Fusing Home Infrastructure Sensors with Smart HomeSensors for Automatic Fixture Level Disaggregation

ABSTRACTLight fixtures account for 11% of energy usage in homesand are good indicators of activities and room usage. De-spite their essential role in the home, however, it remains achallenge to monitor the usage of light fixtures. This paperpresents a new approach called LightSense that uses data fu-sion to cheaply and accurately infer light fixture usage. Itcombines data from a single power meter on the electricalmains and a light sensor in each room, neither of which issufficient alone. LightSense is evaluated using in-situ sen-sor deployments in 4 homes for 10 days each, monitoring 41light fixtures across a total of 27 rooms. It detects light fix-ture events with 81% recall and 86% precision, and estimatesthe total energy used by each fixture with 91.1% accuracy onaverage.

============================

Smart water meters will soon provide real-time access to in-stantaneous water usage in many homes, and disaggrega-tion is the problem of deciding how much of that usage isdue to individual fixtures in the home. Existing disaggrega-tion techniques require additional water sensing infrastruc-ture and/or a manual characterization of each water fixture,which can be expensive and time consuming. In this paper,we describe a novel technique called WaterSense to performfixture-level disaggregation using only a handful of inexpen-sive motion sensors. WaterSense automatically infers howmany fixtures are in each room, and how much water eachfixture uses. We evaluate the system using a 7-day in-situevaluation in 2 diverse multi-resident homes with a total of10 different water fixtures and 467 fixture events and observethat our approach achieves 86% classification accuracy inidentifying individual water fixture events and 80-90% ac-curacy in determining the water consumption of individualwater fixtures.

Author Keywords

ACM Classification KeywordsH.5.2 : .

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, orrepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.UbiComp ’12, Sep 5-Sep 8, 2012, Pittsburgh, USA.Copyright 2012 ACM 978-1-4503-1224-0/12/09...$10.00.

General TermsINTRODUCTIONSet up motivating scenario. Proliferation of smart meters andsmart home sensors. Need for fixture level usage and disag-gregation for energy and activity recognition applications.

Existing research focusses excusively on infrastructure sens-ing, or deploying specialized sensors per appliance alongwith infrastructure sensors. Existing techniques also requiresupervision. Lack of effective data fusion approaches to au-tomatically infer number of fixtures, usage times, and re-source consumption from existing infrastructure and distributedsensors. *also* - our data fusion approaches are not neces-sarily complete substitues for these other approaches, but areorthogonal and could be used in combination with more so-phisticated sensors to improve accuracy or reduce training.

we present a general framework for fusing data from infras-tructure and indirect sensors in homes based on our experi-ences with fixture disaggregation. highlight key challengesfaced and solutions used.

we evaluate our framework through two case studies - lightfixture and water fixture disaggregation. highlight any keydifferences of these two instances of the framework.

summarize deployments, evaluation results.

Light fixtures play an essential role in the home, and mon-itoring their usage can serve many purposes. For example,task lighting near a chair, countertop, or table can be a goodindicator of activities of daily living (ADLs) such as read-ing, preparing food, and eating dinner. In fact, light fixtureusage may be the primary indicator of certain activities, atleast from a sensing perspective. For example, the status ofthe light in a bedroom may be an indicator of either readingor sleeping. Monitoring light usage is also important froman energy conservation viewpoint. Light fixtures accountfor 11% of energy consumption in US homes on average[4]and 25% in California because heating and cooling energyis lower due to the mild climate [20]. This energy usage isequivalent to that used by 15 million homes or about 20%of all automobiles on the road in the US. Homeowners canreduce lighting energy by installing energy-efficient CFL orLED bulbs, daylight harvesting hardware, or occupancy-automatedlighting controls. However, these solutions are costly rela-tive to the energy saved: some of the homes we analyzedcontained over 50 bulbs, including specialized lighting (e.g.recessed lighting or under-cabinet lighting) that is partic-

ularly costly to retrofit. Homeowners need concrete dataabout individual light fixture usage to quantitatively assessthe expected return on investment of any given solution.

Unfortunately, light fixture monitoring remains a challengetoday. Non-intrusive load monitoring (NILM) approacheshave long been used to identify usage of major appliancesby analyzing their electrical signature, using only a singlepower meter on the electrical mains of the home. However,NILM is not effective for low-power appliances such as lightfixtures: the wattage of an incandescent bulb (30-100W) issimilar to the electrical noise levels of a typical home. Re-cently, several alternatives have been developed, but eachhas important limitations. For example, they require a sen-sor to be installed on every fixture [11, 19, 14, 12], requirea manual training and labeling period [17], or are limitedto light fixtures that involve switched-mode power supplies(SMPS) [8] such as those with CFLs or dimmer switches.

In this paper, we present a new approach called LightSensethat uses data fusion to cheaply, easily, and accurately in-fer light fixture usage. In addition to a power meter on theelectrical mains of the home, LightSense uses a light sen-sor in each room to capture both the electrical signature andthe light signature of each fixture. For a typical home, thisapproach requires only 3-8 more sensors than conventionalNILM techniques. Furthermore, light sensors are inexpen-sive and can be easily installed without touching the electri-cal wiring. However, light sensors alone are not sufficientto detect fixture usage: the light levels produced by a typi-cal fixture are similar to or even smaller than the light lev-els caused by window shades opening and closing, peoplewalking past the sensor, and clouds passing overhead. Lightfixtures account for only a small fraction of the light levelchanges in a room. To address this problem, LightSense ex-ploits several insights about light fixture usage: 1) noise onthe power lines and noise in light are typically independent2) light fixtures typically exhibit a similar power/light sig-nature over time, even when using dimmers 3) light usagealways occurs in ON/OFF pairs: a light cannot be turnedON twice in a row. In this paper, we present a probabilisticdata fusion framework that combines power and light datastreams, automatically learns the long-term signature of eachlight fixture, and accurately recognizes the real-time usageevents.

LightSense is completely unsupervised: it does not require atraining process in which users manually label light fixtureusage. Furthermore, it does not require user configuration:users do not need to indicate the number of fixtures in a homeor their individual wattages. These parameters are automati-cally inferred. We envision two types of LightSense deploy-ments. In a short-term energy audit, light sensors would beplaced on a shelf or countertop in each room for a period ofseveral days or weeks to assess the total energy usage of eachlight fixture. Power data would be collected from a smartpower meter, which are already deployment in many partsof the US by power utility companies for billing purposes.For longer-term activity recognition applications, the lightsensors would simply be integrated with other sensors. For

example, several commercial motion sensors already haveintegrated light sensors for calibration purposes [?, ?]. If amotion sensor is placed in every room for activity recogni-tion purposes anyway, the light sensor data can be used byLightSense to infer light fixture usage.

We evaluate LightSense with an in-situ deployment of powermeters and light sensors in four multi-resident homes over10 days each, monitoring a total of 41 light fixtures across32 rooms. We deployed wireless light switches and smartplugs to collect ground truth information about light fixtureusage. We found that the power meter alone detects lightfixture events with only 15% precision, and the light sen-sors alone achieve only 25% precision. By combining thetwo data streams, however, the LightSense data fusion al-gorithm detects ON/OFF events with 81% recall and 86%precision. We also demonstrate that LightSense can disag-gregate whole-house energy usage into fixture-level energyusage, estimating the total energy usage of most the com-monly used light fixtures with 91.1% accuracy on average.

===================

The world’s usable water supply is decreasing at a faster ratethan it can be replenished. Household water conservationis important to ensure sustainability of fresh water reserves,to save energy from water treatment and distribution, and toprevent fresh water habitats from being affected through ex-cessive water use [?]. Residents have a number of practicaloptions to conserve water, ranging from replacing high flowtoilets and showers with low flow replacements, to reduc-ing water used for daily acitivites such as brushing teeth orwashing dishes. To make informed decisions that maximizewater savings, households first need a detailed understand-ing of how much water is used by each appliance and waterfixture in the home.

