1. disaggregated hot and cold water sensing with minimal calibration
semi-supervised training for infrastructure mediated sensing
Eric C. Larson
UbiComp Lab
electrical
engineering
computer
science and engineering
University of Washington
2. how can indirect sensing and machine
learning be used to reduce our
environmental footprint?
6. we are using water faster
than it is being replenished
Pacific Institute for Studies in Development, Environment, and Security, 2011 6
7. we are using water faster
than it is being replenished
Pacific Institute for Studies in Development, Environment, and Security, 2011 6
image: weiku.com
27. meters
flow rate fixture flow
inline water
water
pressure
pressure
sensor
machine
learning
estimated
28. • central sensing point
• easy to install
• low cost
• can observe every fixture
HydroSense
25
29. • central sensing point
• easy to install
• low cost
• can observe every fixture
HydroSense
25
40#
50#
60#
70#
80#
Cold Line Pressure
(Hose Spigot)
0 94.5
time (s)
psi
open close
33. feasibility study
• 10 homes
• staged calibration
• ~98% accuracy
Froehlich, J., Larson, E., Campbell, T., Haggerty, C., Fogarty, J., and Patel, S.N. HydroSense: infrastructure-mediated single-point sensing of whole-home water
activity. Proceedings of the 11th ACM international conference on Ubiquitous computing, (2009), 235–244.
Larson, E., Froehlich, J., Campbell, T., et al. Disaggregated water sensing from a single, pressure- based sensor: An extended analysis of HydroSense using staged
experiments. Pervasive and Mobile Computing, (2010).
28
42. totals
days
33
33
30
27
33
156
events
2374
3075
4754
2499
2578
14,960
events/day
71.9
93.2
158.5
92.6
78.1
95.9
compound
22.2%
21.8%
16.6%
32%
21.3%
21.9%
data collection
Larson, E., Froehlich, J., Saba, E., et al. A Longitudinal Study of Pressure Sensing to Infer Real- World Water Usage
Events in the Home. Pervasive Computing, Springer (2011), 50–69.
most comprehensive labeled dataset
of hot and cold water ever collected
37
58. accuracy
trusted
not trusted
how accurate should the system be?
how can we be sure the user trusts the system?
highly critical
noticeable
Lim, B. and Dey, A. Investigating intelligibility for uncertain context-aware applications. Proceedings of the 13th international
conference on ubiquitous computing, (2011), 415.
~80%
~99%
55
63. 0 0.5 1 1.5 2 2.5 3 3.5 4
46
48
50
52
54
56
58
60
hours
psi finding the dishwasher
0.5 1 1.5 2 2.5 3 3.5 4
hours
template
X
n
tn
X
n
pn
59
dishwasher
64. 0 0.5 1 1.5 2 2.5 3 3.5 4
46
48
50
52
54
56
58
60
hours
psi finding the dishwasher
0 0.5 1 1.5 2 2.5 3 3.5 4
hours
template
pressure difference
time difference
X
n
tn
X
n
pn
59
dishwasher
65. 0 0.5 1 1.5 2 2.5 3 3.5 4
46
48
50
52
54
56
58
60
hours
psi finding the dishwasher
0 0.5 1 1.5 2 2.5 3 3.5
46
48
50
52
54
56
58
60
hours
psi
template
pressure difference
time difference
X
n
tn
X
n
pn
59
dishwasher
66. 0 0.5 1 1.5 2 2.5 3 3.5 4
46
48
50
52
54
56
58
60
hours
psi finding the dishwasher
pressure difference
time difference
X
n
tn
X
n
pn
59
dishwasher
67. 0 0.5 1 1.5 2 2.5 3 3.5 4
46
48
50
52
54
56
58
60
hours
psi finding the dishwasher
pressure difference
time difference
X
n
tn
X
n
pn
X
n
tcyclelaundry machine
59
laundry
dishwasher
70. leveraging unlabeled data
labeled unlabeled
classifier classifier
feature set 2
high confidence high confidence
agree?
feature set 1
self labeled
multi-view classification
62
71. training 48%
49%
labeled unlabeled
multi-view classification
self labeled
TB
SVM+CRF
55%
53%
feature split: hot sensor vs. cold sensor
88%
90%TB
SVM+CRF
99%
dense features
sparse features
63
training
training
training
72. 0 5 10 15 20
46
48
50
52
54
56
58
60
time of day (hours)
psi
kitchen sink, hot master bathroom toilet
multi-view classification
64
85. semi-supervised learning
leveraging the homeowner
can we leverage multi-view models?
