Fault Diagnosis System for

cartcletchΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

120 εμφανίσεις

Fault Diagnosis System for

Wireless Sensor Networks

Praharshana Perera


Supervisors:

Luciana Moreira Sá de Souza



Christian Decker






Page
2

of 23

Outline

Introduction

Sensor Data Analysis


Data Correlation


Time Dependant Sensor Data Analysis

Approaches


Neural Network based Fault Detector


Rule Based fault Detector

Evaluation


Evaluation Neural Fault Detector


Evaluation Rule based Fault Detector

Conclusions and Future Work


Page
3

of 23

Introduction

Wireless Sensor Networks have the potential
to be used in the near future in industrial
applications:

Inventory Management


Items Tracking

Environment Health and Safety


Monitor Storage Regulations


Monitor Patient Conditions


Track Personnel (Workers in Hazardous Areas)




Page
4

of 23

WSN Failure in a Business Process

WSN

Effects of failures in a Business Process:


Economic losses


Contamination of the environment


Human life risk


Quality reduction


Maintenance costs



Page
5

of 23

Our Goal

Automatic
identification of incorrect sensor readings




Called value failures




Provide a higher maintainability to the business process by


Diagnosing failures before they propagate further to the rest of the system


Page
6

of 23

State of the Art Value Fault Detection for WSNs



Depend heavily on model assumptions and expert knowledge



Lack prior data analysis



Perform fault detection in nodes itself



Hierarchical detection does not provide value failure detection


but shift the task of fault detection to a more powerful device (sink)




WSN

WSN

WSN


Page
7

of 23

Neural and Fuzzy Models in Sensor Fault Detection

Advantages



Ability to
learn any complex system model



No assumptions on mathematical/statistical models



Less expert knowledge

Disadvantages



Require training time



Scalability for WSNs


Page
8

of 23

Analysis
-

Sensor Data


Page
9

of 23

Analysis
-

Incorrect Sensor Readings

4 abnormal peaks

of
temperature sensor data

Light sensor stuck in one
value


Page
10

of 23

Sensor Data Correlation

Metrics



Correlation coefficient



Multiple correlation coefficient

Gathered Data



Temperature, Light, and Movement data of 3 neighboring nodes



3 days



To reduce noise (especially movement and light)


Interpolation


Moving Average


Results



Sensor

Multiple correlation coefficient

Temperature

0.91

Light

0.93

Sensor

Sensor

Correlation coefficient

Temperature

Light

0.73

Temperature

Movement

0.69

Light

Movement

0.69

x

x

y

y

High

Low


Page
11

of 23

Time Dependant Sensor Data Analysis


Page
12

of 23

Neural Network based Fault Detector

A neural network has the capability of learning these patterns



Requires training data



A neural network is trained to identify


Too high (incorrect)


Too low (incorrect)


Normal (correct)


Temperature Sensor readings


Page
13

of 23

Rule based Fault Detector

Rule based fault detection algorithm



Rules search phase



Online fault detection phase


Rules are discovered automatically eliminating the need of an expert



Sensor

Data Statistics

σ


μ

R


r


Input

Rule Base



Threshold rules

Fuzzy rules

Fault

Detection

Output

Valid/Invalid


Page
14

of 23

Rules Search Phase

Threshold Rules



Expected values for a node for the time
period T


Mean
μ


Standard deviation
σ



Multiple correlation coefficient R


Correlation coefficient r

Threshold Rules

Search

Fuzzy Rules Search

Input

(Statistics for Time period
T
)

σ

μ

R

r

(Rules for Time period
T
)

Output

If T then
μ



X


If T and
σ

=

low

Then R = high

Fuzzy Rules



Relationships between statistics for a node
for the time period T


μ different sensors


σ

and R same sensor


r different sensors


Page
15

of 23

Fault Detection Phase

Time Period T

Sensor Measurements

μ

σ

R r

Sensor data

Preprocess

Rule Base



Threshold rules

Fuzzy rules

Threshold Rules

If no rule is rejected

If majority of the rules
is rejected

Else

correct

incorrect

Fuzzy Rules

Validate corresponding fuzzy rules

If rejected

incorrect


Page
16

of 23

Evaluation

Experiment setup



32 nodes (uParts) deployed on the ground floor



Data collected for a time period of 23 days (3 for training)

Evaluation Metrics



False positive effectiveness (
FPE
) = actual unreliable / identified unreliable



Fault detection effectiveness (
FDE
) = identified unreliable / unreliable


Page
17

of 23

Evaluation


Neural Fault Detector

Experiment results

Fault detection effectiveness (FDE)

False positive effectiveness (FPE)

0.75

0.80


Page
18

of 23

Evaluation


Rule based Fault Detector

Identified Rules


Temperature






Light


Page
19

of 23

Evaluation


Threshold Rules


Page
20

of 23

Evaluation
-

Number of Rejected Threshold rules


Page
21

of 23

Evaluation


Rule based Fault Detector

Example


Page
22

of 23

Conclusions and Future Work

Conclusions



Proved to be efficient on identification of failures



A new strategy to evaluate sensor readings in WSNs



Require less expert knowledge of the system



Ability to learn environment and system dynamics



Fault detection performed in back
-
end


Without putting burden on the nodes


Independent of any hardware platform :
-

Ideal for enterprise scenarios



Neural fault detector :
-

potential to be used in specialized scenarios



Rule based fault detector :
-

Any WSN scenario supporting the users
(operators)


Page
23

of 23

Conclusions and Future Work

Future Work



Evaluating the approaches within a second application trial


Long period of time


Introducing errors



Neural network to detect failures in light and movement sensors



Enhancements in the decision scheme in rule based detector


Voting or weighting mechanisms