Water utilities are increasingly installing smart water metersthat provide real-time access to household water consump-tion, and 31 million smart water meters are expected to beinstalled by 2016 [?]. However, these meters are installedat the water mains and only provide aggregate water usage,primarily for billing purposes. Disaggregation is the prob-lem of deciding how much of that usage is due to individualfixtures in the home, and is challenging because homes of-ten have multiple sinks, toilets, showers, and other fixturesthat produce similar rates of flow. For this reason, existingdisaggregation techniques require additional sensing on thewater piping infrastructure, and/or a manual characterizationof each water fixture [5, 6, 7, 13, 16]. These techniques canbe expensive, difficult to deploy, and time consuming.

In this paper, we present the WaterSense system that per-forms fixture-level disaggregation of smart water meter datausing only simple motion sensors. Motion sensors are in-expensive ( $5 each), easy to install, and already prevalentin many homes as part of home security or home automa-tion systems [?]. WaterSense does not require any additionalsensing infrastructure on the water pipes or fixtures, and dis-aggregates fixtures in an unsupervised manner that does not

Technique/ Feature Easy to Install? Cost Energy feedback? Accuracy User training? Applicability? In-site evaluation?

Electrisense Yes High No High No CFLs No

Using mechanical properties of switches

Yes High No High Yes Incandescent No

NIALM Yes Low Yes Very Low Yes All Yes

Direct sensing No High Yes Very High No All Yes

Ambient sensing No High Yes High No All No

Light Sense Yes Low Yes High No All Yes

Figure 1. Comparison of existing techniques to infer usage and power consumption of light fixtures

require the collection of training data. The WaterSense tech-nique is based on two basic insights: 1) fixtures with similarflow signatures (e.g. identical toilets) are often in differentrooms, and 2) fixtures in the same room often have differ-ent flow signatures (e.g. a toilet vs. bathroom sink). Basedon these insights, WaterSense clusters all water usage eventsbased on both flow signatures and motion sensor signatures,and each of these clusters represents a unique water fixturein the home. One limitation of this technique is that it isnot likely to differentiate two identical fixtures in the sameroom, such as double sinks in a bathroom. However, suchdistinctions may also be less important for the purpose ofwater conservation decisions. We use a novel Bayesian clus-tering algorithm to create robust clusters despite noise in themotion sensor data caused by both wireless packet loss andmutliple residents moving in multiple rooms simultaneously.We deployed the WaterSense system in two different, naturalhome environments for 7 days each and found that the sys-tem can disaggregate flow at the water mains to individualfixtures with an average 86% classification accuracy.

RELATED WORKSeveral approaches have been developed for monitoring lightfixture usage, but each has important limitations. Figure 1lists the main techniques in this area, and compares them toLightSense in terms of several important metrics.

Non-Intrusive Appliance Load Monitoring (NIALM) [9] tech-niques can be used to identify the usage of electrical fixturesin the home based on electrical signatures. These signaturescan be extracted using only a single power meter [?], whichis already available in many homes today [2]. Some appli-ances have a unique profile of real and reactive power, whileother appliances such as washing machines and dishwash-ers exhibit characteristic electrical patterns over time. How-ever, NIALM techniques not effective for light fixtures thatexhibit constant, low power values due to the large numberof similar, low power appliances in homes, and due to lowpower state transitions from complex appliances such as thetelevision or the HVAC system [9].

An alternative to NIALM is use in-line current sensors to di-rectly measure the amount of energy consumed by individ-ual appliances. Examples include pluggable power meters[?, 11] to monitor lamp fixtures, and smart switches [?] tomonitor lights controlled by wall switches. However, directsensing a large number of sensors to be integrated with thehome’s electrical wiring, which is costly in terms of both

hardware and installation time. In our own deploymentsacross four homes, for example, installing in-line currentsensors for ground truth consumed approximately 19 manhours per home on average, while the LightSense installa-tions only consumed an average of 25 minutes per home.

Recently, new solutions have been developed that requireonly a single sensor on the power line. Gupta et al pro-pose the Electrisense system [8], which uses an easy to in-stall, plug-in sensor that leverages unique high frequencyEMI(Electromagnetic Interference) signals on the power lineto identify light fixtures. However, this approach is lim-ited to appliances that use switched-mode power supplies(SMPS) such as lights with CFL bulbs or dimmer switches;it does not work for conventional incandescent bulbs. Patelet al [17] propose a related approach to identify mechani-cally switched appliances in homes based on a unique highfrequency voltage transients. However, this approach re-quires users to manually train the system by labeling ONand OFF events of individual appliances. If lights are addedor moved throughout the house, training must be performedagain. Furthermore, the authors only evaluate the ability toclassify manually labeled events that are explicitly input tothe classifier(recall and classification accuracy). It is un-clear if the reported upper bound classification accuracy forthese approaches of 75-95% will be retained in a realistic,in-situ evaluation, where false positives in the power linesignals could reduce precision. Moreover, neither of theseapproaches can currently infer the energy consumption ofindividual fixtures; they can only identify which fixture wasused. LightSense is a completely unsupervised techniquethat can identify fixtures and measure their energy consump-tion. LightSense is also evaluated in a in-site experiment.

The systems most similar to LightSense include Ambientsensing approaches that fuse data from whole-house powermonitors with environmental sensors deployed throughoutthe house. For example, Viridiscope [14] uses one special-ized sensor node per appliance containing light, acoustic,and magnetic sensors with whole house power meter data toinfer the energy costs of individual appliances in the home.However, the Viridiscope algorithm makes the strong as-sumption that the power consumed by unmonitored appli-ances is constant, essentially requiring complete coverage ofevery electrical fixture and appliance in the home. A con-tactless sensing system proposed by Rowe et al [19] also re-quires specialized environmental sensors per appliance, andadditionally uses high frequency circuit level power meter

data, which requires high installation effort at the breakerbox. Both of these approach require far more sensors and amore involved installation process than LightSense. Light-Sense can detect light fixture usage without the need to mon-itor any other fixtures or appliances in the home. Further, itdoes not require a sensor on each fixture, but only requiresfull coverage of all light fixtures of interest (typically, onesensor per room).

=============

Several approaches have been proposed in the literature toinfer fine-grained water fixture use in homes. The most ba-sic approach is to use flow signatures, such as flow rate, flowduration or, in the case of washing machines and dishwash-ers, patterns of flow to identify types of fixtures and appli-ances [16]. However, flow signatures alone cannot disam-biguate between different instances of identical sinks, toilets,or showers in the same home. Fogarty et al [5] use micro-phones installed in a basement to classify water fixture use,and achieve good accuracy in identifying high consumptionappliances, but low accuracy in differentiating between dif-ferent instances of the same fixture category. Another recentapproach [13] uses vibration sensors on pipes to disaggre-gate total flow measured at a central location in a novel, un-supervised manner. Both of these techniques, however, re-quire additional sensing of the water infrastructure, whichmay require access to pipes in crawl spaces or walls. Fur-thermore, microphones and accelerometers are more powerintensive than motion sensors and would either have a shortbattery life or would require wired power. Finally, micro-phones may be considered privacy-invasive sensors.

Patel et al [6] avoid extensive sensing by using a single waterpressure sensor that samples at 500Hz, plugged into a freespigot or water outlet in the home. However, this approachrequires significant training data on the order of several daysin a real world setting [7] to achieve high accuracy in in-ferring individual water fixture events. Our approach is anunsupervised technique that does not require training data.

A DATA FUSION ALGORITHM(flow diagram architecture figure)

Tier I: Event DetectionDo automatic edge detection on each infrastructure and in-direct sensor. highlight key common challenges.

Light Fixturesdiscuss briefly about edge detection challenges for light sen-sor and power meter.

Water Fixturesdiscuss briefly about edge detection for water flow meter andmotion sensor.

Tier II: Data FusionMotivate two main reasons why data fusion is important: 1.eliminate false positives 2. infer correct power/water con-sumption for a fixture event for energy applications

use temporal distance as a criteria for data fusion, since theassumption is that fixture usage causes both indirect and in-frastructure sensor events within a small time window. butindirect sensor events such as motion sensors are not alwaysin sync with infrastructure events , or exact time sync maynot be possible because infrastructure and smart home sen-sor might be from different systems.

