 3 ost kyuma wiy n E ph Z 08 Z 0 gt 06 gt 8 o threshold g 0 fi fi L 1 0 1000 2000 3000 4000 5000 6000 Time s Fig 6 Detecting stove burner 1 Red curve Total household power readings when a burner is working Blue curve The reconstructed lower envelope 2 Standard deviation of power readings and threshold based detection results detection window size 100s D Energy Calculation Supero adopts a simple approach to calculate the energy consumed by each appliance based on the detected events and estimated powers For a single phase appliance the energy consumption is simply the product of power and the on time For a multi phase appliance its power can be updated according to the associated power changes Inte grating the power over time yields the energy consumption The evaluation will show that this approach leads to satis factory results Based on the association results regression approaches e g 9 can also be integrated with Supero to improve the robustness in the case of false alarms and misses VII DUTY CYCLED HEATING APPLIANCES As discussed in Section HI B
17. cent measurements when Z n Z n lt nz The threshold nz is continuously updated according to the noise model to achieve a low false alarm rate e g 5 The settings of as a and 7 will be discussed in Section VIII Fig 2 shows the operation of the EDF on the readings of a photodiode when two lights are turned on off and a person moves around It can be seen that the light events can be accurately detected and the human movements do not trigger false alarms B Acoustic Event Detection A challenge in acoustic sensing is that high sampling rate is often required to extract acoustic features of interest Motivated by the observation that many appliances remain off in most of the time Supero adopts an adaptive sam pling scheme to reduce computation overhead for sensors Initially the sensor samples acoustic signal at 1 kHz for 0 05 seconds i e 50 samples every 2 seconds When the signal energy exceeds a threshold 774 the sensor switches to a high sampling rate of 12 5kHz to capture more details of the possible event In the fast sampling mode the sensor samples for 0 08 seconds i e 1024 samples every 2 seconds A series of software filters decompose the signal of 1024 samples into low pass band pass and high pass signals Signal energy and zero crossing count of the signals in the whole band and the three subbands are computed and transmitted to the base station Note that zero crossing count characterizes frequency and
18. cled electricity vampires VIII IMPLEMENTATION CONFIGURATION AND DEPLOYMENT A System Implementation Sensors and smart meter The sensors are implemented using TelosB and Iris motes 23 TelosB only has light sensors while Iris has both light and acoustic sensors According to our lab tests the light sensors on TelosB and Iris have satisfactory isotropic sensitivity in a considerably large incoming angle which can mitigate the impact of sensor orientation on the association algorithm presented in Section VI A The light sensors on TelosB and Iris have also been calibrated such that they can be used in the same deployment The signal sampling and event detection algo rithms described in Section IV are implemented in TinyOS 2 1 To reduce computation overhead these algorithms are carefully implemented using integer arithmetic All sensors use 802 15 4 channel 11 The sensors communicate directly with the base station Such a single hop topology suffices for our deployments in three apartments and two multi story houses TED5000 17 is used to measure the total household power consumption Base station The base station is a TelosB mote connected to a netbook computer A daemon service on the computer retrieves real time power readings from the TED5000 and stores the event messages received by the base station mote in a database The data correlation clustering and association algorithms are implemented in GNU Octave Groundtruth Ki
19. communication Due to the router failures and lost groundtruth information from 2 and seg 3 shown in Fig 10 a The total length of the three segments is more than 7 days The three data segments are concatenated and then fed to the clustering and association algorithms Power spikes Power spike is a typical dynamics in power lines which can be caused by bad weather conditions and turning on off appliances in the tested home and even neigh bor homes Power spikes may cause errors in the appliance power estimation In the controlled experiment we can see a few power spikes in the top chart of Fig 9 when an appliance changes state As we apply a guard region for computing the power change as discussed in Section IV D the power spikes do not affect the results However in the 10 day experiment we observe excessive power spikes as shown in Fig 10 b that can affect the calculation of power changes for the detected events We suspect that the power spikes observed on September were caused by the thunderstorms during the period of experiment A zoomed in view of the power trace on that day is shown in Fig 10 c Almost all power spikes can be removed by a median filter with a window size of 7 seconds We also apply the median filter with the same setting to the power traces collected in other experiments 2 Evaluation Results Table II shows the results based on the data of 7 days During this period 713 false alarms out of t
20. ction IV D Unsupervised event clustering By leveraging unsuper vised clustering algorithms the events generated by an appliance can be classified into the same cluster and the power of the appliance can be estimated by correlating with the measurements of the smart meter Section V Autonomous event appliance association Supero asso ciates the classified events with respective appliances based on event features and prior information Based on the clus tering and association results Supero calculates the energy consumption of each appliance Section VI IV EVENT DETECTION AND DATA CORRELATION In this section we first describe the event detection algorithms for sensors and then present a multi modal data correlation algorithm to reduce sensing errors A Light Event Detection Light sensors detect the state changes of lights from the changes of light readings However besides electrical light 2600 l sensor readi TERN zaoo zsif ir E a 5 2200 Lo ph pH cheb IR Light 1 on 2000 eh Nese eile 100 S I 1800 am a 1600 C L 1 L L 1 L 1 ma 0 20 40 60 80 100 120 140 Time second Fig 2 Operation of EDF Tr 4 Green vertical lines represent detections A person passes by Light 1 at the 31 and 53 second events the change of sensor readings can also be caused by sensor noise and natural ambient light change We present a light event detection algorithm that is resilient to these d
21. deployment strategy the signal emitted by an appliance can be sensed by multiple sensors More over due to the spatial distribution of sensors appliances and environmental dynamics the group of sensors that can detect each appliance is different For a particular appliance although the features measured by sensor 7 are dynamic due to noise their variance typically falls within a small range Hence the feature vectors of the events caused by the same appliance are clustered in the feature space Fig 3 shows the feature vectors of intensity change measured by two light sensors when three standing lights nearby were turned on and off We can clearly see that the feature vectors are clustered The light event features will be clustered into Mz clusters The Euclidean distance between two feature vectors can be small when non zero vector entries are measured by completely different light sensors leading to potential false clustering result Hence the Euclidean distance is not a desirable dissimilarity metric for the light modality Supero adopts a new dissimilarity metric that incorporates sensor location information Let bx 0 1 denote the detection decision made by light sensor i regarding event k where bki 1 means that sensor i detects on off event of some appliance The decision vector denoted by Bx is given by By bk 1 bk 2 bk N The dissimilarity between two decision vectors B and B is defined as d Bk B Di bk