• select low confidence examples
• ask homeowner for label
AT&T LTEAT&T LTE 5:23 PM
did you just use water?
YesNo
75
86. semi-supervised learning
leveraging the homeowner
can we leverage multi-view models?
• select low confidence examples
• ask homeowner for label
AT&T LTEAT&T LTE 5:23 PM
did you just use water?
YesNo
75
AT&T LTEAT&T LTE 5:23 PM
Fixture Selection
Do Not Disturb
Select Notification Times
OFF
Master Toilet
Recently Used:
Half Bath Toilet
Dishwasher
Master Sink Select Temp.
Kitchen Sink Select Temp.
Master Shower Select Temp.
More
Half Bath Sink Select Temp.
87. simulating labels from homeowner
AT&T LTEAT&T LTE 5:23 PM
did you just use water?
YesNo
• ask for two labels every other day
• one morning and one evening
• only from 8AM-9PM
• randomly ask for previous event
76
AT&T LTEAT&T LTE 5:23 PM
Fixture Selection
Do Not Disturb
Select Notification Times
OFF
Master Toilet
Recently Used:
Half Bath Toilet
Dishwasher
Master Sink Select Temp.
Kitchen Sink Select Temp.
Master Shower Select Temp.
More
Half Bath Sink Select Temp.
88. 10 15 20 25 30 35 40 45
0.65
0.7
0.75
0.8
0.85
Co−Labeling in H1
Number of Labels
ValveLevelAccuracyofCoLabel−HMM
Co−Labeling
Random Labeling
co-labeling
random labeling
iteration 1
iteration 3
iteration 5
iteration 10
simulating labels from homeowner
co-labeling for H1minimaltrainingset totals
days
33
33
30
27
33
156
events
2374
3075
4754
2499
2578
14,960
events/day
71.9
93.2
158.5
92.6
78.1
95.9
compound
22.2%
21.8%
16.6%
32%
21.3%
21.9%
77
91. implications for homeowner
week one
• homeowner installs system
• 1-2 examples per fixture
• show sparse classes:
dishwasher, shower, laundry
80
92. implications for homeowner
week one
• homeowner installs system
• 1-2 examples per fixture
• show sparse classes:
dishwasher, shower, laundry
week two
• 2-4 labels, every 2 days
• fixture category: 85%
80
93. implications for homeowner
week one
• homeowner installs system
• 1-2 examples per fixture
• show sparse classes:
dishwasher, shower, laundry
week two
• 2-4 labels, every 2 days
• fixture category: 85%
week three
• 9-12 more examples
• fixture: 82%
80
94. implications for homeowner
week one
• homeowner installs system
• 1-2 examples per fixture
• show sparse classes:
dishwasher, shower, laundry
week two
• 2-4 labels, every 2 days
• fixture category: 85%
week three
• 9-12 more examples
• fixture: 82%
end of week three
• fixture: 87%
• valve: 80%
80
95. summary contributions
• comprehensive disaggregated dataset
• multi-view classification
• expert knowledge
• compressed sensing
• framework for virtual evidence in IMS
• co-labeling with multi-view
• idea: inception to industry ready
81
96. how can indirect sensing and machine
learning be used to reduce our
environmental footprint?
97. disaggregated hot and cold water sensing with minimal calibration
semi-supervised training for infrastructure mediated sensing
UbiComp Lab
electrical
engineering
computer
science and engineering
University of Washington
eclarson.com
eclarson@uw.edu
@ec_larson
Eric C. Larson
98. disaggregated hot and cold water sensing with minimal calibration
semi-supervised training for infrastructure mediated sensing
UbiComp Lab
electrical
engineering
computer
science and engineering
University of Washington
eclarson.com
eclarson@uw.edu
@ec_larson
Eric C. Larson