Key challenge: dealing with a large number of possibili-ties for computing ¡infrastructure sensor,direct sensor¿ eventpairs due to high false positive rates from either direct sen-sors or infrastructure sensors or both.

Solution: Bayesnet approach to automatically keep track ofcommonly occuring pairs of ¡infrastructure, direct sensor¿events incorporating both instantaneous and historical sensorevidence

Light Fixturesdata fusion eliminates many false positives from both lightsensor and power meter. bayesnet addresses high false posi-tives from power meter and computing correct ¡light,power¿event pairs

(example data figure from light sense)

Water Fixturesbayesnet helps disambiguate from simultaneous events frommotion sensors in fixture rooms.

(example data figure from water sense)

Tier III: Event Matchinggoal: match ON and OFF fixture usage events. discuss bi-partite matching approach.

Light Fixturesuse both probability of power and light reading matchingalong with joint prob from bayesnet in tier II

Water Fixturesmatching is more trivial since simultaneous use is rare

Tier IV: Fixture Clusteringgoal: automatically compute set of fixtures sensed by eachindirect sensor using clustering approaches.

Light Fixturesgives set of light fixtures. possible to compute nominal wattage,power consumption, usage.

Water Fixturesgives set of water fixtures. use simple heuristics to classifyas sink or flush or shower.

A CASE STUDY: MOTION AND LIGHT SENSORS

Inferring Water Fixture Usage

Inferring Light Fixture Usage

Tier I: Light

Edge Detection

Light Sensors

Data

Power Meter

Data

Tier I: Power

Edge Detection

Tier II: Data Fusion and

Bayesian Matching

Tier III: Fixture

Identification

Power Edges Light Edges

Room Level ON-OFF Events

Fixture Level ON-OFF events

Figure 2. LightSense users a 3-tier data fusion framework to combinedata streams.

LIGHTSENSE DESIGNThe LightSense algorithm uses a three-tier data fusion frame-work to combine raw light sensor and power meter data andgenerate the usage times and energy costs of individual lightfixtures. The framework is illustrated in figure 2. The Tier Iapplies edge detection algorithms to identify both rising andfalling edges in both data streams. LightSense uses severalnovel edge detection algorithms tailored to the noise typicalof light and power timer series. Tier II combines the edgesdetected in both data streams to generate a single series ofevents. Tier II applies three basic principles to decide whichedges to keep: a) the edge must appear in both the light andpower data b) the light and power values must be typical forthat room c) there must be a matching between rising edgesand falling edges. Once the light fixture events are identi-fied, Tier III uses a clustering algorithm to infer the individ-ual light fixtures present in each room, their usage times, andtheir total energy costs. Figure 3 shows a sample data tracefrom 6AM to 4PM in one home. The output of Tier I pro-duces a large number of false edges for both the light andpower data, and the data fusion algorithm in Tier II elimi-nates almost all of them to identify only 4 correct ON/OFFevents.

Tier I: Light and Power Edge detectionThe goal of Tier I is to detect rising and falling edges inthe power and light data streams, despite several types ofnoise. The inputs to Tier I are the time series data fromeach light sensor i denoted by Li, and the time series datafrom the whole house power meter P . Using this input, theedge detection algorithms output a sequence of rising andfalling edges for each light sensor i, denoted by RLi andFLi, respectively, and a sequence of rising and falling edgesfor whole house power consumption, denoted by RP andFP , respectively.

Power Edge Detection

To detect whole house power consumption edges from thepower meter time series P , we apply a sliding window basededge detection algorithm, that computes all possible poweredges with a maximum window bound. In particular, eachpower edge e is defined by a time window [es, ef ], and theedge value is given by ev = (P [ef ] − P [es]). We placetwo main constraints on time windows which can constitutea power edge e:

1. (ef − es) ≤ maxwinP2. |ev| ≥ dP

Condition 1 ensures that all edges have a time window lengthbounded by maxwinP . A small parameter for maxwinPparameter ensures that we eliminate a significant number ofslow power intensity changes that do not originate from artif-ical lighting. In our current implementation, we fixmaxwinP =5 seconds across all homes. Condition 2 determines the low-est wattage light fixtures detected by LightSense, parame-terized by dP ; in section , we evaluate the recall precisiontradeoff as dP is varied.

Our power edge detection algorithm is a multi-edge detec-tion approach, since multiple power edges could simulta-neously be detected at any time instant from the differenttime window lengths bounded by maxwinP . We adopt thisapproach because the power meter data typically containsseveral closely spaced simultaneous low power events, anda single edge window is insufficient to achieve high recall;a larger edge window is ineffective when there are closelyspaced events ON or OFF events, while a smaller edge win-dow does not capture slowly increasing power events fromdimmable lights as reported by the power meter, which re-ports rms power every second from the AC current and volt-age inputs. We address the problem of pruning power edgesin section , by leveraging the joint light and power signa-tures of light fixtures. Also, we bound the maximum poweredge magnitude |ev| to 1000W to prevent interference fromhigh power appliances such as the stove or the washing ma-chine. The edges output by our edge detection algorithm arepartitioned to rising and falling power edges RP and FP ,respectively, and the edge value ev is set to |ev| in the parti-tioned subsets.

Figure 3 shows the timestamps of edges output by the poweredge detection algorithm. We observe that the power edgedetection approach shows poor precision in detecting lightfixture events, due to false positives from numerous low powerappliances, and low power state transitions of complex ap-pliances such as the HVAC system or the television.

Light Edge DetectionTo detect light edges from each light sensor stream, we de-sign the light edge detection algorithm outlined below. Fig-ure 3 shows the timestamps of rising(ON) and falling(OFF)edges output by the two steps in the light edge detection al-gorithm. The first step uses a sliding window-based edgedetection algorithm similar to the power edge detection algo-rithm to eliminate noise due to slowly changing light events,such as daylight changes from the movement of the sun. Thesecond step uses three novel adaptive noise filters to filter the

typical noise events seen in home light sensors.

for each light sensor i ∈ SLSi = medfilt(Li, window)

Ei = windowdetect(LSi,maxwin, dL)

F i = null

for each e ∈ Ei

b = noise filter(e) ∧ onoff filter(e) ∧ consensus filter(e)

if b == 1 then add e to F i

partition F to (RL,FL)

In the algorithm above, for every light sensor i ∈ S, whereS is the set of light sensors in the home, we first performmedian filtering on the light sensor time series Li with afixed window size of 10 samples across all four homes. Me-dian filtering eliminates impulse noise from our light sen-sors. We then apply a window-based edge detection algo-rithm windowdetect on the smoothed time series LSi tooutput a sequence of edges Ei. This algorithm uses the twoconditions mentioned above for the window-based poweredge detection algorithm, with new parameters dL andmaxwinLinstead of parameters dP and maxwinP . Setting a smallbound on the maxwinL parameter ensures that we elim-inate a significant number of slow light intensity changesthat do not originate from artifical lighting. In our currentimplementation, we fix maxwin = 2.5 seconds across allhomes, since that is the maximum time required for typicalresidential dimmer switches to reach their target light inten-sity level. Paramter dL determines the lowest intensity oflight edges detected by LightSense; in section , we evaluatethe recall precision as dL is varied. We also add two addi-tional constraints on edges e ∈ Ei to ensure that multiplelight edges are not detected simultaneously:

1. ∀d ∈ Ei, (e 6= d)→ ([es, ef ] ∩ [ds, df ] = {})

2. ∀[x, y], (|Li[y] − Li[x]| ≥ dL) ∧ ([x, y] ⊂ [es, ef ]) → (ev ∗(Li[y]− Li[x])) > 0

The above conditions together ensure that the edge windowsizes are adjusted automatically depending on the variationin the light signal; smaller edge windows are used when lightintensity changes are dense, while larger edge windows areused when the light intensity changes are sparse. Due tospace constraints, we do not discuss how the windowdetectalgorithm ensures the constraints above to output Ei, but wenote that it is a straightforward algorithm that is in O(|Li|).