22. ercentage of residential electricity has increased from 17 to 31 1 while the cost of energy has also been on the rise Consumers have become more interested in reducing their energy usage by appliances If the home owner could have a better understanding of the energy consumption of each appliance the waste of power could be identified Research 2 has shown that providing users information concerning their fine grained power usage in the home fosters conservation Previous systems for residential power usage monitoring can be broadly classified into two basic categories The first category direct sensing measures power usage through in line power meters Examples include Kill A Watt 3 Watts Up 4 and radio enabled ACme 5 As these meters are connected in between the appliance and power outlet they cannot be used on the appliances permanently connected to the power lines such as ceiling lights The second category indirect sensing is less intrusive as it infers the working states and energy consumption of individual appliances by detecting their power usage patterns 6 7 or the ambient signals they emit during operations 8 However the accu racy of these techniques can be influenced by the physical characteristics of the electrical wiring and appliances As a result many of them need in situ supervised training The first two authors are listed in alphabetic order in which the information about energy consumption
23. erm monitoring e g a few weeks such that the generated report is instructive and meaningful for identifying wasteful energy usage Four major challenges are brought by the above design objectives First due to the ad hoc deployment a sensor may pick up the signals emitted by multiple appliances making it difficult to differentiate which appliance is consuming power For instance a light sensor can sense the signals emitted by various lights and an acoustic sensor in the kitchen can hear the sounds from exhaust fan disposer microwave and etc Second without careful installation sensors typically suffer sensing errors caused by interference from environment and human activities For instance light sensors likely report false alarms when nearby window blinds are opened during the day time and acoustic sensors may pick up sounds unrelated to power consumption such as human conversation Third without in situ system training more prior information is often required to bootstrap the unsupervised learning approaches We strive to reduce the difficulties for non professional in obtaining the prior infor mation required by Supero while maintaining satisfactory monitoring accuracy Finally to extend system lifetime wireless sensors must adopt lightweight sensing algorithms and send the least amount of data which however imposes challenges to accurate appliance working state monitoring B Motivation To meet the aforementioned object
24. ers in the collected acoustic features We propose the following approach based on advanced pattern recognition algorithms to address the above challenges The dimensionality of acoustic feature vector is 8N a which will incur heavy computation overhead in clustering even when a few acoustic sensors are deployed Supero first applies principal component analysis PCA to reduce the dimensionality In our experiments to keep 99 sample variance the dimensionality can be reduced from 40 to 8 when 5 acoustic sensors are deployed Supero then estimates the number of clusters as kop arg max wee B1 where S k and Su k are the between cluster and within cluster variance matrices when the specified cluster number is k For each given k the k means algorithm is executed to cluster the features into k clusters and calculate S k and Sw k With the clusters under kopt Supero detects the acoustic transitions by identifying the transitions between clusters over time Specifically by dividing time into small windows edges between two consecutive windows having different largest clusters are detected as acoustic transitions The window size is selected to minimize the product of the number of acoustic transitions and the sum of misclas sification rates in all windows The misclassification rate in a window is the ratio of the number of events that do not belong to the largest cluster in the window to the total number of events in the window
25. eviation of the power readings in the top part and the detection result We can see that the time duration that the burner is working can be accurately detected For the power readings in a window that has a positive detection we apply the k means algorithm with k 2 and then interpolate the power readings in the cluster with smaller average to reconstruct the lower envelope of power consumption i e the background power as shown in the top part of Fig 6 With the lower envelope it is easy to calculate the energy consumption of the duty cycled appliance In typical U S homes stove burner and oven are the major duty cycled heating appliances and they are often the components of the range Supero does not differentiate the duty cycled heating appliances and attributes all energy consumption to the range To address multiple simulta neously working duty cycled appliances the number of clusters i e k can be first determined by the technique presented in Section V B The rapid duty cycling can cause significant errors to the EDF based power event detection cf Section IV C and the second tier of the multi modal data correlation cf Section IV D Hence when a duty cycled appliance is detected Supero disables these two components and the power changes of the light acoustic events in this period are set to be missing values Although such a design can cause errors to other appliances it is worthwhile to give priority to the duty cy
26. exibility in choosing the sensor positions Compared with the existing approaches that require careful sensor installation this feature is highly desirable to facili tate the large scale deployment in practice 2 Deployment Time We now present two case studies on how easily Supero can be deployed and configured by non professionals We recruited two homeowner volunteers to deploy Supero in their homes including a single bedroom apartment Apartment 3 and a two story house with base ment House 2 We first introduced Supero and explained the deployment strategies to the volunteers which took less than one hour They then installed the sensors and configured the system using the web interface without any instructions from us For safety reasons they did not install TED5000 In Apartment 3 the volunteer deployed 5 TelosB and 3 Iris motes to monitor all appliances including 5 lights a fridge a microwave and a fan The deployment and configuration only took about half an hour In House 2 the volunteer took about one hour to survey the appliances and another hour to install the sensors He finally deployed 12 TelosB and 10 Iris motes to monitor 12 lights an exhaust fan in kitchen a waste disposer a dish washer a fridge a microwave and three fans in three bathrooms The base station on the first floor can reliably receive the packets from the sensors distributed on two floors and basement After the system deployments we conducted con
