Hydro pervasive2011 eric_larson

A Longitudinal Study of Pressure
Sensing to Infer Real-World Water
Usage Events in the Home

Eric Larson, Jon Froehlich, Elliot Saba,
Tim Campbell, Les Atlas, James Fogarty
and Shwetak Patel

design: 
use: 
build:  ubicomp lab
university of washington 
university of washington  sustainability research

today‘s usage

refrigerator
0.3 gallons
kitchen sink
28 gallons
dishwasher
6.5 gallons

today‘s usage

shower
62.4 gallons bathroom sink 1 bathroom sink 2
4.2 gallons 0.8 gallons

bath
6.5 gallons

toilet
78.4 gallons

today‘s usage: hot vs. cold
shower
52.4 gallons
shower
10.4 gallons bathroom sink 1 bathroom sink 1 bathroom sink 2
3.2 gallons 1.2 gallons 0.8 gallons

bath bath
bathroom sink 2
6.5 gallons 0.0 gallons
2.4 gallons

toilet
78.4 gallons

assisted living applications

last active:
10:09PM
last active:
10:08PM
last active:
10:03PM

last active:
last active:
10:33PM
7:20AM

at ubicomp09, we introduced hydrosense

hydrosense
•  single, screw-on sensor
•  identifies fixture usage
•  estimates flow

Froehlich et al., UbiComp2009; Larson et al., PMC2010

hydrosense
uses pressure waves to
identify usage
upstairs toilet flush
bath open

kitchen sink
kitchen sink
hotdishwasher open
open
cold open

downstairs shower flush
downstairs toilet
open

ubicomp2009  
feasibility study

controlled experiments
•  2 researchers per site
•  5 trials per valve

experimental script
•  valve opened full stop
•  pause for ~5 seconds
•  valve closed

ubicomp2009 paper
successfully demonstrated the
potential of using pressure waves to
identify fixture usage
evaluation method
algorithm
staged experiments
each pressure wave treated

independently
all faucet handles were operated at
approximately the same flow rates
did not consider context of usage

all fixtures were tested in isolation
was not probabilistic

what we’re really interested in…

how well will hydrosense
perform on real-world
water usage data?

water tower 
water tower

bathroom 1
hose 
spigot 

compound events kitchen

incoming cold 
water from 
dishwasher
supply line 

thermal  
expansion  
utility water pressure
regulator tank 
meter
hot
water
heater
laundry
bathroom 2

pervasive 2011 contributions
1 longitudinal study of real-world water
usage and the resulting dataset

2 a new probabilistic approach to water
usage classification

3 demonstrate that this new approach
can accurately classify real world data

ground truth 5-week classification classification
sensors
deployment
algorithm
results

ground truth labels
manual
automatic
kitchen sink
kitchen sink
bathroom sink
cold cold hot hot cold cold
open close open close open close

water tower 

bathroom 1
hose 
spigot 

kitchen

incoming cold 
water from 
dishwasher
supply line 

thermal  
expansion  
utility water pressure
regulator tank 
meter
hot
water
heater
laundry
bathroom 2

how can we obtain water usage
labels at the valve-level?

this is actually a challenging question…

function across fixtures

kitchen sink
bathroom sink

bath shower

toilet

laundry basin washing machine
dishwasher

custom ground truth data
collection system

xbee wireless modem fixture usage sensor board

hall reed 3-axis unidirectional ball omnidirectional
effect switch accelerometer switch ball switch

collection system
“wake up” sensors



collection system
fixture handle
position sensors



collection system
xbee
push wireless
button ter
transmit

sage
fixture u ard
o
sensor b


modified or
kill-a-wa
tt thermist


deployment sites

residents
2
2
4
2
2

size
3000 sqft
750 sqft
1200 sqft
700 sqft
750 sqft

floors
3
2
2
3rd flr
6th flr

fixtures
17
8
13
8
8

valves
28
13
21
13
13

deployment infrastructure
ground truth sensor
on every valve
backend python
a laptop running hydroserver
hydrologger

http

pressure sensor

two pressure sensors per home
home 1
apartment 1

pressure
sensor 1
(cold point)
pressure
sensor 2
(hot point)

pressure
pressure sensor 2
sensor 1
(cold point)
(hot point)

pressure stream
red = hot line
blue = cold line
reed switches
high = active
low = inactive

hydrosense annotations
1. ground truth sensor

2. semi-automated label

3. review annotator
4. verification

5. final label

dual pressure sensor

toilet
bathroom faucet
kitchen faucet
dishwasher
washing machine
bath/shower
bath/shower diverter