For each edge e ∈ Ei output by the windowdetect algo-rithm, we apply three additional adaptive filters to eliminatenoisy edges, namely the noise filter, the onoff filter, and theconsensus filter; in figure 3, we see that these adaptive filterseliminate a significant number of noisy edges output by thewindowdetect algorithm.

Firstly, lights do not turn on and off very frequently; if wedetect a large number of similarly sized light edges, we as-sume they are ambient noise. For each edge e, the noise filteralgorithm first counts the number of comparable, temporallyadjacent edges x within time dTa of edge time es, whosemagnitude is at least α times the current edge value ev un-der consideration. The filter returns false if the number of

kcl

kcp

kl

kp

Figure 4. The lpbayes Bayesnet finds the maximum likelihood clusters(top nodes) given the power and light edges (lower nodes).

such comparable edges exceeds a maxcount parameter. Wefix parameters dTa = 20 minutes, maxcount = 20, andα = 0.8 across all homes.

Secondly, when people or clouds pass by the light sensor, arising edge and falling edge occur closely together in time;we filter these pairs of events out. For each edge e, theonoff filter algorithm returns false if an opposing, compara-ble, temporally adjacent edge y exists within time windowdTb of edge time es, whose magnitude yv is at least α timesthe current edge e under consideration, and whose polarityis opposite to edge e. We fix parameter dTb = 12 seconds.

Thirdly, redundant light fixture events from adjacent roomsare detected in multiple light sensors. So, when we ob-serve light edges in multiple light sensors simultaneouslywith the same polarity, we only retain the light edge withmaximum intensity, since that edge is most likely from thesensor in the same room as the light fixture. Assuming acoarse time synchronization among the multiple light sensorstreams based on timestamps at the receiver base station, theconsensus filter eliminates redundant light edges within onesecond of each other. Finally, the filtered light edges are par-titioned to rising and falling edges based on their polarity toRL and FL, respectively.

Tier II: Data Fusion and MatchingThe goal of Tier II is to combine the low precision light andpower edges from Tier I to compute a sequence of matchedON-OFF events M i for each light sensor i. Each ON-OFFevent m ∈M i is defined by a four tuple (ms, mf , mp, ml);ms andmf denote the ON and OFF time, respectively, whilemp and ml denote the average power consumption and lightintensity increase for event m. Firstly, we use data fusion toeliminate temporally isolated power or light edges as falsepositives. Secondly, we use Bayesian matching to prune un-matched ON or OFF light and power edges as false posi-tives. Finally, we use a novel Bayesian clustering approachto accurately assign power costs to each ON-OFF event, byautomatically learning the typical power and light signaturesin a room.

Data FusionThe Data Fusion step combines the rising and falling poweredges {RP,FP} and light edges {RLi, FLi} to compute aset fused light-power edges {RLP i, FLP i} for each lightsensor i ∈ S. We first describe how the rising light-power

Raw Light Data

Power Consumption in Watts

Light Edges

Filtered Light Edges

Power Edges

Data Fusion and Bayesian Matching

Matched Light Power Edges

Power Edge Detection

Window based edge Detection

Adaptive filters

ON

OFF

ON

OFF

ON

OFF

ON

OFF

Figure 3. Sample Data Trace from a bedroom light sensor and whole house power meter from 4AM to 4PM in House 2 along with ground truthfixture events. LightSense eliminates false positive light and power Edges by performing data fusion and matching.

edgesRLP i are computed fromRP andRLi. We first com-pute a power edge set PSe for each light edge e ∈ RLi,where PSe denotes all the power edges from RP withintime dTadd of the light edge timestamp es. If no poweredges are found, we discard the light edge e, otherwise weadd the light-power edge (e, PSe) to RLP i. Thus, tempo-rally isolated light or power edges are eliminated as falsepositives, while temporally adjacent light and power edgesare added to RLP i. We similarly compute the falling light-power edges FLP i from FP and FLi. We currently setdTadd to 12 seconds across our deployments to account fortime synchronization errors between the light sensor and powermeter data streams. Since the power meter data is noisy, mul-tiple power edges in PSe are typically associated with a sin-gle light edge e; the problem of assigning the correct poweredge to each light edge is addressed as part of our BayesianClustering step below.

Bayesian MatchingFor each light sensor i ∈ S, the Bayesian matching stepmatches rising light-power edgesRLP i to falling light-poweredges FLP i to compute the sequence of matched light fix-ture events M i. Bayesian matching prunes false positiveedges from RLP i and FLP i that cannot be matched withhigh probability. The algorithm is shown below:

for each light sensor i ∈ SWeightsi = null

for each edge (d, PSd) ∈ RLP i

for each edge (e, PSe) ∈ FLP i

c1 = (ds > es)

c2 = min(Li[ds : es]) > γ ∗ dvif (c1 ∨ c2)pmatchid,e = 0

elsecompute lprobid,e, pprob

id,e

pmatchid,e = lprobid,e ∗ pprob

id,e

Weightsi[d, e] = −log(pmatchid,e)

Mi = Hungarian(Weightsi) with weight threshold wmax

The Bayesian Matching algorithm shown above first con-structs a weighted bipartite graph with edges from the set ofrising light-power edges RLP i, to the set of falling light-power edges FLP i. The edge weight Weightsi[d, e] be-tween any rising edge d ∈ RLP i and any falling edge e ∈FLP i is set to −log(pmatchid,e), where pmatchi(d,e) rep-resents the probability of edge d being matched to edge e.We perform a min-cost bipartite matching on the weightedbipartite graph represented by Weightsi, which returns anoptimal matching M i with maximum match likelihood, i.e.M i = argmaxx(

∏(d,e)∈x pmatch

id,e).

In the algorithm above, we see how the match probabilitiespmatchid,e are computed. Firstly, we set pmatch

id,e = 0

if any of the two conditions c1 or c2 are satisfied. Condi-tion c1 ensures that rising edges in an ON-OFF event occurbefore falling edges. Condition c2 ensures that rising andfalling edges from two different ON-OFF event pairs are notmatched together by by leveraging the observed additive na-ture of light intensities in a room; if a +100 rising edge and a−100 falling edge are separated by a time interval where thetotal light intensity is only 5, then the two edges are likelyfrom different ON-OFF events. In the algorithm, γ is setbased on empirical experiments to 0.8 across all our deploy-ments. If conditions c1 and c2 are false, the two edges d, eunder consideration could potentially be matched.

As shown in the matching algorithm, the match probabilitiespmatchid,e are computed as lprob

id,e ∗ pprobid,e, where the

light match probability lprobid,e denotes the probability thatthe light edges d and e are matched, while the power matchprobability pprobid,e denotes the probability that the two cor-responding power edge sets PSd and PSe are matched.

We compute the light match probability lprobid,e as the like-lihood that ON and OFF light edges d, e belong to the same

light fixture; we observe that ON and OFF light edges fromthe same two-state light fixture have similar edge values. Toleverage this observation, we first cluster all the light edgevalues from RLP i and FLP i using the QT (Quality Thresh-old) clustering algorithm [10] with a relative distance errorthreshold of 0.25 on the light edge values. QT Clusteringoutputs a set of light edge clusters CLi with a fixed clusterassignment for each edge. We assume each light edge clus-ter has a normally distributed light intensity distribution, andcompute the mean and standard deviation for each light clus-ter cl ∈ CLi as clmean and clstd, respectively, from the edgesassigned to cl. The probability lprobid,e that the two lightedges d, e belong to the same light cluster cl ∈ CLi is thencomputed as:

lprobid,e =∑

cl∈CLiN (dv , clmean, clstd) ∗ N (ev , clmean, clstd) (1)

Computing the power match probabilities pprobid,e is lessstraightforward since there are typically multiple rising andfalling edges in the power edge sets PSd and PSe. To sim-plify this problem, we choose the assignment x ∈ PSd, y ∈PSe that maximizes pprobid,e as follows:

pprobid,e = maxx,y(vprobix,y ∗ pi(x, dv) ∗ pi(y, ev)) (2)

In equation 2, the power value match probability vprobix,ydenotes the probability that the two power edges x, y arefrom the same light fixture, while the joint light-power prob-abilities pi(x, dv), pi(y, ev) denote the probability that a lightfixture with light edge value dv or ev has a power consump-tion of x or y, respectively. To compute the power valuematch probability vprobix,y , we leverage the observation thattwo-state light fixtures produce ON and OFF power edgeswith similar values. We cluster all the power edge valuesin edge sets from RLP i and FLP i using the QT (QualityThreshold) clustering algorithm with a relative distance er-ror threshold of 0.25 to obtain the set of power edge clustersCP i; from CP i , vprobix,y is computed similar to lprob

id,e

in equation 1. In equation 2, the light-power probabilitypi(x, dv) denotes the joint probability of observing poweredge value x together with light edge value dv from lightsensor i. We use the lpbayes Bayesian Clustering approachdescribed below to compute the joint light-power probabilityvalues.