27. h In addition we leverage TPCDB 20 which is an online collaboratively edited database of appliance powers to help the user input the required rated powers Currently TPCDB comprises the information of more than 500 appliances Fig 7 b shows our interface of querying TPCDB through its web service API where the user can find the rated power by appliance type manufacturer and model IX EXPERIMENTAL EVALUATION A Deployments and Methodology We deployed and evaluated Supero in five typical house hold environments We first deployed Supero in a 40 m single bedroom apartment Apartment 1 and an 150m three bedroom ranch house House 1 to evaluate the perfor mance of Supero As most appliances in Apartment 1 can be monitored by groundtruth KAW meters the Apartment 1 deployment allows us to extensively evaluate the accuracy of Supero In House 1 most appliances are hardwired to power lines and hence we cannot collect complete groundtruth information using KAWs The major purpose of the House 1 deployment is to evaluate the portability of Supero to larger home environment We further deployed Supero in other three homes to evaluate the impact of sensor placement on the sensing results and how easily Supero can be deployed by non professional volunteers We compare Supero with two baseline approaches These two baselines are based on a state of the art residential power monitoring system called ViridiScope 9 ViridiS cope estimate
28. heating appliances such as stove burner and oven are major electricity consumer in homes Most modern heating appliances duty cycle to achieve the desired heat level For instance the top part of Fig 6 shows the total household power readings when a GE JB710ST2SS burner is working As the cycle can be short e g several seconds the EDF based detector discussed in Section IV C will have poor performance In this section we propose a new approach to detect the duty cycling pattern from the total power readings and calculate the related energy consumption Duty cycled appliance rapidly switches between on and off causing large variation in power readings Hence we detect the duty cycling pattern based on the standard devi ation of the windowed power readings By denoting P and y 0 1 as the power and duty cycle of the appliance the standard deviation of the power readings can be derived as P y 72 We choose a threshold of PV0 05 0 057 by conservatively assuming that the duty cycle is greater than 5 When P is unknown we can choose a default value of 1 5kW for P because most duty cycled heating appliances have a rated power around 1 5kW 20 As a result the default threshold is 0 327 kW To suppress the false alarms caused by other high power non duty cycled appliances we further require that the zero crossing count of the mean removed power readings in a window is at least 2 The bottom part of Fig 6 shows the standard d
29. i S gt bji ae bki g bj is where represents exclusive OR 7 4 bk b is the number of sensors that can only detect either event k or j but not both and D br gt 6 4 is the number of sensors that can detect both event k and j Hence d Bp Bj quantifies the net difference between the sets of sensors observing the two events By denoting X X as the Euclidean distance between the feature vectors X and X in event k and j the new dissimilarity metric is defined as sS X Xill d Br Bj lt do XX Xi Xj 4 d Bp Bj gt do V det S k det Sw k 1 OF 41 1 L 0 2 4 6 8 10 b Time minute g 5 Acoustic event clustering and transition detection for a 3 speed fan ng The number of 3 hases is identified as 3 b Clustering and transition detection results ere Y axis is the major principle component PC vertical lines represent the detected acoustic transitions where do is a threshold and 6 is a large constant that can separate the feature vectors observed by very different subsets of sensors The setting of dg will be discussed in Section VIII Clustering algorithms based on the Euclidean distance e g k means cannot be applied due to the use of the metric defined in 1 Supero adopts the merging based clustering algorithm 21 which is applicable to nonlinear dissimilarity measures to group the feature vectors into Mz clusters Due to space limitation
30. ight Acoustic Smart Meter Sensors Fig 1 Architecture of Supero C System Architecture Supero is composed of a number of wireless sensors distributed in the home a wireless smart meter and a base station receiving the information from the sensors and the smart meter Fig 1 illustrates the architecture of Supero In this work we only consider light and acoustic sensors while other sensing modalities such as infrared can be easily incorporated by Supero Supero has a two tier architecture as follows In the first tier sensors sample signals and detect the events that are possibly caused by turning on off appliances Specifically if a sensor detects a significant change in the received signal it extracts various relevant features and sends an event message to the base station The details of the first tier will be presented in Section IV The base station provides a graphic configuration interface that allows user to input prior information such as sensor appliance distances and appliances rated powers When Supero is requested to generate a power usage report for a specified time period the base station executes the following algorithms based on the collected data and the prior information input by user Multi modal data correlation The base station corre lates sensor events and power readings to differentiate the events generated by turning on and off appliances and the false alarm events unrelated to power consumption Se
31. ilot deployment The second group of parameters are do and outlier cluster size in the light event clustering presented in Section V A We set do 2 i e two feature vectors should be classified into two distinct clusters if the number of sensors that can only detect the first event is 2 more than that of the second event Moreover we set the outlier cluster size to be 2 i e we ignore the appliances that only generate less than 2 events in a long period such as several days As other parameters can be either easily set e g a for acoustic sampling and 6 in 1 or autonomously optimized e g and p in Section VI A we omit the details here C System Deployment In this section we first discuss the sensor deployment strategies and then summarize the user inputs to Supero A necessary condition for correct clustering and association is that every appliance can be detected which is referred to as the coverage requirement A conservative and intuitive deployment strategy is to place a sensor close to each appliance The user manual can provide a table of detection ranges for typical household appliances which are measured in offline experiments For instance a 60 W incandescent bulb can be reliably detected by a TelosB mote within 5 m In addition we also discuss an incremental deployment strategy that can possibly reduce the number of sensors Initially a sensor is deployed for each appliance that emits dim signal Appliance P
32. is motes as shown in Fig 11 and then select four different subsets of them to evaluate the impact of sensor placement The subsets are S All 11 Iris motes S2 10 12 14 15 16 18 20 S3 10 12 14 19 and S4 10 14 The acoustic appliances covered in the experiment include exhaust fan over the range waste disposer in the sink dish washer and vacuum During the experiment we used the vacuum both in the dinning area and living area S and S2 use redundant sensors and hence are conservative S3 basically follows the incremental deployment strategy As there is no sensor in the living area S4 does not follow any deployment strategy All subsets ensure the coverage requirement As the exhaust fan has two phases low and high speed sensor 10 that is closest to the fan is designated as the primary sensor of the fan For other appliances the order rather than the actual values of their power consumption is provided to Supero The event detection and association results for S1 S2 and S3 are correct For S4 although all acoustic events can be successfully detected some of them cannot be correctly associated For instance when the vacuum ran in the living area sensor 10 received the highest signal energy which is inconsistent with its designation as the primary sensor of the exhaust fan From the above results we can see that by following the deployment strategies discussed in Section VII C the user has considerable fl