5-week dataset

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%

22% of all events were compound

natural water use
toilet
70 kitchen sink kitchen sink
bathroom sink
pressure (psi)

50

30

bayesian inference
signal
behavior
P (S|V ) P (V )

R−1 N−1 K−1
ˆ ˆ
fr ( Sr |V r ) P (vn |vn−1 ) fp (vi ) f k ( v a , vb )
r=0 n=0 i∈β
/ k=0 a,b ∈β

(i) templates and (ii) bigram (iv) paired valve
signal features language model (iii) grammar
priors

term(i) template matching
unclassified event
event library

shower 
kitchen  faucet 
dishwasher 
toilet 
compare across multiple
signal transformations

signal transforms

unclassified event
unclassified event
event library

shower 
dishwasher 
toilet

unclassified event
event library

shower 
unclassified event
assess similarity

dishwasher 
toilet 
15%
signal transforms

55%

18%

9%

hydrosense
example pressure waves

upstairs toilet flush
bath open

kitchen sink
kitchen sink
hotdishwasher open
open
cold open

downstairs shower flush
downstairs toilet
open

filter transforms
50 

48 
raw pressure 46 
(psi)
44 
detrended
42 
0  5  10  15  20 
48 
signal transforms

derivative smoothed
pressure 46 
(psi)
44 
2 
derivative
(psi/s) 0 

‐2 

time (s)

filter transforms
50 

48 
raw pressure 46 
(psi)
44 
detrended
42 
0  5  10  15  20 
48 
signal transforms

derivative smoothed
pressure 46 
(psi)
bandpass 44 
derivative 4 
bandpass
derivative 0 
(psi/s)

‐4 
time (s)

filter transforms
50 

48 
raw pressure 46 
(psi)
44 
detrended
42 
0  5  10  15  20 
signal transforms

derivative
frequency

bandpass
derivative
30 
cepstral
cepstrum amplitude 10 

‐10 
index

unclassified event
event library

shower 
unclassified event

kitchen faucet 
dishwasher 
toilet 
detrended
signal transforms

derivative

bandpass
derivative

cepstrum

unclassified event
event library

unclassified event

kitchen faucet 
dishwasher 
toilet 
detrended detrended
55%
signal transforms

derivative derivative
85%
bandpass bandpass
65%
cepstrum cepstrum

90%

unclassified event
event library

unclassified event

kitchen faucet 
dishwasher 
template comparisons

toilet 
P (S|V )
detrended detrended
55%
R−1 N −1
ˆ ˆ
fr ( Sr |V r ) P (vn |vn−1 )
signal transforms

derivative r=0 derivative n=0
85%
(i) templates and (ii) bigram
signal features language model
bandpass bandpass
65%
cepstrum cepstrum

90%

term(i) signal features
event library
unclassified event

shower 

toilet 

bath

pressure drop
unclassified event

shower  toilet 

toilet 

bath

resonance tracking
unclassified event

shower  toilet 

toilet 

bath

resonance tracking
unclassified event
template comparisons
P (S|V )

R−1 N −1
ˆ ˆ
fr ( Sr |V r ) P (vn |vn−1 )
r=0 n=0

shower  toilet 

signal feature comparisons
toilet 

bath

term (i): templates and signal features
70
pressure

50
30 P (S|V )

R−1 N −1
P(kitchen sink hot open) 14%
ˆ ˆ
fr ( Sr |V r ) P (vn |vn−1 )
P(kitchen sink cold open) 3%
r=0 n=0
P(toilet open) 15%

P(kitchen hot/cold close) 2%

P(kitchen hot close) 6%

P(toilet close) 1%

term (ii): bigram language model
70
pressure

50
30 P (S|V )
P (S|V

R−1
R−1 N −1
−1
P(kitchen sink hot open) 14%
fr ( ˆ ˆ
fr ( Sr |V r ) P (vn |vn−1 ))
P (vn |vn−1
r=0
r=0 n=0
n=0
P(toilet open) 15%
(i) templates and
(i) templates (ii) bigram
(ii) bigram
signal features
language model



P(toilet close) 1%

70
pressure

50
30

P(kitchen sink hot open) 14% 4.6%
sequence 1

P(toilet open) 15% 4.3%
sequence 2

4.1%
sequence 3


P(toilet close) 1%

70
pressure

50
30
sequence 1
kitchen sink kitchen sink dishwasher bathroom sink kitchen sink toilet toilet kitchen sink
hot open cold open open hot close hot close open close hot close
sequence 2
kitchen sink shower toilet bathroom sink kitchen sink kitchen sink kitchen sink kitchen sink
hot open cold open open hot close hot close hot open hot close hot close

sequence 3
kitchen sink (S|Vtoilet
P ) bathroom sink bathroom sink kitchen sink (V kitchen sink
P ) kitchen sink toilet
hot open (S|V open
P ) hot open hot close hot close (V )hot open
P hot close close