Bayesian ClusteringOur Bayesian Clustering approach computes the joint prob-ability pi(pk, lk) of observing any power edge value pk andany light edge value lk together in the same light-power edgefrom light sensor i ∈ S. The joint probability must effec-tively capture the typical light and power signatures that oc-cur together in a room. For example, consider an exampleON-OFF match from a 100W light bulb, containing a ris-ing edge of d = (40L,[40W, 103W, 500W]) and a correspondingfalling edge of e = (41L,[40W, 95W, 1000W]) sensed by lightsensor i. We expect lprobid,e to be high for this match, since

the two light edges are very likely to be from the same 40Llight cluster. Among the multiple power edge assignmentspossible, the highly likely assignments based on vprobi fromequation 2 are (40W, 40W ) and (103W, 95W ). If we useonly vprobi, we incorrectly assign the (40W, 40W ) poweredges to the ON-OFF event, since they are more likely to befrom the same cluster than the (103W, 95W ) edges whichare more noisy. However, by leveraging the joint probabil-ity of occurance of the light and power meter data streamsusing lpbayes, we still make the correct 100W assignmentto the ON-OFF match because the 103W event has a muchhigher likelihood than a 40W event of co-occuring with the40L light event, i.e. pi(103W, 41L) >> pi(40W, 41L) .

Our Bayesian Clustering approach uses the lpbayes Bayesnetshown in figure 4 to compute the joint probabilities pi(pk, lk);in the Bayesnet, clk and cpk are hidden random variablesthat denote the cluster to which the light edge value lk andpower edge value pk belong to, respectively. The Bayesneteffectively captures the probabilistic relationship betweenindividual light or power values and their cluster assignments,and more importantly, the conditional relationship betweenthe light and power clusters, indicating how likely an ele-ment from power cluster cpk occurs given that an elementfrom light cluster clk has occured. From the basic propertiesof Bayesnets, the joint light-power probabilities pi(pk, lk)are computed as follows:

pi(pk, lk) =∑clk

∑cpk

p(lk|clk) ∗ p(pk|cpk) ∗ p(cpk|clk) (3)

The conditional probabilities p(lk|clk) and p(pk|cpk) are com-puted using the Gaussian distributions computed for the lightand power edge clusters during the QT clustering process de-scribed in the previous section, i.e. (p(lk|clk) = N (lk, clkmean, clkstd)and (p(pk|cpk) = N (pk, cpkmean, cpkstd) . The conditional proba-bility P (cpk = a|clk = b) is computed by frequency countingthe number of individual light edges z assigned by QT clus-tering to light cluster b that have at least one edge in PSzassigned to power cluster a. In other words, if we denote theset of light-power edges assigned to light cluster b as CAb,and the set of light-power edges in which at least one ele-ment in the power edge set is assigned to power cluster a asCAa, then:

p(cpk = a|clk = b) = |CAa ∩ CAb|/|CAb| (4)

The power assignments x, y that maximize equation 2 arefixed as the optimal power assignments for the current match(d, e) under consideration. From equations 1, 2, 3, and 4,we calculate the match probabilities pmatchid,e. After com-puting the match probabilities, the edges from RLP i andFLP i are matched using the min-cost bipartite matching al-gorithm (Hungarian algorithm [15]). We limit the maximummatch weight wmax to 70 in our deployments to eliminatehighly noisy matches which are unlikely to be true light fix-ture events. For each match m ∈ M i, the start and endtimes ms and mf are set to the timestamps of the rising and

House # # of residents

# of rooms

# of light fixtures

House Type # of rooms with windows

1 3 8 12 3 floor House 8

2 3 6 6 3 bedroom student apartment

6

3 2 8 9 1 bedroom condo

6

4 4 9 14 3 floor house 7

Figure 5. Details of the four homes chosen for deploying LightSense

falling edge, respectively. The light intensity value ml andthe power value mp are set to the mean of the light and opti-mal power edges in the rising and falling edge of match m,respectively.

Tier III: Fixture IdentificationThe goal of Tier III is to identify the individual light fixturesin each room, their nominal wattage values, usage times andtotal energy costs. Using the ON-OFF matches M i fromTier II as input, Tier III outputs a set of light fixtures LF ifor each light sensor i; for each light fixture lf ∈ LF i weoutput: (a) the sequence of ON-OFF event times given bylfm ⊆ M i, (b) the nominal wattage value lfw, and (c) thetotal energy consumption lfe.

To compute LF i, we cluster the ON-OFF matchesM i basedon the match light intensity ml using QT clustering with arelative distance error threshold of 0.25. For each light fix-ture cluster lf ∈ LF i output by QT clustering, we compute(i) the usage times based on the start and end times of thematches lfm ⊆ M i assigned to cluster lf , (ii) the total en-ergy consumption based on the ON-OFF event timings andpower consumptions as lfe =

∑m∈lf mp ∗ (mf − ms),

and (iii) the nominal wattage value of the light fixture asthe median of the power assignments mp made to matchesm ∈ lf . We perform a simple threshold based filtering toeliminate any degenerate light fixtures df ∈ LF i, if the num-ber of events assigned to the light fixture df is less than 2,or if the total energy contribution of the light fixture to theroom containing the light sensor i is less than � %, i.e. ifdfe < (�/100) ∗ (

∑y∈LF i ye).We set � to 10% in our imple-

mentation across all deployments. The resulting set of lightfixtures after eliminating degenerate light fixtures is outputas the final set of light fixtures LF i in the room containinglight sensor i.

EXPERIMENTAL SETUPIn this section, we discuss details of our underlying sensingand ground truth hardware, and their deployment in homes.

Physical Sensor ComponentsOur WaterSense approach requires a water flow meter at thewhole house water input line, and also motion sensors ineach room that contains water fixtures. In our home deploy-ments, we use a Shenitech Ultrasonic water flow meter [?]

Tier 1: Detecting

Water Flow Events

Tier 2: Creating Room Clusters

Water fixture activities

Water flow data Motion Sensor

Data

Activity Room Assignments

Tier 3: Differentiating Fixture Types

Fixture level usage information

Figure 8. WaterSense uses a three tier inference algorithm to find 1)water flow events 2) clusters of flow volumes that often co-occur withcertain motion sensors, and 3) different types of fixtures in the sameroom.

that uses the doppler effect to measure the velocity and re-sulting flow of water through the pipeline. The flow meterreports instantaneous water flow (in cubic meters per hour)at a frequency of 2Hz using the home’s Wifi connection totransmit data. We expect that utility water flow meters beingdeployed in a large scale [?] in homes will have a similarsetup. Figure ?? shows the installation of the flow meter inone of our home deployments. In addition to the flow meter,WaterSense requires at least one motion sensor in each roomcontaining water fixtures. In our deployments, we use off theshelf X10 motion sensors [?] inside rooms to detect occu-pancy, as seen in figure ??. The X10 motion sensors send abinary ON message whenever motion is seen with a minimaldamping interval of 7 seconds between ON messages.