33. ives Supero utilizes a household power meter and a small number of inexpensive light and acoustic sensors that are deployed in an ad hoc manner in the home Based on an unsupervised approach it does not require any in situ system training but leverages a small amount of prior information that can be easily obtained by non professional users We now discuss several important observations that motivate our approach Real time total household power metering Nowadays the real time total household power consumption can be easily measured by installing a commercial off the shelf smart meter e g TED 17 and AlertMe 18 on the main circuit panel These meters are inexpensive and most of them can be easily installed without hardwiring with the power lines 18 Moreover as the coverage of smart grid increases the real time total household power readings are increasingly available to the homeowners without resorting to a personal smart meter Sensing modalities According to a survey of U S Depart ment of Energy 19 the average distribution of electricity consumption in household is heating 24 lights 24 air conditioners ACs 20 refrigerators 15 dryers 9 and electronics 9 As most heating appliances consume sub stantially more power than other appliances their consump tion trace often can be identified from the real time total household power readings Most lights ACs refrigerators and dryers emit detectable light and acous
34. lding These two studies exploited the tree topology of the power supply system to reduce the number of sensors 13 and derive estimation quality 14 In 7 an electrical event detection and classification approach was developed based on the frequency patterns of the transient noises generated by switching on off appliances and measured by a single in line sensor However the transient signature is heavily influenced by the physical characteristics of the electrical wiring which results in the need of post deployment training In 8 15 appliances were recognized based on their electromagnetic interference 8 and acoustic signals 15 However both two approaches need labor intensive in situ training for the particular home they are deployed in A typical training process involves switching different appliances collecting and labeling signals In a recent work 16 a thermal camera is employed to detect the on off states of the appliances in its field of view which are utilized to infer the per appliance energy consumption Compared with tiny wireless sensors the thermal camera is cumbersome hard to install and can raise privacy concerns in residential environment ViridiScope 9 is a fine gained power usage monitoring system that is closest to Supero It features an autonomous regression based calibration framework that can calculate the energy consumption of each appliance A fundamental requirement of the regression approach i
35. ll A Watt meters In order to evaluate the accuracy of Supero we build 14 power meters based on the P3 Kill A Watt KAW Model P4400 3 to provide groundtruth power usage data of individual appliances We connect two ADC channels of a Senshoc mote to two pins on the internal circuit board of a KAW to sample the voltage and current signals Senshoc is a TelosB compatible mote implementation with significantly reduced cost Fig 8 a shows a modified KAW The Senshoc mote computes and transmits the real time power usage data to the base station for storage Each modified KAW is carefully calibrated to output accurate power readings B System Configuration The parameters of algorithms in Supero are determined by offline experiments Note that this process does not need to be repeated for different deployment environments All the deployments in our experimental evaluation use the same parameter settings The first group of parameters are the coefficients of the EDF for light and power event detection presented in Section IV By setting a 0 18 a 0 074 and 7 4 light sensors are resilient to sensor noise and normal human movement By setting a 0 31 a 0 08 and T 4 power changes as small as 50 W can always be detected As the above settings depend on sensor noise and reading calibration they are sensor specific but environment independent The above settings are obtained by extensive evaluation on raw sensor data collected in a p
36. oO amp i yap Hanah vin 830 8 31 9 1 9 2 93 94 95 96 97 98 99 Date from 1AM Aug 30 2011 to 11AM Sep 8 2011 1 b Power trace 15 raw power reading aa 1L tered eG indow size x ost Mii Ml z 0 lp init too j amp ost 1 i caused by appliances 4AM 10AM 4PM 10PM 4AM Time starts at 09 01 2011 4AM c 12 hours of power trace on September 1 2011 Fig 10 PRR and power traces in 10 days Second we add a random bias to a particular distance in each test The result shows that if the bias is within 70 of the true distance the association remains correct These results demonstrate that Supero is robust to the errors in the light sensor distances Finally when we exclude Node 2 from the evaluation the results remain the same as long as the order of the distances from Node 1 to Light 1 and Light 3 is consistent with reality i e Light 1 is farther from Node 1 than Light 3 C 10 Day Experiment in Apartment 1 To evaluate the performance of Supero in long period we conducted an uncontrolled experiment that lasts for 10 days in Apartment 1 with the deployment shown in Fig 8 b During the 10 days two residents led normal life in the apartment In this section we first discuss our experiences and learned lessons and then present the evaluation results 1 Experiences and Learned Lessons Router failures The probe of TED5000 installed on the power panel sends real time
37. omes Apartment 2 Apartment 3 House 2 to evaluate the impact of sensor gt I kitchen light standing light lt jdish washer ceiling light TelosB drop cord light O Iris CH wall light Fig 11 Sensor placements in Apartment 2 The numbers in the squares and circles are the sensor IDs of TelosB and Iris If a TelosB does not face upward the arrow represents its facing direction Node 2 2nd placement Node 5 4th placement Fig 12 Sensor installation examples Sensors were placed on the ground in the corner of walls on the fan of a range and on a table placement on the sensing results and provide case studies on whether non professionals can successfully deploy Supero in their homes and how much time they need 1 Impact of Sensor Placement We deployed 6 TelosB and 11 Iris motes in Apartment 2 which is a 80m two bedroom apartment to evaluate the impact of sensor place ment on the sensing results Sensors were only deployed in the doorway living room and kitchen area as shown in Fig 11 As the doorway light 1 and doorway light 2 are in series they are regarded as one light As shown in Fig 12 sensors were placed or attached on ground walls appliances and furniture using mounting tape We separately evaluate the impact of sensor placement on the light and acoustic sensing algorithms We conducted five sensor placement trials to