R−1
R−1 N −1
N −1 K−1
K−1
ˆ ˆ
fr (Sr ||Vr )
ˆ ˆ
fr ( Sr V r ) P (vn |vn−1 )
P (vn |vn−1 ) fp (v )
fp (vii) fk ( va ,, vb )
f k ( v a vb
r=0
r=0 n=0
n=0 ∈β
i/
/
i∈β k=0 a,b ∈β
k=0 a,b ∈β
(i) templates and
templates and
(ii) bigram
(ii) bigram (iv) paired valve
signal features
features language model (iii) grammar
(iii) grammar
(iii) grammar (iv) paired valve
(iv) paired valve
language model priors
priors
priors

term(iii): grammar

1 opened closed
2 opened closed
3 temperature consistency
P (S|V )
P (S|V ) soft penalty
))
P (V
P (V
R−1
R−1 N −1
N −1 K−1
K−1
ˆ ˆ
fr (Sr ||Vr )
ˆ ˆ
P (vn |vn−1 ) fp (v )
f k ( v a vb
r=0
r=0 n=0
n=0 ∈β
i/
/
k=0 a,b ∈β
(i) templates and
templates and
(ii) bigram
signal features
(iii) grammar
(iv) paired valve
priors
priors

term(iii): grammar
70
pressure

50
30
sequence 1
sequence 2
bathroom sink shower toilet bathroom sink shower kitchen sink toilet kitchen sink
hot open cold open open hot close cold close hot open close hot close

sequence 3
hot open (S|V open
P hot close close

R−1
R−1 N −1
N −1 K−1
K−1
ˆ ˆ
fr (Sr ||Vr )
ˆ ˆ
P (vn |vn−1 ) fp (v )
f k ( v a vb
r=0
r=0 n=0
n=0 ∈β
i/
/
k=0 a,b ∈β
(i) templates and
templates and
(ii) bigram
signal features
(iii) grammar
(iv) paired valve
priors
priors

term(iii): grammar
70
pressure

50
30


hot open (S|V open
P hot close close

R−1
R−1 N −1
N −1 K−1
K−1
ˆ ˆ
fr (Sr ||Vr )
ˆ ˆ
P (vn |vn−1 ) fp (v )
f k ( v a vb
r=0
r=0 n=0
n=0 ∈β
i/
/
k=0 a,b ∈β
(i) templates and
templates and
(ii) bigram
signal features
(iii) grammar
(iv) paired valve
priors
priors

term(iii): grammar
70
pressure

50
30

kitchen sink toilet bathroom sink bathroom sink kitchen sink kitchen sink kitchen sink toilet
hot open open hot open hot close hot close hot open hot close close

kitchen sink (S|V ) sink
P kitchen dishwasher bathroom sink kitchen sink (V ) toilet
P toilet kitchen sink
P (S|V )open
hot open cold open hot close hot close (V ) open
P close hot close

R−1
R−1 N −1
N −1 K−1
K−1
ˆ ˆ
fr (Sr ||Vr )
ˆ ˆ
P (vn |vn−1 ) fp (v )
f k ( v a vb
r=0
r=0 n=0
n=0 ∈β
i/
/
k=0 a,b ∈β
(i) templates and
templates and
(ii) bigram
signal features
(iii) grammar
(iv) paired valve
priors
priors

term(iii): grammar
70
pressure

50
30


kitchen
P P (S|V dishwasher bathroom sink kitchen sink (V ) toilet
P (S|V )open)
hot open cold open hot close
P
hot close (V ) open
P close hot close

R−1
R−1R−1 N −1
N−1
N −1 K−1
K−1
K−1
fr (Srr(|Sˆr|) r )
ˆ ˆ
fr (f r |ˆrrV
ˆ V ˆ
S V ) P P n |vn−1 )
P (v(v|v|vn−1 )
(vn n n−1 ) fpp(v )
ffp(vii)
(v ffk (va ,av,bvb )
f( ( v a, v)
kk v b
r=0
r=0 r=0 n=0
n=0
n=0 ∈β
i/∈β
/
i/
k=0 a,b ∈β
k=0 a,b ∈β
(i) templates and
templates and
(i) templates and and
bigram
(ii)(ii) bigram
(iii) grammar
(iv) paired valve
(iv) paired valve
signal features
features
signal features
language model
language model (iii) grammar
(iii) grammar
priors
priors
priors