Real World DeploymentsTo evaluate the LightSense system, we deploy it in four multi-resident homes for 10 days each. We use the TED powermeter [3] along with the cheap Hamamatsu photo diodes [1]interfaced with the telosb mote [18], as shown in figure 7aand figure 7b. The power meter samples at 1Hz to match thecapabilities of prevalent smart utility meters, while the lightsensors sample only at 2Hz to have a minimal impact onthe lifetime of the sensing node. To measure ground truth,we replace the existing switches in each home with wire-less ZWave smart switches [?], and connect each lamp to awireless ZWave smart plug [?] that measures the power con-sumption of the appliance plugged into it. Figure 5 showsdetails of the four homes in which LightSense and the groundtruth sensing system were deployed. We see that each homehas a high proportion of rooms with windows, which intro-duces higher noise levels due to natural light changes for ourLightSense approach to handle. During our deployments, wealso encountered multiple light fixture types such as incan-descent bulbs, CFLs, and halogen lights, and both dimmable

Figure 6. Energy Costs at $0.15 per KWh projected over 5 years, for individual light fixtures across all four homes, as computed by LightSense andthe ground truth system. LightSense accurately reports the energy costs of the top energy consumers in each home.

(a) TED Power Meter

(b) Light Sensing Mote

(c) ZWave Switch

(d) ZWave Smart Plug

Figure 7. Our system uses (a) a smart power meter on the electrical mains, and (b) light sensors in each room. For experimental ground truth, wealso deployed (c) Z-wave light switches for in-line fixtures, and (d) Z-wave plug load metering for plug-in fixtures.

and non-dimmable switches with different light intensitiesand wattages as seen in figure 11.

Energy Usage Ground TruthBefore we can evaluate the energy feedback accuracy of Light-Sense, we first need to map the set of ground truth light fix-tures GF monitored using the ZWave devices, to the set oflight fixtures LF inferred by LightSense. To perform thismapping, we first compute a match recall metric mrg,l be-tween each fixture g ∈ GF and each fixture l ∈ LF , thatdenotes the proportion of ON-OFF events from g that aredetected by at least one ON-OFF event from l. An eventx from fixture g is said to be detected by an event y fromfixture l if both the ON and OFF times of events x and yare within dTm = 10 seconds of each other. We map eachground truth fixture g to an unmapped LightSense fixturemap(g) ∈ LF with the maximum match recall metric, if themaximum match recall metric mrg,max(g) > .5, i.e. at least50% of the ON-OFF events from fixture g are detected fromfixture l. At the end of this matching process, some lightfixtures in GF and LF may be unmapped, resulting in falsenegatives and false positives, respectively.

To provide an easily understandable economic interpretationof light fixture energy consumption, we transform the rawenergy in KWh for each light fixture to a projected energycost over 5 years, assuming a cost of $0.15 per KWh, and

extrapolating our 10 day sample period to a 5 year period.We use the ZWave ON and OFF times to compute the energycost of each ground truth light fixture. We set the thresholdparameters dL(in raw ADC units), and dP (in watts), to bothbe 10 in order to identify even low intensity light fixtures; weperform a sensitivity analysis on these parameters in section.

EVALUATIONTo evaluate our WaterSense approach, we deploy our sys-tem in two homes for 7 days each. Both homes had multipleresidents, multiple bathrooms, and a wide array of water fix-tures and appliances. Details of the two deployments aresummarized in Figure 10. To execute WaterSense in thesehomes, we deployed a single water flow sensor on the watermains and a motion sensor in each room. In Home 2, one ofthe motion sensors in a bathroom malfunctioned during ourweek long deployment, so we used a motion sensor in anadjacent bedroom with a partial view of the bathroom in ouranalysis. To evaluate the system, we deployed ZWave con-tact reed switches [?] , as shown in Figure ?? to record theactual times that each fixture was used. We compute groundtruth water consumption for each fixture by integrating totalwater flow into the house when each fixture was used. Inthe case of simultaneous toilet and sink events when we donot observe explicit sink edges, we ignore the short durationusage of the sink. These cases constitute a small fraction of

Home #

Number of Residents

Number of rooms with fixtures

Number of monitored fixtures

Pipe material and width

Home 1

2 3 5 ¾” Copper

Home 2

4 3 5 ¾” PEX

Figure 10. Our deployments involved two homes with multiple resi-dents. Both homes had a kitchen sink and two bathrooms with bothsink and toilet.

events and we do not believe that it significantly changes theoverall water consumption profiles. In this experiment, weinstalled ground truth sensors on only 3 sinks and 2 toiletsin each home and only evaluate accuracy for these fixtures.Other fixtures and appliances were also used during the ex-periment period, and we observe several anecdotal instancesof fixtures such as the shower or the sprinker system beinginferred in these homes. However, we currently limit ourevaluation only to those ground truth events observed fromthe ZWave system. In the future, we intend to expand ourevaluation to a larger set of fixtures and appliances.

The true and estimated water consumption levels for eachfixture are shown in Figure ??. The order of the fixturesis sorted based on their actual flow. WaterSense estimatedwater flow at each fixture with a median accuracy of 81.5%in Home 1 and 89.9% in Home 2. More importantly, how-ever, WaterSense preserves the relative ordering of fixturesin terms of maximum consumption; it correctly indicates, forexample, which sink or toilet causes the most water usage.End users can use the output of WaterSense to understandwhat the high consumption fixtures in their home are. Ifcombined with intuitive displays that show users when eachfixture was being used, and in which daily activity, we ex-pect to provide users with actionable recommendations tosave water, such as leaving the sink closed while washingor brushing, or providing low flow replacements for certainshowers and flushes. For example, in Homes 1 and 2, res-idents might be able to infer that it might be most efficientto replace the high flow toilet in bathroom 1, while in Home2, residents can also infer that replacing the high flow toiletin bathroom2 would have about the same effect as using thesinks in bathroom1 and the kitchen more efficiently with abetter understanding of how they are used.

Figure 9 shows the confusion matrix for classifying individ-ual fixture usage events. We observe that classification ac-curacy is high for the majority of fixtures in the home. Mostmisclassifications that did occur were due to simultaneousoccupancy in multiple rooms from fixtures with similar flowsignatures, such as high confusion between sink usage in thekitchen and bathroom2 in Home 1. In Home 1, we findzero confusion between the flush fixtures in the two bath-rooms even in the presence of simultaneous occupancy asseen in the example data trace in Figure 3, while in Home 2,

there is about 7% confusion between the two flush fixtures.The reason is that in Home 1, the two flush fixtures have alarger difference in flow signatures (0.3 and 0.6 kl/hour flowrates), while in Home 2, the two flush fixtures have simi-lar flow signatures of 0.6-0.7 kl/hour; the Bayesnet in Home2 relies more heavily on the noisy instantaneous occupancythan in Home 1, resulting in slightly more errors. In general,these misclassifications do not cause significant degradationin water consumption accuracy because they are infrequentand roughly symmetric across the diagonal, and so they of-ten cancel each other out.

RESULTS

Light Fixtures(figures 6,10,11) from lightsense

Water Fixtures(figures 6) from watersense - briefly discuss confusion ma-trix, classification accuracy

ASSESSING ENERGY USAGEGiven the above experimental setup, we assess the accuracywith which LightSense infers the energy usage of individuallight fixtures. Figure 6 shows the projected energy cost ofeach light fixture for the ground truth approach and Light-Sense, across all four houses over the 10 day deploymentperiod. For each light fixture in figure 6, figure 11 shows thecorresponding room locations, wattages, and light intensitychanges observed at our light sensors.

We see from figure 6 that our LightSense approach accu-rately computes the energy costs of the top energy consumersin each home; we observe an average accuracy of 91.1% indetermining the energy cost of light fixtures consuming 90%of the home’s lighting energy. All of the false negatives andpositives have very low energy consumption and do not sig-nificantly interfere with decision making about which lightfixtures to optimize. Interestingly, House 1, which is a large3 floor house, uses less lighting energy than House 2, whichis a 3 bedroom student apartment; the reason is that House1 has large wall-sized windows, which negate the need forlighting during most of the day. Our feedback can also beused to apportion the lighting bill based on energy usage; forexample, the resident in Bedroom1 of House 2 might be mo-tivated to reduce her high lighting energy given a personalcost incentive to do so.