38. onitored by dedicated sensors Before running the Hungarian algorithm each dedicated cluster m is associated with the appliance that is closest to the sensors in Rm The unassociated clusters and appliances are then fed into the Hungarian algorithm The association algorithm requires the knowledge of sensor appliance distances that can be estimated by a sonic laser tape arm span or even rough visual estimation The association algorithm is robust regarding the distance es timation As long as the relative order of the distances is correct the association results are not affected Moreover the optimization framework finds a and 8 by jointly ac counting for the detected event features and sensor appliance distances As a result the ranges of a and limit the impact of inaccuracy of sensor appliance distances on association results These observations are confirmed in Section IX B Acoustic Transition Appliance Association For electrical lights most power is consumed in the form of light Hence the scaling factor 6 in the power law does not vary substantially across different lights Although acoustic signal also follows the power law in contrast to light it is typically a by product in the operation of appli ances As a result the scaling factor can vary significantly across different acoustic appliances and the association al gorithm developed for light modality is not well applicable to acoustic modality We propose a heuristic
39. ons in Power vol 12 no 4 1999 12 L Farinaccio and R Zmeureanu Using a pattern recognition approach to disaggregate the total electricity consumption in a house into the major end uses Energy and Buildings vol 30 no 3 1999 13 X Jiang M Van Ly J Taneja P Dutta and D Culler Experiences with a high fidelity wireless building energy auditing network in ACM SenSys 2009 14 D Jung and A Savvides Estimating building consumption breakdowns using on off state sensing and incremental sub meter deployment in ACM SenSys 2010 15 Z C Taysi M A Guvensan and T Melodia Tinyears spying on house appliances with audio sensor nodes in ACM BuildSys 2010 16 B Ho H Kao N Chen C You H Chu and M Chen Heatprobe a thermal based power meter for accounting disaggregated electricity usage in UbiComp 2011 17 The Energy Detective http www theenergydetective com 18 AlertMe http www alertme com 19 U S Dept of Energy Annual energy outlook 2006 20 The power consumption database http www tpcdb com 21 R Duda P Hart and D Stork Pattern Classification Wiley 2001 22 R Burkard M Dell Amico and S Martello Assignment problems SIAM 2009 23 Memsic Corp TelosB Iris MTS MDA datasheets 2011
40. or KAWs we only use three data segments seg 1 seg mance of Supero Each Supero sensor only sends a packet when an event is detected while each KAW continuously transmits groundtruth power usage to the base station by the attached Senshoc mote equipped with a CC2420 radio Therefore we use the data traces of KAWs to examine the packet reception ratio PRR Fig 10 a shows the PRR of a KAW during the 10 days We can see that the communication performance significantly degraded and fluctuated between the evening of September 1 and the noon of September 3 As the residents watched online videos during this period we suspect that the poor link quality was caused by the interference from WiFi We also examined the traces of other KAWs Similarly their link quality degraded during this outrage period We were able to repeat this phe nomenon in an extra experiment using Senshoc motes and two laptops that transferred a large file over WiFi Although the channel of Senshoc was set to 11 which is well separated from channel 6 used by WiFi the PRR of Senshoc still significantly degraded However we did not observe signif icant degradation of PRR when experimenting with TelosB and Iris motes Hence we suspect that the performance degradation is caused by the imperfect antenna design of Senshoc Nevertheless after the 10 day experiment we have enabled packet acknowledgment and added retransmission mechanism to enhance the reliability of
41. otal 859 light events were raised by light sensors where 703 false alarms are identified by the multi modal data correlation All the remaining false alarms are iden tified as outliers by the clustering algorithm presented in Section V A In addition to the acoustic transitions generated by fridge 60 acoustic transitions were detected We can see from Table II that Supero can accurately estimate the energy consumption of lights The tower fan was turned on and off twice in the experiment and all its transitions were detected However two bath fan transitions were incorrectly associated with the tower fan because Node 13 i e the primary sensor for tower fan heard loud noises in living room at the same times The two false associations introduce errors to the energy estimates of tower fan and hair dryer From Table II the overall performance of Supero is close to that of Oracle Baseline fails to estimate the energy consumption of several appliances due to excessive false alarms D Controlled Experiment in House 1 House 1 is an one story three bedroom ranch house with living space of about 150m Compared with Apartment 1 it has more lights in various types incandescent bulbs standard compact fluorescent lamps and an electrical stove burner Hence the House 1 deployment evaluates whether Supero can be ported to a different household environment The deployment consists of 7 TelosB and 3 Iris motes The Iris motes detect both ligh
42. ower Database Data source www tpcdb com Part 2 Acoustic Sensi Type Manufacturer EA merenna Monitors 7 Asus v fi Det Model Power 12 Del Del alll VE276Q 52 17 W Ne Add MS246H 33 0 W b Acoustic appliances 24T1E 75 0 W Name Primary muttiphase Power op MS238H 33 0 W sensor order bath fan None No 1 De MT276H 70 0 W hair i 7 F 22T1E 75 0 W dryer None No 7 27 Del tower fan 11 v Yes N A Del Other online databases iE e California Energy Commission AED e Energy Star e Natual Resources Canada enerGuide lia Add a Acoustic configuration Fig 7 b Rated power database Web configuration interface The user then switches each other appliance to check if it can be detected by any already deployed sensor by looking at sensor s LED which blinks to indicate a detection If not an additional sensor is deployed for the uncovered appliance This process repeats until the coverage requirement is met Finally a few extra sensors may be deployed at random locations in the regions e g living room dining area without any sensors We now summarize the user inputs to Supero First Supero needs a list of monitored appliances which are categorized as lights acoustic appliances and unattended appliances Supero also needs to know whether an appliance has multiple working states while
43. possibly multiple sensors with the appliance that generates the event Finally to make the system practical it is desirable to minimize the amount of prior information about the appliances that needs to be collected by users At the same time the system must ensure the accurate power usage monitoring based on the limited prior information without any supervised post deployment training This paper presents the design and implementation of Su pero a System for Unsupervised PowER mOnitoring using inexpensive wireless sensors that are ad hoc deployed in the home Supero utilizes a power meter to measure the real time total household power consumption and inexpensive light and acoustic sensors to detect the events of appli ances Supero adopts a multi sensor fusion scheme where the data collected by power light and acoustic sensors are correlated to mitigate the impact of noise and remove possible sensing errors By using advanced unsupervised clustering algorithms Superio analyzes the signal signatures of different appliances and identifies the events generated by the same appliance Moreover Supero autonomously associates the classified events with appliances through an optimization framework that accounts for environment dependent factors like light signal propagation Provided with a small amount of easily obtained prior information such as sensor appliance distances and the rated powers of a small subset of appliances these unsuper