term(iv): paired valve priors
paired
70 

60  estimated flow volume
open close
50 
0  20  40  60 
bin frequency

toilet
toilet bath
faucet bath faucet

30 60 90 120 150 1 3 6 9 12
seconds estimated gallons

fixture usage duration
flow volume

70
pressure

50
30


kitchen
P (S|V )open)
P
hot close (V ) open
P close hot close

R−1
R−1R−1 N −1
N−1
N −1 K−1
K−1
K−1
fr (Srr(|Sˆr|) r )
ˆ ˆ
fr (f r |ˆrrV
ˆ V ˆ
S V ) P P n |vn−1 )
P (v(v|v|vn−1 )
ffp(vii)
f( ( v a, v)
kk v b
r=0
r=0 r=0 n=0
n=0
n=0 ∈β
i/∈β
/
i/
k=0 a,b ∈β
k=0 a,b ∈β
(i) templates and
templates and
bigram
(ii)(ii) bigram
(iii) grammar
(iv) paired valve
(iv) paired valve
signal features
features
signal features
language model
(iii) grammar
priors
priors
priors

toilet
kitchen sink kitchen sink
70 bathroom sink
pressure

50
30

P (S|V )
P (S|V ) )
P (S|V P (V )
P
P (V )
R−1
R−1R−1 N −1
N−1
N −1 K−1
K−1
K−1
fr (Srr(|Sˆr|) r )
ˆ ˆ
fr (f r |ˆrrV
ˆ V ˆ
S V ) P P n |vn−1 )
P (v(v|v|vn−1 )
ffp(vii)
f( ( v a, v)
kk v b
r=0
r=0 r=0 n=0
n=0
n=0 ∈β
i/∈β
/
i/
k=0 a,b ∈β
k=0 a,b ∈β
(i) templates and
templates and
bigram
(ii)(ii) bigram
(iii) grammar
(iv) paired valve
(iv) paired valve
signal features
features
signal features
language model
(iii) grammar
priors
priors
priors

three levels of granularity

1 valve level
e.g., upstairs bathroom faucet hot water activated

2 fixture level
e.g., upstairs bathroom faucet activated

3 fixture category level
e.g., faucet activated

hydrosense classification results
real-world water usage data
valve fixture fixture category
100%

80%

60%

40%

20%

0%
term (i)

100%

80%

60%

40%

20%
70.2%83.42642
70.2% 83.4% 93.6% 87.99408
95.4%
72.4% 88.0% 95.3%
72.5% 75.5% 89.4856
75.5% 89.5% 95.9%
0%
term (i) term (i)-(iii) terms (i)-(iv)

*error bars = std error *10-fold cross validation

100%

90%

80%

70%

60%
70.2% 83.4% 93.6% 72.5% 88.0% 95.3%
72.4% 95.4% 75.5% 89.5% 95.9%
75.5% 89.5%
50%


compound events
100%

90%

80%

70%

60%
65.6% 80.8% 93.6%
94.1%
70.2% 83.4% 94.0% 72.8% 87.2% 95.4%
96.1%
72.5% 88.0% 96.0% 72.3% 86.2% 95.2%
75.5% 89.5% 95.9%
50%


one sensor, terms(i)-(iv)
100%

80%

60%

40%

20%
75.5% 89.5% 95.9%
0%

one sensor, terms(i)-(iv) two sensors, terms(i)-(iv)
100%

80%

60%

40%

20%
75.5% 82.4% 89.5% 93.5% 95.9% 97.7%
0%
*error bars = std error *10-fold cross validation *terms (i)-(iv)

hydrosense training results
100

80

60

40
terms (i)-(iv)
20

0
0 2 4 6 8 10
Days of Training Data
*error bars = std error

hydrosense training results
100

90

80
terms (i)-(iv)
70

60
0 2 4 6 8 10
Days of Training Data
*error bars = std error

future work
1 additional features

2 segmentation

3 ease of training

A Longitudinal Study of Pressure
Sensing to Infer Real-World Water
Usage Events in the Home

Eric Larson
eric.cooper.larson@gmail.com

Jon Froehlich, Elliot Saba, Tim Campbell,
Les Atlas, James Fogarty and
Shwetak Patel

design: 
use: 
build:  ubicomp lab
university of washington 
university of washington  sustainability research

Hydro pervasive2011 eric_larson

Recommandé

Recommandé

Contenu connexe

Plus de Eric Larson

Plus de Eric Larson (20)

Dernier

Dernier (20)

Hydro pervasive2011 eric_larson