FixtureFinder: Discovering the Existence of Electrical and Water Fixtures Vijay Srinivasan*, John Stankovic, Kamin Whitehouse University of Virginia *(Currently affiliated to Samsung) Motivation For Fixture Monitoring Home Healthcare Applications Cooking Toileting Resource conservation applications 7 KW hours 400 liters Fixture Monitoring Using Smart meters Whole house power or water flow Water meter Power meter Time • Poor accuracy for low power or low water flow fixtures • False positive noise • Identical fixtures 100 litres/hour Bathroom 100 W Bedroom 100 W 2000 W Kitchen 100 litres/hour 100 W Livingroom Existing Fixture Monitoring Techniques Direct metering on each fixture Single-Point Infrastructure sensing Images courtesy: HydroSense and Viridiscope (Ubicomp 2009) Indirect sensing + smart meter Requires users to: • Identify each fixture, and for each fixture: • Install a sensor, or • Provide training data FixtureFinder • Automatically: Light and motion + – Identify fixtures Lights, sinks and toilets – Infer usage times Home security or automation sensors – Infer resource consumption Water meter Power meter 2 PM 5 PM … Training data 400 liters Bathroom Kitchen Single-Point Infrastructure sensing 7 KW hours Bedroom Livingroom FixtureFinder Insights Fixtures identical in meter data Unique in (meter, sensor) data Power meter Bathroom Water meter Kitchen Light sensor Bedroom 100 W, 30 lux 100 W 100 W, 50 lux Livingroom FixtureFinder Insights 1. Eliminate noise False positive events in one noise in meter stream when no and sensor activity in other data stream 2. Eliminate unmatched noise Power meter Bathroom Water meter Kitchen Light sensor Power meter data Bedroom light sensor data ON-OFF pattern Bedroom 100 W, 30 lux 100 W, 50 lux Livingroom Outline • • • • • FixtureFinder algorithm Case studies Experimental setup Evaluation results Conclusions FixtureFinder Algorithm Inputs Light or motion sensors Stream 1 • Four step algorithm or Stream 2 Power meter Water meter Step 1 – Event Detection For example: 40 lux 100 Watts Stream 1 Light sensor ON 40 40 Stream 1 Edge detection algorithms Key challenge: Large number of false positives OFF 40 140 Time ON 100 500 60 40 Stream 2 100 OFF False positives events: Stream 2 True positive events: Power meter 200 60 Step 2 – Data fusion For example: 40 lux 100 Watts Stream 1 Light sensor ON 40 40 Stream 1 Fixture use creates events in multiple streams simultaneously Compute event pairs OFF 40 140 Time ON 100 500 60 40 Stream 2 100 OFF Eliminate temporally isolated false positives Stream 2 Power meter 200 60 Step 3 – Matching For example: 40 lux 100 Watts Stream 1 Light sensor ON Fixture use occurs in an ON-OFF pattern Match ON event pairs to OFF event pairs High match probability 40 Stream 1 OFF 40 140 Time ON 100 500 60 40 Stream 2 100 OFF Eliminate unmatched false positives Stream 2 Power meter 200 60 Step 3 – Matching For example: 40 lux 100 Watts Two ON-OFF event pairs: (40,100) or (40,60) ? Stream 1 Light sensor ON High match probability 40 Stream 1 OFF True event pairs are more likely than noisy event pairs Use both match and pair probabilities to compute ON-OFF event pairs Soft clustering and Min Cost Bipartite matching (Described in paper) 40 All false positives eliminated Low pair probability in this example! High pair probability Time ON 100 60 Stream 2 100 OFF Stream 2 Power meter 60 Step 4 – Fixture Discovery Step 3: Matching ON-OFF events Fixtures discovered Stream 1 (Light) intensity Stream 2 (Power) intensity ON Time OFF Time 41 102 5 PM 6 PM 62 103 5:30 PM 6:15 PM 43 99 8 PM 10 PM 60 101 7 PM 8 PM 61 100 9 PM 10 PM Clustering based on: (stream 1 intensity, stream 2 intensity) 40 lux, 100 watts Clustering 60 lux, 100 watts Outline • • • • • FixtureFinder algorithm Case studies Experimental setup Evaluation results Conclusions Light Fixture Discovery Power meter Apply FixtureFinder algorithm on every (light sensor, power meter) Unique fixture usage defined by: Light sensor location Light intensity Power consumption Water meter Bathroom Kitchen 40 lumens, 100 watts Bedroom 40 lumens, 150 watts Livingroom Light Fixture Discovery False positives eliminated after steps 2 and 3 Bedroom light fixture ONOFF events Large number of false positives after step 1 Bedroom light sensor data Power meter data Water Fixture Discovery Power meter Apply FixtureFinder