44. r m Supero associates the clusters with appliances by matching P s with the rated powers The association can be formulated as a linear assignment problem where the error of associating cluster m with appliance j is defined as em j Pm P and P is the rated power of 7 As the association is accomplished by an optimization algorithm it is resilient to small deviations between the true working power and rated power We propose to create a virtual background appliance to represent all the appliances that consume little but variable powers such as laptop computers In the association algorithm the association error of the background appliance is always Zero 1 m j 0 for any cluster m In other words the background appliance can be associated with any cluster such that it will not affect the association of other unattended appliances Our pilot deployments show that for various acoustic appliances that have complex signal patterns the sensors may miss important events For instance the sound of a water boiler becomes detectable in a couple of seconds after turned on The delayed acoustic event may be falsely removed by the data correlation due to little power change To address this issue we treat such an acoustic appliance as an unattended appliance as well and then merge the acoustic transitions and power events In practice the user may not know which acoustic appliances might be missed by sensors A conservative str
45. re ground truth power usage of appliances It consists of a Sensehoc mote and a KAW b Apartment 1 deployment always on Supero can estimate its energy consumption Hence Supero can estimate the energy consumption of the laptops as the difference between the total energy consump tion and the sum of estimated energies consumed by all other appliances A KAW is connected to each appliance to provide groundtruth power usage except for the bath fan that is hardwired on the ceiling The rice cooker 500 W water boiler 1500 W and fridge about 150 W are treated as unattended appliances The water boiler and fridge are also monitored by acoustic sensors However there are numerous miss detections due to sound delay and low sound level as discussed in Section VI C Fig 8 b shows the floor plan of the apartment and the sensor positions The sensors are placed on the floor nearby table chairs and toilet The positions of sensors are not carefully chosen except for tower fan fridge and water boiler As the tower fan and fridge are quiet they cannot be detected even when the sensor is just several centimeters away The water boiler is also quiet for a few seconds after it is turned on Therefore sensors are deployed close to these appliances As the bathroom has complex sound patterns we deployed two acoustic sensors in bathroom Note that they are not dedicated sensors because each of them can hear all appliances and water facilities in
46. readings through power lines to the TED5000 gateway which was attached on a power outlet and wired to the WiFi router TP Link WR740N to deliver readings to the base station computer However the router failed twice during the 10 days leading to disruptions to the collection of power readings We had to reset the router manually to restart the data collection We suspect that the failures were caused by bugs in the router As power readings are crucial information to Supero various improvements can be made in the future work For instance when the base station fails to receive power readings for a while it can raise alarm sound to remind the user to reset the router Communication performance The quality of wireless links Table I ENERGY BREAKDOWN IN THE 1 HOUR CONTROLLED EXPERIMENT IN APARTMENT 1 KAW measurements upero Name Power Energy Power Energy W kW h W kW h ight 0 030 4 0 0309 Light 2 0 0298 0 0300 Light 3 0 0300 0 0304 Light 4 0 0211 0 0210 Light 5 0 0207 0 0205 Water boiler 1472 1524 0 0490 1479 0 0456 Tower fan 23 40 0 0031 N A 0 0029 Rice cooker 498 0 0163 508 0 0168 Hair dryer 442 0 0158 462 0 0150 Fridge 117 146 0 0784 129 0 0841 Bath fan 60 0 0020 Router p 12 0 0142 3 Laptops 0 0468 36 0 0430 Oracle Baseline Error Power Energy Error Energy Error W kW h kW h 0 0 030 j 0 0310 F 7 150 0 0300 0 0305 23 3 153 0 0306 2 0 0 0307 2 3 Be 60 0 0210 0 5 0 0219 3 8 Be 100 0 0
47. s We first define the following notation e Nz and N4 are the total numbers of light and acoustic sensors Mz Ma and My are the total numbers of light acoustic and unattended appliances A denotes the absolute power change on the k light power event or acoustic transition e xi denotes the feature of sensor i in an event For light modality x is the absolute change of light intensity measured by sensor 7 which can be calculated from the current reading and the long term average For acoustic modality x is the acoustic feature sent from sensor i which is composed of the signal energies and zero crossing counts in the subbands For unattended power events by letting the index of smart meter be 0 xp 1For acoustic modality event refers to the start or end of a detected sound signal Acoustic event will be formally defined in Section V B 300 T T T T T a 8 A x 250 4 a i a amp 200 J 3 L E e VY 2 1504 a oe s PRE ST Ge A Light 3 100 I I L fi L 3 5 L L fi 50 100 150 200 250 300 350 4 4 5 5 5 5 6 Feature of Sensor 1 x1 In Distance from light source Fig 3 Light feature vectors of Fig 4 Light intensity vs distance two sensors cm in log scale Ax For light and acoustic modality the feature vector is defined as X x1 22 2N 7 where N Nz or N4 A Light Event Clustering Due to the ad hoc
48. s that the working states of each appliance can be accurately sensed This was achieved by installing dedicated sensors for each appliance 9 For instance magnetic sensors were carefully attached to the power cords of the monitored appliances and light sensors were positioned in close proximity to the monitored light and must not be triggered by other lights Such an approach incurs high cost sensor installations especially for the difficult to access appliances such as ceiling lights fans In contrast Supero allows ad hoc and non dedicated sensor deployment which can significantly reduce installation cost and improve usability II OVERVIEW OF SUPERO A Design Objectives and Challenges The main goal of Supero is to provide a fine grained electrical power usage report for a specified time duration in a household The report includes the details of the energy consumption of a particular appliance as well as when it was turned on and off Supero is designed to meet the following three objectives First the sensors should be deployed in an ad hoc and non intrusive manner A non professional should be able to deploy the sensors with intuitive instructions such as place a light sensor with no obstruction to lights and place an acoustic sensor on top of a microwave Second we aim to reduce the system configuration efforts by avoid ing labor intensive training and extensive user input Third Supero should be able to achieve long t
49. s the power of each appliance 7 denoted by pi by the regression arg ming yi P t doy Pisilla where s t is the state 0 or 1 of appliance i and P t is the total household power at time t In our evaluation the first baseline approach referred to as Oracle uses groundtruth information to generate the state of each appliance and then applies the regression in ViridiScope The results of Oracle allow us to evaluate the accuracy of Supero with respect to the state of the art approaches In the second baseline approach referred to as Baseline the state of each appliance is detected by the sensor closest to the appliance and then the regression in ViridiScope is applied In the implementation of ViridiScope each appliance s state was obtained by dedicated sensors that were carefully installed on the appliances 9 The results of Baseline will help us understand the challenges brought by the ad hoc sensor deployment B Controlled Experiment in Apartment 1 1 Experimental Settings All electrical appliances in Apartment 1 include 5 standing lights fridge water boiler 3 speed tower fan rice cooker bath fan hair dryer 3 laptop computers and a WiFi router The apartment uses natural gas range and steam based central heating unit that do not draw electrical power The deployment consists of 4 TelosB and 5 Iris motes The Iris motes only detect acoustic events The laptops and router cannot be easily detected by sensors Howe
