WATERSENSE: WATER FLOW DISAGGREGATION USING MOTION SENSORSVijay Srinivasan, John Stankovic, Kamin WhitehouseDepartment of Computer ScienceUniversity of Virginia
Water Monitoring
World’s usable water supply decreasing
Household water conservation can save fresh water reserves
Before you can conserve it, measure it first!
1000 gallons
1000 gallons
200 gallons
800 gallons
Water Monitoring Fixture level
usage Change Behavior Change Fixtures Activity
Recognition
Water Meter Data Aggregate water
consumption
1000 gallons
1000 gallons
200 gallons
800 gallons
Water
Meter
3000 gallons
Disaggregation problem
Background Flow Profiling
Ambiguity with similar sinks, flushes
Direct flow metering Expensive, In-line
plumbing
Accelerometers Sensors on all fixtures
Single point water pressure sensor High training cost
Water
Meter
5 gallons/min1 minute
1 gallon/min.5 minutes
1 gallon/min.5 minutes
WaterSense Data Fusion Approach Combine water
meter with motion sensors
Key Insight Fixtures with the
same flow profile may have unique motion profiles
Use <flow + motion> profile
Water
Meter
5 gallons/min1 minute
1 gallon/min.5 minutes
1 gallon/min.5 minutes
WaterSense Data Fusion Approach WaterSense
advantages Easy to install Cheap ($5) No Training
Water
Meter
5 gallons/min1 minute
1 gallon/min.5 minutes
1 gallon/min.5 minutes
Rest of the talk WaterSense Design WaterSense Evaluation Conclusions
WaterSense Data Fusion Approach
Kitchen motion
Bathroom1 motion
Bathroom2 motion
Water Flow rate in kl/hour
Time in HoursThree Tier Approach
WaterSense Data Fusion Approach - Tier I Flow Event Detection
Kitchen motion
Bathroom1 motion
Bathroom2 motion
Water Flow rate in kl/hour
Time in Hours
Flow event 1
Flow event 2
Canny Edge Detection Rising and falling
edges Bayesian matching
Flow events
0.75 kl/hr, 35 seconds
0.75 kl/hr, 45 seconds
WaterSense Data Fusion Approach - Tier II Room Clustering
Kitchen motion
Bathroom1 motion
Bathroom2 motion
Water Flow rate in kl/hour
Time in Hours
Flow event 1
Flow event 2
Flow profile ambiguous
Look at which motion sensors occur at the same time as the flow event Temporal
distance feature for each room
0.75 kl/hr, 35 seconds
0.75 kl/hr, 45 seconds
Kitchen motion
Bathroom1 motion
Bathroom2 motion
Water Flow rate in kl/hour
Time in Hours
Flow event 1
Flow event 2
0.3 kl/hr, 90 seconds
0.6 kl/hr, 40 seconds
Temporal distance feature ambiguous? Simultaneous
activities Missing activity
WaterSense Data Fusion Approach - Tier II Room Clustering
Kitchen motion
Bathroom1 motion
Bathroom2 motion
Water Flow rate in kl/hour
Time in Hours
Flow event 1
Flow event 2
0.3 kl/hr, 90 seconds
0.6 kl/hr, 40 seconds
Temporal distance feature ambiguous? Simultaneous
activities Missing activity
Cluster flow events by flow profile
Learn cluster to room likelihood
WaterSense Data Fusion Approach - Tier II Room Clustering
Cluster 1 Cluster 2
Cluster 1
Cluster 2
Kitchen motion
Bathroom1 motion
Bathroom2 motion
Water Flow rate in kl/hour
Time in Hours
Hidden variables
Evidence variables
Room
Temporal
Distance
Flow rate,
duration
Flow cluster
P(Room | Temporal Distance, Flow rate, Duration)
Bayesnet to label each flow event
Cluster 1
Cluster 2
Cluster 1 Cluster 2
Flow event 1
Flow event 2
0.3 kl/hr, 90 seconds
0.6 kl/hr, 40 seconds
WaterSense Data Fusion Approach - Tier II Room Clustering
- Use a binary temporal distance feature
- Use quality threshold clustering for flow profiles
- Maximum likelihood estimation
Kitchen motion
Bathroom1 motion
Bathroom2 motion
Water Flow rate in kl/hour
Time in Hours
Cluster 1
Cluster 2
Cluster 1 Cluster 2
Flow event 1
Flow event 2
0.3 kl/hr, 90 seconds
0.6 kl/hr, 40 seconds
WaterSense Data Fusion Approach - Tier III Fixture Identification
Use simple flow profiling to identify fixture E.g.) Flush events
different from sink events
Tier III fixture type + Tier II room assignment results in a unique water fixture
Rest of the talk WaterSense Design WaterSense Evaluation Conclusions
Home Deployments Two homes for one
week each
Ultrasonic water flow meter (2 Hz)
X10 motion sensor ($5)
Ground Truth Zwave reed switch
sensors
Flow meter
X10 motion sensor
Zwave reed switch sensor
Water Consumption Accuracy 90% Water Consumption Accuracy Use Accurate feedback to improve water
usage
B – BathroomK – KitchenS – SinkF – Flush
86% classification accuracy Errors have reduced effect on
consumption accuracy
Water Usage Classification
B – BathroomK – KitchenS – SinkF – Flush
Rest of the talk WaterSense Design WaterSense Evaluation Conclusions
Limitations and future work Current evaluation limited to simple
fixtures Include all fixtures, including washing
machines, sprinklers, and dishwashers, in future evaluation
Extend evaluation period
Current system uses binary motion data Explore joint clustering of infrared motion
readings and water flow profiles
Conclusions WaterSense – Practical data fusion
approach to water flow disaggregation Cheap Unsupervised
Water consumption accuracy of 90%
High Enough Classification accuracy for activity recognition applications
Thank YouFeedback or Questions?