Soft Sensor for Faulty

Measurements Detection and

Reconstruction in Urban Traffic

Department of Adaptive systems, Institute of Information Theory and Automation, June 2010, Prague
Outline

Problem description
Soft sensors
Gaussian Process models
Soft sensor for faulty measurement
detection and reconstruction
Conclusions
Outline

Problem description
Soft sensors
Gaussian Process models
Soft sensor for faulty measurement
detection and reconstruction
Conclusions
Problem description

Traffic crossroad - count of vehicles

Inductive loop is a popular choice
Devastating for traffic control system
Failure detection and recovery of sensor
signal
Example of controlled network (Zliin shopping
centre, Prague)

Sensors on
crossroads
Failure: control
system has no means
to react
Possible solution: soft
sensor for failure
detection and signal
reconstruction
Soft sensors
Models that provide estimation of another
variable
`Soft sensor: process engineering mainly
Applications in various engineering fields
Model-driven, data-driven soft sensors
Issues:
missing data, data outliers, drifting data, data
co-linearity, different sampling rates,
measurement delays.
Outline

Problem description
Soft sensors
Gaussian Process models
Soft sensor for faulty measurement
detection and reconstruction
Conclusions
 GP model
Probabilistic (Bayes) nonparametric model.
GP model determined by:
Input/output data (data points, not signals)

(learning data identification data):

Covariance matrix:
 Covariance function

Covariance function:
functional part and noise part
stationary/unstationary, periodic/nonperiodic, etc.
Expreses prior knowledge about system properties,
frequently: Gaussian covariance function
smooth function
stationary function
Hyperparameters
Identification of GP model =
optimisation of covariance function
parameters
Cost function: maximum likelihood of data
for learning
GP model prediction

Prediction of the output based on similarity

test input training inputs
Output: normal distribution
Predicted mean
Prediction variance

-2 +2
Static illustrative example
Nonlinear function to be modelled from learning points
8 Nonlinear fuction and GP model
10
y=f(x)

Static example: 86
Learning points

6
4
4

y
2

y
2

9 learning points: 00

-2 Learning points
-2 2

Prediction
-4
f(x)
-6-4
-1.5
-1.5 -1-1 -0.5
-0.5 00 0.5
0.5 11 1.5
1.5 22
xx
Prediction error and double standard deviation of prediction

Rare data density 6 2

|e|

4
e

increased variance 0
-1.5 -1 -0.5 0 0.5 1 1.5 2

(higher uncertainty). x
GP model attributes (vs. e.g. ANN)
Smaller number of parameters
Measure of confidence in prediction, depending on data
Data smoothing
Incorporation of prior knowledge *
Easy to use (engineering practice)

Computational cost increases with amount of data

Recent method, still in development
Nonparametrical model

* (also possible in some other models)

Outline

Problem description
Soft sensors
Gaussian Process models
Soft sensor for faulty measurement
detection and reconstruction
Conclusions
The profile of vehicle arrival data
Modelling
One working day for estimation data
Different working day for validation
data
Validation based regressor selection
the fourth order AR model
(four delayed output values as regressors)
Gaussian+constant covariance function
Residuals of predictions with 3 band
 20
3
15

measurements
Queue length 10

0 Estimation
-5

-10
50 100 150 200 250 300 350 400 450
Samples
10
3
5 |e|
e

0
50 100 150 200 250 300 350 400 450
Samples

Residuals distribution
120

100

80

60

40

20

0
-8 -6 -4 -2 0 2 4 6 8
Value of residual
 20
3
15

measurements
10
Queue length

0 Validation
-5

-10
50 100 150 200 250 300 350 400 450
Samples
10
3
5 |e|
e

0
50 100 150 200 250 300 350 400 450
Samples

Residuals' distribution
100

90

80

70

60

50

40

30

20

10

0
-8 -6 -4 -2 0 2 4 6 8
Value of residual
Proposed algorithm for detecting
irregularities and for reconstruction the
data with prediction

Sensor fault:
longer lasting
outliers.
The comparison of MRSE for k-step-
ahead predictions
0.62

0.6

0.58

0.56
MRSE

0.54

0.52

0.5

0.48

0.46
0 5 10 15 20 25
Prediction steps

Purposiveness of the obtained model

(the measure of measurement validity, close-enough prediction,
fast calculation, model robustness)
Soft sensor applied on faulty data

80
3
60

Queue length

40 measurements
20

-20
50 100 150 200 250 300 350 400 450
Samples

15
corr. measurements
Queue length

10

0
50 100 150 200 250 300 350 400 450
Samples
Conclusions
Soft sensors: promising for FD and
signal reconstruction.
GP models: excessive noise, outliers, no
delay in prediction, measure of
prediction confidence.
The excessive noise limits the possibility
to develop better predictor.
Traffic sensor problem successfully
solved for working days.