50. t and acoustic events From the clustering result there is no dedicated light sensor i e each light sensor can detect multiple lights As almost all monitored appliances are hardwired to power lines we do not install KAW We conducted a controlled experiment for more than 5 hours Groundtruth information was manually recorded and then rectified by checking the total power readings In the experiment 40 false alarms out of total 127 light events were raised by light sensors where 38 false alarms are identified by the multi modal data correlation The left two false alarms are identified as outliers by the clustering algorithm Table III shows the results In a dining light event a sensor monitoring the dining light missed the event which results in a misclassification and introduces error to the energy estimate of dining light From the background cluster of unattended power events we observe an unknown appliance with a power of 140 W was turned on every about 10 minutes and lasted one minute It turns out to be a hot water dispenser on a sink Moreover it is the dispenser that caused a miss detection of guest bed light as they were once turned on off at the same time The average error of Supero is 6 1 which is consistent with the experiments in Apartment 1 E Sensor Placement As sensors are deployed in an ad hoc manner Supero is designed to be robust to the high variability in sensor place ment We deployed Supero in three more h
51. t features sensing models estimated sensor appliance distances and other prior information A Light Cluster Appliance Association The decay of light intensity follows the power law which can be exploited to associate light appliances and clusters However in complex household environment the decay of light intensity is affected by several factors such as the reflection of furniture and walls We conducted extensive measurements using light sensors to verify the decay model in various household environments Due to space limitation we only report one set of results Fig 4 plots the light intensity readings of a photodiode in a 5 x 3 2m living room where the line of sight distance between the light bulb and the sensor ranges from 60 cm to 3 m Both axes of Fig 4 are in log scale We have two observations from the figure First the linear relationship conforms to the power law Second at a certain distance from the light bulb the intensity measured by the sensor is proportional to the power of the light bulb Therefore we assume that the intensity measured by sensor i denoted by y is given by y BP dis where Pj is the power of light appliance j dij is the line of sight distance between sensor 2 and light appliance 7 a is the path loss exponent of the power law and is a scaling factor The a and 8 can vary with deployment environment but have bounded ranges The a typically ranges from 2 0 to 5 0 The is the ratio of
52. the battery voltage of the Iris with Alkaline batteries The tested Iris kept working from the 4 to the 9 day Regression analysis shows that the projected lifetime is 40 days by conservatively setting the MOV of Iris to be 2 2 V since the MOVs of the RF230 The TED5000 probe needs to be hardwired to the electrical service wires to get powered and connected to the gateway We are building a battery powered Zigbee smart meter based on TelosB and Fluke i2000 Flex current clamps to replace TED5000 which does not need the hardwiring on 3 3 T T te i T eE T_T T T T S 3 2 TelosB w Lithium 3 2 Iris w Alkaline 4 gt a LTelosB w Alkaline z4 L 4 BO 3 eee ee eae on L 4 23 ina NL 229 L 4 gt 5701 1 T A a E E 0 5 10 15 20 25 30 0 2 4 6 8 10 12 14 a Days b Days Fig 13 Battery voltage traces of TelosB and Iris radio chip and ATmega1281 8MHz MCU on Iris are 2 1 V and 1 8 V We note that the lifetime can be further extended by simply using Lithium batteries and duty cycling the CPU of motes X CONCLUSION AND FUTURE WORK This paper presents Supero a sensor system for un supervised residential power usage monitoring In Supero the multi sensor fusion can effectively reduce sensing errors in complex household environments By using unsupervised event clustering algorithms and a novel event appliance as sociation framework Supero can autonomously estimate the power and energy of each moni
53. the exact number of working states is optional Second for light modality Supero requires roughly estimated line of sight distances between sensors and lights which can be measured by a sonic laser tape arm span or even visual estimation As discussed in Section VI A Supero is robust regarding the distance estimation Third for acoustic modality Supero needs to know whether an acoustic appliance has a primary sensor All non primarily monitored acoustic appliances need to be sorted according to their powers cf Section VI B Such a ranking is usually straightforward based on common sense This can also be done based on their rated powers Finally Supero requires the rated powers of unattended appliances We note that the number of unattended appliances is small in a typical household Rated powers can be obtained by reading the labels on the appliances or from a database of rated powers The information described above can be easily obtained by non professionals and input to the system after deployment Supero only needs to be reconfigured occasionally e g when sensors appliances are relocated We have developed a web configuration interface using JavaServer Pages served by the base station computer to help the user input all the required information For instance Fig 7 a shows the configuration for the acoustic sensing where the user can input the acoustic sensor IDs appliance names and other information described in last paragrap
54. tic signals As a result on average more than 90 power consumption of a typical household can be captured by a combination of a smart meter and a set of light and acoustic sensors Useful prior information To avoid expensive in situ system training Supero leverages unsupervised learning techniques and a small amount of prior knowledge including rough sensor appliance distances and the rated powers of a small subset of appliances As the light acoustic signal decays with the distance from the source appliance the distances between sensors and appliances provide important hints for associating the detected events to the right appliances Moreover although the rated power of an appliance often has small discrepancy with the actual power consumption it helps identifying the consumption trace of a small number of difficult to detect appliances from the household power readings Rated powers are often available from the labels on the appliances or the user manuals Moreover there exist a few publicly available databases e g 20 which provide rated power based on the appliance brand and model Base Station Unsupervised Graphic Config Interface Cluster Appliance Gs 1 Light sensor Event Clustering Association distances 2 Acoustic appliances Multi modal Data Power Event fine grained properties Detecti power usage 3 Appliances rated Correlation SECUN powers events features power readings L