algorithm on (fused motion sensor, power meter) 300 litres/hour Fused motion sensor stream Water meter 100 litres/hour Bathroom Kitchen 100 litres/hour Unique fixture usage defined by: Motion sensor signature Flow rate Bedroom Livingroom Water Fixture Discovery Two toilets with the same flow signature but different motion signatures Water Fixture Discovery Use event pair probability to pair simultaneous toilet events with correct rooms Two toilets with the same motion signature but different flow signatures Outline • • • • • FixtureFinder algorithm Case studies Experimental setup Evaluation results Conclusions In-Situ Sensor Deployments in Homes X10 motion Custom light sensing mote One per room in a central location (Except in 3 large rooms where two sensors were used) One per home Power meter (TED 5000) Water meter (Shenitech) In-Situ Sensor Deployments in Homes Ground truth for light fixtures Smart switch Smart plug Ground truth for water fixtures Contact switches on water fixtures All sensors deployed in 4 homes for 10 days (Except water meter deployed in 2 homes for 7 days) Outline • • • • • FixtureFinder algorithm Case studies Experimental setup Evaluation results Conclusions Fixture Discovery Results Discovered all sinks and toilets across 2 homes Discovered 37 out of 41 light fixtures across 4 homes Undiscovered lights: - All in large kitchens - Task lighting or under-cabinet lighting - Used rarely (1-3 times) - Low energy consumption One false positive light with negligible energy consumption Fixture Usage Inference Results Results shown for light fixtures True positive ON-OFF events from fixtures Precision: % of detected fixture events that are supported by ground truth Training data 99% precision 64% recall High precision usage data Recall: % of ground truth fixture events detected by Fixture Finder Single-Point Infrastructure sensing Fixture Usage Inference Results Results shown for light fixtures Home Activity Monitoring applications Precision: % of detected fixture events that are supported by ground truth 92% precision 82% recall Balanced precision and recall Recall: % of ground truth fixture events detected by Fixture Finder Analysis of FixtureFinder Steps • Step 1: Event Detection Results shown for light fixtures – ME: Meter event detection – SE: Sensor event detection Small reduction in recall • Step 3: Matching – MM: Meter event matching – SM: Sensor event matching • Step 2: Data Fusion – SMF: Sensor meter data fusion • FixtureFinder Significant increase in precision with steps 2, 3, and FixtureFinder Light Fixture Energy Estimation • 91% average energy accuracy for top 90% energy consuming fixtures Water Consumption Estimation • 81.5% accuracy in Home 3 • 89.9% accuracy in Home 4 Home 3 B – Bathroom K – Kitchen S – Sink F – Flush Home 4 Outline • • • • • FixtureFinder algorithm Case studies Experimental setup Evaluation results Conclusions Conclusions • FixtureFinder combines smart meters with existing home security sensors to automatically: – Identify fixtures – Infer usage times – Infer resource consumption • Demonstrated for light and water fixtures • Complements other fixture monitoring techniques by providing training data without manual effort Future Improvements • Expand scope to include: – Additional electrical appliances and water fixtures – Additional sensing modalities such as routers, smart switches, infrastructure sensors • Extend algorithm to multi-state appliances – Not just two-state ON-OFF • Explore temporal co-occurrence over multiple timescales Thanks Questions? FixtureFinder Approach Light and motion Home security or automation sensors • Automatically discover low power or low water flow fixtures – Lights, sinks, and toilets + Power meter Bathroom Bedroom Water meter Kitchen Livingroom Step 3 – Bayesian Matching • Two matches possible Stream 1 – (40,100) or (40,60) • Assumption: Edge pairs from true fixtures are more frequent than noisy edge pairs ON 40 OFF 40 – P(40,100) >> P(40,60) Time ON Stream 1 cluster Stream 2 cluster 100 60 100 OFF Hidden variables Stream 2 Stream 1 edge Stream 2 edge Observed variables 60 Step 3 – Bayesian Matching • Incorporate edge pair probability into a match weight function • Perform optimal bipartite matching based on match weight function • Eliminate unlikely matches Stream 1 ON 40 OFF 40 Time ON 100 60 100 OFF Stream 2 60