For each light fixture, LightSense uses Tier III to provideusers with the (1) nominal wattage, (2) room location basedon the light sensor room location, and (3) the total energycost. LightSense reports the nominal wattage of each lightfixture shown in figure 11 within ±5W accuracy for all thetop 90% energy consuming light fixtures in each home. Us-ing these three inputs, residents can optimize high energylight fixtures by replacing incandescent bulbs with CFLs orLED bulbs, by replacing traditional lights with motion acti-vated lights, by implementing a daylight harvesting system,or just by using the lights in a more conservative fashion.Each of these techniques requires a different input cost and

Fixtures K S B1 S B1 F B2 S B2 F

K S 8 0 0 8 1

B1 S 1 19 0 2 0

B1 F 0 1 49 0 0

B2 S 3 1 0 16 2

B2 F 0 0 0 1 14

(a) Home 1

Fixtures K S B1 S B1 F B2 S B2 F

K S 81 10 0 3 0

B1 S 7 78 0 5 0

B1 F 1 0 85 0 5

B2 S 0 1 0 6 0

B2 F 0 0 2 1 13

(b) Home 2

Figure 9. WaterSense accurately classifies water flow events for most of the monitored water fixtures across the two homes as seen in the fixture levelconfusion matrices. Confusion between fixtures of the same type in different rooms is common due to overlapping occupancy. Confusion betweenfixtures of different types, such as a sink and a flush, due to overlapping flow signatures, is less common. (B stands for bathroom, K for kitchen, S forsink, and F for flush)

Light

#

Room Wattage Light

change

1 MasterBed 135 23

2 Livingroom 35 193

3Livingroom 40 29

4Kitchen 90 20

5 MasterBath 20 36

6 MidBathroom 10 62

7BottomBath 40 41

8MasterBath 42 0

9BottomBath 40 27

10 Kitchen 5 -

11TopRoom 80 923

12Kitchen 98 -

Light

#

Room Wattage Light

change

1 Livingroom 185 366

2 Bedroom1 90 55

3Kitchen 120 71

4Bedroom2 35 50

5 Bedroom3 50 414

6 Bathroom 50 260

Light

#

Room Wattage Light

change

1 Livingroom 95 155

2 Livingroom 115 59

3Bathroom2 220 1187

4Bedroom 95 70

5 Kitchen 110 181

6 Officeroom 90 55

7Bathroom1 305 1157

8Diningroom 200 182

9Livingroom 55 239

10* Diningroom 100 19

Light

#

Room Wattage Light

change

1 Frontroom 55 311

2 Basement 325 353

3Kitchen 250 60

4Bathroom 395 1028

5 Kitchen 280 125

6 Livingroom 80 50

7Dining Room 200 110

8Bedroom 95 69

9Bathroom1 95 649

10 Nursery 55 129

11Bedroom 60 32

12Kitchen 70 -

13Kitchen 30 -

14Kitchen 110 -

House 1 House 2 House 3 House 4

Figure 11. Nominal Wattages and Room Locations for the light fixtures in the four houses along with the light intensity increase or decrease observedat our light sensor as the fixtures are switched ON or OFF, respectively. The light fixture numbers here match the light fixture numbers in figure 6.

0 2 4 6 8 10 12 140

50

100

150

200

250

300

350

400

# of lights to optimize

cost

sav

ings

ove

r 5

year

s in

$

House 1 OptimalHouse 1 LightSense

House 2 OptimalHouse 2 LightSense

House 3 OptimalHouse 3 LightSense

House 4 OptimalHouse 4 LightSense

Figure 12. Projected cost savings based on LightSense recommenda-tions and an optimal system with ground truth data.

reduces energy by a different amount, so the energy feed-back from LightSense is valuable to the end user in order tomaximize net cost savings.

Given a limit n on the number of light fixtures that userscan optimize, we compute the maximum projected energycost savings achievable over 5 years (the typical lifetime ofa CFL), assuming we optimize the top n light fixtures basedon the energy cost ordering computed by LightSense; we as-sume 75% energy reduction for each light fixture after opti-mization, which is similar to the energy savings obtained onswitching from incadescent lights to CFLs. Figure 12 showsthe maximum projected cost savings of LightSense for eachhouse as n is increased along the x-axis. We also show thecost savings for an optimal system that uses the ground truthenergy costs. We see that LightSense closely tracks the op-timal recommendation system for optimizing light fixturesacross all houses. Also, in each house, the cost savings arequite different and become very small after 6-7 fixture arereplaced; this suggests that users should carefully weight thecost vs. benefit of optimizing light fixtures, before decidingon a suitable n for each house.

We see the value of LightSense by combining the resultsfrom figures 6, 11 and 12 ; many of the low energy, low re-turn on investment light fixtures in each house are in similarrooms as the top energy consumers, and in some cases evenhave a higher or similar nominal wattage compared to thetop energy consumers. Thus, optimizing light fixtures basedon their actual energy usage is more profitable compared tousing only their nominal wattages or room locations.

ANALYSISIn this section, we provide detailed insights on the high ac-curacy achieved by LightSense in inferring the energy costof individual light fixtures in the home. Firstly, we show thatboth the data fusion and Bayesian matching steps contributeequally to aggresively filter out false positives. Secondly,we show that our Bayesian Clustering approach is needed toaccurately assign power measurements to ON-OFF events.Finally, we perform a sensitivity analysis of LightSense tothe power and light thresholds dP and dL.

House 1 House 2 House 3 House 40

20

40

60

80

100

Tot

al R

elat

ive

Pro

ject

ed C

ost E

rror

%

Power Only Assignmentlpbayes Power Assignment

Figure 14. Relative Total Cost Error for power only assignment andlpbayes power assignment.

Impact of Data Fusion and MatchingTo understand the impact of Data Fusion and Bayesian Match-ing on improving LightSense accuracy, we implement fivedistinct inference approaches with increasing sophistication,namely: Light Edge Detection (LE), Power Edge Detec-tion (PE), Light Edge Matching (LM), Power Edge Match-ing (PM), and finally Light Power Data Fusion (LP), thateliminates temporally isolated light or power edges from byapplying our temporal intersection step. We also considerthe LightSense approach, which effectively uses all the in-ference algorithms and sensor sources used in the five ap-proaches above. To measure the accuracy of each infer-ence approach above, we use the common evaluation metricsof edge detection recall and precision for light fixture ONand OFF edges. Figure 13 shows these metrics for all ap-proaches; we fix the light and power thresholds dL = dP =10, to be low enough to detect the low intensity fixtures seenin figure 11. We see that approaches LM and LP improveprecision significantly compared to the simpler approachesLE and PE in each house, by eliminating unmatched or tem-porally isolated false positive edges. Interestingly, approachPM only has a marginal improvement over approach PE dueto the large number of false positive power edges. Finally,we see that in each house, LightSense improves upon bothLM and LP, by combining both noise elimination techniquesto aggressively filter out false positives from the light andpower meter data streams.

Impact of Bayesian ClusteringTo understand the impact of using Bayesian Clustering in ac-curately assigning power costs to ON-OFF events, we con-sider two different power assignment approaches: (1) PowerOnly assignment uses only the power meter data to assignpower edges to ON-OFF events, i.e. by only using vprobix,yin equation 2, (2) Bayesian Clustering assignment also usesthe joint probability of observing the light and power me-ter data as shown in equation 2 using our lpbayes Bayesnet.For each approach, we consider the total energy cost error,normalized as a percentage by the total energy cost of allfixtures in the home. In figure 14, we observe that leverag-ing the long term temporal correlation between the light andpower meter data reduces the total normalized energy costerror from 40-50% to less than 10% in multiple homes. Wesee almost no impact of our Bayesian Clustering approach inHouse 2, because this small apartment has a little noise onthe power lines due to the lack of HVAC, television, or other

House 1 House 2 House 3 House 40

10

20

30

40

50

60

70

80

90

100

Edg

e D

etec

tion

Rec

all %

PELE

PM

LM

LPLightSense

(a) Edge Detection Recall

House 1 House 2 House 3 House 40

10

20

30

40

50

60

70

80

90

100

Edg

e D

etec

tion

Pre

cisi

on %

PELE

PM

LM

LPLightSense

(b) Edge Detection Precision

Figure 13. Edge Detection Recall and Precision for LightSense and its individual inference components. LightSense outperforms the individualcomponents by effectively fusing the noise filtering properties of data fusion and ON-OFF event matching.

complex appliances.

