Download CF for coordinated monitoring in WSN

Collaborative Filtering for
Coordinated Monitoring in
Sensor Networks
Janne Toivola & Jaakko Hollmén
-Department of Information and Computer Science,
Aalto University School of Science,
Finland
1
lauantaina 10. joulukuuta 2011
An idea...
Collaborative filtering useful in a wireless
sensor network?
Recommendation
system
Wireless sensor
network
???
Content based...
Collaborative filtering
Variety of methods...
2
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Outline
Background
-Structural Health Monitoring (SHM)
-Wireless Sensor Networks (WSN)
Coordinated monitoring problem
-Collaborative Filtering (CF)
-Random assignment and Majority voting
Experiments and results
-Wooden model bridge
-Aluminum bookshelf
Summary
3
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Background
Multidisciplinary research project
-Intelligent Structural Health Monitoring System (ISMO)
-Applied mechanics, data mining, sensor networks...
Structural Health Monitoring
-Assessing the condition of man-made structures
Data mining
-Extracting relevant information from data
Wireless Sensor Networks
-Distributed computing environment
4
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Structural health monitoring
No sensors that directly indicate damage
-Extracting features from vibration measurements
-Sensitive to damages, not to environmental variability
-Transmissibility magnitude: how well vibrations of
certain frequency propagate between two sensors
Novelty detection problem
-Not allowed to damage the (unique) structure to collect
training samples
-Supervised feature selection not possible
-Maybe some additional information available?
5
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Wireless sensor networks
Eliminate costly and error-prone wires
Restricted bandwidth and computation
-Simple features can be computed locally on a node
-Number of transmitted features should be low
Each sensor node should concentrate on
measuring relevant features!
Relevance depends on what the other nodes
are measuring => coordination problem!
6
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Coordinated monitoring
Monitor at most D local features / node
-An upper limit on computation & communication
-Sparsity helps also the detection algorithm
Each local feature can be given a rating
-Encodes additional application-specific information
-The above restriction holds: only D ratings at a time
Computation of combined features
-Requires measuring certain sets of local features
-Local ratings combined to select monitored features
7
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Collaborative filtering
Traditionally in recommendation systems
-Users give ratings to items
-Recommend new items to a user based on ratings
from other similar users
Assumes structure in the rating data
-Similar users like prefer same items
Needs to deal with the sparsity of ratings
WSN: recommend features to nodes!
-some similarities & differences to conventional CF...
8
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Baseline: random & voting
What if it doesn’t matter which features to
monitor / all are equally good?
-Let’s try also random selection, at most D/node
What if the sensors don’t need to
specialize / just one global choice enough?
-Let’s try majority voting based on the same ratings
-Number of combined features may still be high due to
the number of combinations...
9
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
The proposed architecture
Accelerometers:
Distributed sensor nodes
Acceleration
time series
-time series xs[n]
DFT
Local features:
Coordinated by CF
-power spectrum Xs[k]
Transmissibility
Combined features:
-transmissibility
T[s1,s2,k] = Xs1[k] / Xs2[k]
Novelty detector
Centralized user
Novelty detector:
k-NN or Gaussian
10
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Frequency domain ratings
-run FFT etc.
-over t time windows
Sensor 8
0.2
Acceleration
Mean power
spectrum Xs[k]
0.1
0
−0.1
−0.2
0
0.1
0.2
0.3
0.4
0.5
Time (s)
0.6
0.7
0.8
0.9
1
−4
x 10
Power
Sort by power
15
10
5
For top-D bins
20
40
60
Frequency (Hz)
80
100
120
0
20
40
60
Frequency (Hz)
80
100
120
10
Rating
-r[s,k] += 1
-Default vote 0
0
5
0
Repeat MCF times
11
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Combined ratings and CF
After collecting local ratings r[s,k]
Compute ratings for the combined features
-w[s1,s2,k] = corrj(r[s1,j], r[s2,j]) * r[s1,k] * r[s2,k]
-similarity of sensors s1 and s2 measured by correlation
-symmetric, since transmissibility is ”symmetric”
Sort and select top-D combined features for
each sensor node
-some sensors may be left with less than D features, if
other nodes are fully utilized
-no eliminated items like in recommendation systems
12
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Data: Wooden bridge [19]
A
1
A
7
4
10
2
B
13
weights
A
8
5
11
3
A
6
12
9
shaker
14
15
C
15 accelerometers and a shaker
Added weights to simulate damages
1998 + 265 time series measurements
-32 seconds • 256 Hz = 8192 samples each
13
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Example: Random, D=4
14
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Example: Majority vote, D=4
15
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Example: CF, D=4
16
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Example: CF, D=20
17
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Novelty detection accuracy
AUC values with kïNN
1
Iterated 10
times
0.9
CF, vote,
random
k-NN
detector
min, med,
max AUC
18
AUC
D=2...20
0.8
0.7
0.6
0.5
Random
Vote
CF
0.4
0
50
100
150
200
250
300
number of transmissibility features
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
350
400
450
Novelty detection accuracy
AUC values with Gaussian
1
Iterated 10
times
0.9
CF, vote,
random
Gaussian
detector
min, med,
max AUC
19
AUC
D=2...20
0.8
0.7
0.6
0.5
Random
Vote
CF
0.4
0
50
100
150
200
250
300
number of transmissibility features
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
350
400
450
Data: LANL bookshelf [20]
18
17
19
20
10
9
11
12
2
1
3
4
24 accelerometers
damages
24
23
15
16
22
21
14
13
6
5
7
8
A
20
C
shaker
Loosened and
removed bolts
Shaker at the bottom
150 + 120 time series
measurements
-5.12 sec • 1600 Hz =
8192 samples each
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
Novelty detection accuracy
AUC values with kïNN
1
Iterated 10
times
0.9
CF, vote,
random
k-NN
detector
min, med,
max AUC
21
AUC
D=2...20
0.8
0.7
0.6
0.5
Random
Vote
CF
0.4
0
200
400
600
800
number of transmissibility features
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
1000
1200
Novelty detection accuracy
AUC values with Gaussian
1
Iterated 10
times
0.9
CF, vote,
random
Gaussian
detector
min, med,
max AUC
22
AUC
D=2...20
0.8
0.7
0.6
0.5
Random
Vote
CF
0.4
0
200
400
600
800
number of transmissibility features
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011
1000
1200
Summary
Problem of coordinated monitoring
Collaborative filtering as a solution
Applied as part of SHM system
Demonstrated with two data sets
-8192•S acceleration samples => one detection result
-CF performed well with wooden bridge data
-Differences smaller with the LANL bookshelf data
Some CF problems remain: coverage
23
IEEE ICDMW DaMNet 2011: Collaborative Filtering for Coordinated Monitoring in Sensor Networks
lauantaina 10. joulukuuta 2011