55. tored appliance Extensive evaluation in five homes shows that Supero estimates per appliance energy consumption with an average error of less than 7 5 Case studies demonstrate that the users can successfully deploy Supero in short time As discussed in Section III B electronics like TV and stereo consume 9 electricity in typical U S homes It is challenging to monitor the activities of many electronics as they emit either no or complex light acoustic signals and often consume variable power From the fact that a home typically has a limited number of high power electronics and most of them are plugged into power outlets a direct sensing approach e g by applying radio enabled KAW developed in this work can be integrated with Supero to achieve complete coverage of power usage monitoring Our evaluation does not cover all possible appliances due to the specific settings of the tested homes For instance in the apartments evaluated in this paper the laundry machines are shared among multiple homes and do not draw power from the tested apartments We also acknowledge that there may exist a small number of appliances that have highly complex signal characteristics and power consumption pro files e g a modern furnace with variable speed motor In our future work we will study these appliances and explore the use of other sensing modalities such as infrared seismic magnetic and imaging sensors to detect their activities REFERENCES
56. trolled experiments to evaluate the de ployments and configurations We generate fake total power readings according to the groundtruth to run the algorithms The event detection clustering and association results of the controlled experiment in both deployments are correct These two case studies show that the non professional users are able to quickly deploy Supero and ensure correct sensing results We also find that the non professional users tend to adopt the conservative deployment strategy discussed in Section VIH C F System Lifetime This section evaluates the lifetime of the battery powered Supero sensors In this experiment we force the CPUs of the motes to stay active even though they would operate in low duty cycles e g lt 5 for Iris in Supero The radios are turned on only when there are packets to transmit The TelosB motes report their battery voltages to the base station every minute Fig 13 a plots the battery voltages of two TelosB motes with Alkaline and Lithium batteries respectively over time The projected lifetime with Alkaline batteries is 79 days by conservatively setting the minimum operating voltage MOV to be 2 2 V although it is 2 1 V in datasheet 23 With the high capacity Lithium batteries there is no observable voltage drop in one month For the tested Iris mote we enforce it to always work in the fast sampling mode It piggybacks voltage reading to the acoustic feature packet Fig 13 b plots
57. upero the base station timestamps all sensor event messages and the real time power readings using its system clock The window based data correlation can fully tolerate small de lays of the real time sampling and event detection algorithms of sensors The appliances that cannot be easily or reliably detected by light and acoustic sensors e g rice cookers are referred to as unattended appliances A power event is regarded to be caused by an unattended appliance if there is no simultaneous light event or acoustic transition We refer to such power events as unattended events A challenge in power event detection is to deal with the power transients and delay effect of inductive capacitive loads We set a guard region centered at the time instance of the event and adopt the averages of the power readings before and after the guard region to calculate the power change Note that the length of the guard region dictates the time granularity at which Supero can differentiate two events happening close to each other In our implementation a guard region of 6 seconds well handles power transients and inductive capacitive delays and effectively identifies false alarms V UNSUPERVISED EVENT CLUSTERING A novel feature of Supero is that it automatically classifies the events detected by the sensors and associates them with the right appliances without any in situ system training This section presents the unsupervised event clustering al gorithm
58. ver as the router s rated power is known and it is 1 6 5555555 1 2 Sabari BE 1 2222223 i 0 8 oiei b Coni 0 6 B E E gi 0 4 Bt 0 2 mo Bi 0 Major PC Detection Total power kW i ms e 2 boiler T fridge water boiler rice cooker 20 30 40 50 60 70 Time minute Fig 9 Results of the controlled experiment in Apartment 1 1 The top chart shows the power readings labeled with groundtruth of events 2 The bars in the second chart show the detections of light sensors Two black bars at around the 35 minute are the false alarms labeled FA in the chart identified by multi modal data correlation Clusters are differentiated by colors and the overhead numbers are the IDs of the associated light 3 The third chart shows the major principle component given by PCA and the acoustic transitions detected by the window based algorithm The acoustic transitions with the same color are associated with the same appliance 4 The bottom chart shows the clustered and associated power events of the unattended appliances 8m Light 1A Ly kitchen counte 1 lege 1 Tower fan Node 1 Alight 3 Rice cooker bedroom Node 12 wg o r Node 2 bathroom Node4 Light 5A Bath Hair fan dryer Light 4 Dee CAe Legend TelosB Iris AO Appliances a b Light 2A Fig 8 a The power meter used to measu
59. vised algorithms work together to disaggregate the total household energy consumption to individual appliances As Supero does not require any post deployment in situ training it facilitates the deployment by non professional users We implemented Supero using TelosB Iris motes and a wireless power meter and evaluated Supero in five homes with significantly different square footage and electric power consumption A 10 day extensive evaluation in an apartment and a ranch house shows that Supero estimates the energy consumption with errors less than 7 5 Our results also demonstrate that Supero can be easily deployed by non professional users in short time II RELATED WORK This section discusses representative indirect sensing ap proaches for appliance power usage monitoring and identi fies the differences from Supero Early work 6 11 12 utilized per appliance power operating characteristics measured at the power panel to disaggregate the total energy consumption using pattern recognition algorithms To correctly identify appliances these approaches need either post deployment training 6 12 or a comprehensive database of priori power character istics of appliances 6 11 A recent paper 13 presented the experiences of monitoring power usage of a lab using 38 ACme meters 5 and 6 light sensors In 14 binary sensors were employed to help deploying power meters to estimate energy breakdowns for major devices in a bui
60. ynamics Light sensors may also pick up events unrelated to power consumption referred to as non power events such as those caused by human movement and opening closing window blinds The non power events will be identified by the multi modal data correlation discussed in Section IV D Light sensors sample light intensity periodically 4 Hz in our implementation and detect light events by an exponen tial difference filter EDF which is a lightweight and yet effective detection algorithm By denoting x n as the sensor reading at time step n the exponential moving average denoted by Z n is computed by Z n a x n 1 a Z n 1 where a 0 1 By setting a a or a a where s gt aj we have the short term and long term moving averages denoted by zs n and Z n The changes of n and z n capture the transient light changes and natural ambient light dynamics respectively Given two positive thresholds 77 and 7 the sensor counts the number of continuous samples satisfying s n Z n gt nz and raises a detection once the count exceeds 7 The sign of zs n z n indicates whether the appliance is turned on or off Whenever the sensor raises a detection it reports a light event message including current reading and the two averages Moreover it sets Z n Z n to quickly adapt the long term average to the most recent light intensities The sensor maintains a Gaussian noise model based on the re

A Sensor System for Unsupervised Residential Power Usage

Contents

Download Pdf Manuals

Related Search

Related Contents