Parameter Sensitivity AnalysisThe main variable parameters in our approach that affectclassification accuracy are the light and power edge thresh-olds dL(in ADC units) and dP (in Watts). Figure 15 showsthe recall-precision graph for all the five additional inferenceapproaches from section and LightSense as dL and dP aresimultaneously decreased from 500 to 10 in steps of 50 each;we set the the minimum values of the threshold parametersto be 10 instead of 0, and make the last step size 40. We plotthe edge detection recall and precision for each parametersetting averaged across the four homes.

As the threshold parameters are decreased, we observe that(1) edge recall for all approaches increases as low intensityevents greater than the decreasing threshold are detected,and (2) precision decreases as more false positives are intro-duced from low intensity edges. While approaches LM andLP achieve only about a 5-10% worse F1 score compared toLightSense at a threshold setting of 50, they achieve muchlower precision when detecting low intensity events at lessthan 50 ADC light units or less than 50 W. However, iden-tifying low intensity events is important, since we see fromfigure 11 that several light fixtures in the top 5 or 6 con-sumers in each home have either a wattage less than or closeto 50W, or a light intensity less than 50 ADC units. Thus,our LightSense approach combines both noise eliminationtechniques, matching and data fusion, to accurately identifyeven important low intensity light fixture events.

CONCLUSIONSIn this work, we presented the LightSense system to infer thefine-grained energy usage of light fixtures in homes. Through10-day deployments in four natural home environments, weshow that LightSense effectively combines two noisy sensorsources, namely light sensors with 25% precision and wholehouse power consumption data with 15% precision, to in-fer the energy costs of individual light fixtures with 91.1%accuracy.

=====================

In this paper, we present an unsupervised approach to infer

0 10 20 30 40 50 60 70 80 900

10

20

30

40

50

60

70

80

90

100

Recall (in %)

Pre

cisi

on (

in %

)

LEPE

LM

PM

LPLightSense

Figure 15. Recall Precision tradeoff for LightSense and its individualinference components as the light and power level thresholds are de-creased from 500 to 10 in steps 50.

the fine-grained fixture level breakdown of water consump-tion in homes by effectively combining cheap occupancysensors and whole water flow meter data. Our approachshows significant promise in an early evaluation carried outin two homes over 7 days each, and is able to accurately in-fer both the time of use and water consumption of individualfixtures. In the future, we expect to build upon the exist-ing approach by performing an extensive evaluation of oursystem that includes ground truth sensing on fixtures suchas the shower, washing machine, and dishwasher to addresspotential challenges posed due to simultaneous fixture usefrom these high consumption fixtures. There are also severalinteresting alternatives to explore in our multi-tier inferencearchitecture to improve accuracy in a complete evaluation,including improvements to fixture identification in Tier III,improvements to the temporal distance features used in thebayesnet, and additions to the algorithm to handle simulta-neous or compound flow events. With these additions, weexpect our WaterSense approach to be a viable approach toprovide fine grained water consumption information and rec-ommendations to end users to help conserve water.

REFERENCES1. Hamamatsu light diodes. http://www.hamamatsu.com.

2. Pike research smart grid deployment tracker.http://www.pikeresearch.com/research/smart-grid-

deployment-tracker-3q10.

3. Ted the energy detective.http://www.theenergydetective.com.

4. Energy Information Administration. 2005 residentialenergy consumption survey.http://www.eia.doe.gov/emeu/recs/contents.html.

5. J. Fogarty, C. Au, and S. Hudson. Sensing from thebasement: a feasibility study of unobtrusive andlow-cost home activity recognition. In Proceedings ofthe 19th annual ACM symposium on User interfacesoftware and technology, pages 91–100. ACM, 2006.

6. J. Froehlich, E. Larson, T. Campbell, C. Haggerty,J. Fogarty, and S. Patel. Hydrosense:infrastructure-mediated single-point sensing ofwhole-home water activity. In Proc. UbiComp,volume 9, pages 235–244. Citeseer, 2009.

7. J. Froehlich, E. Larson, E. Saba, T. Campbell, L. Atlas,J. Fogarty, and S. Patel. A longitudinal study ofpressure sensing to infer real-world water usage eventsin the home. Pervasive Computing, pages 50–69, 2011.

8. S. Gupta, M. Reynolds, and S. Patel. Electrisense:Single-point sensing using emi for electrical eventdetection and classification in the home. In Proceedingsof Ubicomp, 2010.

9. G. Hart. Nonintrusive appliance load monitoring.Proceedings of the IEEE, 80(12):1870–1891, 1992.

10. L. Heyer, S. Kruglyak, and S. Yooseph. Exploringexpression data: identification and analysis ofcoexpressed genes. Genome research, 9(11):1106,1999.

11. X. Jiang, S. Dawson-Haggerty, P. Dutta, and D. Culler.Design and implementation of a high-fidelity acmetering network. In Proceedings of IPSN, 2009.

12. D. Jung and A. Savvides. Estimating buildingconsumption breakdowns using on/off state sensing andincremental sub-meter deployment. In Proceedings ofSensys 2010.

13. Y. Kim, T. Schmid, Z. Charbiwala, J. Friedman, andM. Srivastava. Nawms: nonintrusive autonomous watermonitoring system. In Proceedings of the 6th ACMconference on Embedded network sensor systems,pages 309–322. ACM, 2008.

14. Y. Kim, T. Schmid, Z. Charbiwala, and M. Srivastava.Viridiscope: design and implementation of a finegrained power monitoring system for homes. InProceedings of Ubicomp, 2009.

15. H. Kuhn. The hungarian method for the assignmentproblem. Naval research logistics quarterly,2(1-2):83–97, 1955.

16. P. Mayer and W. DeOreo. Residential end uses ofwater. American Water Works Association, 1999.

17. S. Patel, T. Robertson, J. Kientz, M. Reynolds, andG. Abowd. At the flick of a switch: Detecting andclassifying unique electrical events on the residentialpower line. Pervasive, 2007.

18. J. Polastre, R. Szewczyk, and D. Culler. Telos: enablingultra-low power wireless research. In Proceedings ofIPSN, 2005.

19. A. Rowe, M. Berges, and R. Rajkumar. Contactlesssensing of appliance state transitions through variationsin electromagnetic fields. In Proceedings of Buildsys,2010.

20. San Diego Regional Energy Office. Residential EnergyEfficiency Measures. http://energycenter.org/uploads/Residential%20Efficiency%20Measures.pdf.

http://energycenter.org/uploads/Residential%20Efficiency%20Measures.pdfhttp://energycenter.org/uploads/Residential%20Efficiency%20Measures.pdfhttp://energycenter.org/uploads/Residential%20Efficiency%20Measures.pdf

IntroductionRelated WorkA Data Fusion AlgorithmTier I: Event DetectionLight FixturesWater Fixtures

Tier II: Data FusionLight FixturesWater Fixtures

Tier III: Event MatchingLight FixturesWater Fixtures

Tier IV: Fixture ClusteringLight FixturesWater Fixtures

A Case Study: Motion and Light SensorsInferring Water Fixture UsageInferring Light Fixture Usage

LightSense DesignTier I: Light and Power Edge detectionPower Edge DetectionLight Edge Detection

Tier II: Data Fusion and Matching Data FusionBayesian MatchingBayesian Clustering

Tier III: Fixture Identification

Experimental SetupPhysical Sensor ComponentsReal World DeploymentsEnergy Usage Ground Truth

EvaluationResultsLight FixturesWater Fixtures

Assessing Energy UsageAnalysisImpact of Data Fusion and MatchingImpact of Bayesian ClusteringParameter Sensitivity Analysis

ConclusionsREFERENCES