27
1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of California, Berkeley

1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

  • View
    220

  • Download
    0

Embed Size (px)

Citation preview

Page 1: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

1

Distributed Online Simultaneous Fault Detection for Multiple Sensors

Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya

EECS, University of California, Berkeley

Page 2: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

2

Plan

1. Introduction

2. Problem Statement

3. Proposed Solution

4. Analysis and Implementation

5. Experiments

6. Conclusions

Page 3: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

4

Application: Freeway Traffic Management

ACCIDENT!

Measurement Backhaul Processing

Internet

Internet

Control & Info

Cellular

Cellular

Traffic Management

Center

TrafficControl

PeMShttp://pems.eecs.berkeley.edu

Page 4: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

5

Sensor System State

large oscillations

Page 5: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

6

Mean days before failure or working continuously (D4)

55% of loops work continuously for fewer than 20 days; none works for more than 50 days in 2004 vs. 20% in 2005.

0 20 40 60 80 100 120 140 160 1800

10

20

30

40

50

60

70

80

90

100

Days

% F

ails

Befo

reMean Days to Fail (per Sensor) Distribution

2004

2005

Page 6: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

7

Motivation: freeway monitoring sensors

One sensor per lane every 2 miles

Measures flow, occupancy every 30 seconds

Sensor failures are frequent

Non-stationary environment

Events: onset of traffic jam, accidents, sudden slowdowns

Page 7: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

8

Problem statement

Detect faulty sensors that report plausible values

Distinguish events from faults

– Events temporary sudden changes in measurements

– Faults lasting sudden changes in measurements

Real time detection

Each sensor uses only local data

Page 8: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

9

Proposed approach

Sensor Network Fault Graph Change Point Model

Score S is correlation with block length T samples

Change times have some known priors

Page 9: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

10

Model details

Change times have priors

Scores have joint change distributions

Link information strength

Page 10: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

11

Preview of results

Accounting for average time scale of physical events

Combining multiple sources of weak evidence

Importance of feedback for detection algorithms

Statistical modeling = feasible implementations

Page 11: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

12

Does it make sense?Empirical distributions from highway deployment

Working Faulty

Page 12: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

13

Does it make sense?Empirical distributions from highway deployment

Use Box-Cox transformation or conditional normal distribution (Kwon, Rice and Bickel, 03)

Page 13: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

14

Selection of block length T

Distinguish events from faults :

Rule: T > Average event duration

Tradeoff: T = minimum waiting time to detect

Page 14: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

15

Measuring the performance

Control false alarm:

Minimize Average Detection Delay (ADD):

time (n)

time (n)

Page 15: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

16

Single change point review

For minimize ADD

Single change point optimal rule [Shyrayev (1978)]:

Performance [Tartakovsky and Veeravali (2005)]:

Minimum delay achievable for all procedures with false alarm

At time n test:

Page 16: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

17

Model for analytic problems

Two sensors:

For each proposed procedure:

– Achieved false alarm

– Delay

X and Y represent aggregates of many links to working sensors

Among all procedures with false alarm , minimum delay?

Page 17: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

18

Delay performance lower bound

Theorem 1: For all procedures with false alarm for each sensor:

Page 18: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

19

Multiple sensor posterior rule (no feedback)Direct extension of single change rule:

Common link does not help

ZX Y1 2

Theorem 2:

Page 19: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

20

Multiple sensor rule (with one bit feedback)

Use shared link until either sensor thinks it has failed

ZX Y1 2

Page 20: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

21

What is procedure doing?

Over time, implicit averaging

ZX Y1 2

Over sensors, 1 bit summarizes other links information

Page 21: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

22

False alarm boundConfusion probabilities

Theorem 3 [Rajagopal et al, 2008]:

Page 22: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

23

Confusion probabilityTheorem 4 [Rajagopal et al, 2008]:

For example (using some simplifications):

and

Guarantee that

and

Page 23: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

24

Delay guarantee

Theorem 5 [Rajagopal et al, 2008]:

Page 24: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

27

Delay estimates

Symmetric (X and Y same distribution) method is optimal:

Fully connected i.i.d network:

Page 25: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

28

Two sensor network: confusion probability

Theory predicts covariance ratio > 2

Page 26: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

31

Fully connected network: fixed false alarm

Small False Alarm (theory is close!)

= 0.1

= 0.0001

Page 27: 1 Distributed Online Simultaneous Fault Detection for Multiple Sensors Ram Rajagopal, Xuanlong Nguyen, Sinem Ergen, Pravin Varaiya EECS, University of

33

Conclusions and future work

Change point framework is good for building algorithms for fault detection

Currently Caltrans collecting data by visiting sensors predicted broken

Developed tools for analysis of multiple change point problems

Simultaneous online multiple event detection