You are on page 1of 23

Fault Detection and Diagnosis Of

Air-Conditioning Systems Using


Residuals

Ammu Chathukulam
2016EEA2472
Content
Objective
Introduction
Model based fault diagnosis scheme
Classification approaches for residual patterns
LS-SVM Classifier
Hexadecimal residual code approach
Hybrid Classification Approach
Results and Discussions
Conclusion
Objective
To develop Model based fault diagnosis
schemes using residuals for diagnosing
major AC faults.

To classify the residual patterns for


identifying faults using LS-SVM
classification, hexadecimal residual code
and hybrid classification approach.
Introduction
A fault is detected when the observed behaviour
of a system differs from the expected behaviour
by some threshold.

The main purpose of fault detection and


diagnosis (FDD) is to detect, locate and if
possible, predict the presence of the defects
causing faulty operation.
Model based fault
diagnosis scheme
Model based fault diagnosis
scheme
Primary requirement: Fault free model of the plant.

RESIDUAL : Difference between output variables of the


actual plant and fault free model of the plant under same
input excitations.

The output variables, sensitive to the faults are selected .

For each faulty condition residuals form a particular pattern.

Classification of the residual patterns is required for


diagnosing the fault.
Classification approaches for
residual patterns
LS-SVM Classification Approach

Hexadecimal Residual Code approach

Hybrid Classification Approach


LS-SVM Classifier

y ( x) sign[ k 1 k yk K ( x, xk ) b]
N

{ yk , xk } Training set

k Support
value(scalar)

K Kernel function(RBF Kernel )


Start

Pre-processed
Residual Pattern

Layer1 F0

Layer13 F12

Preventing
Actions
LS-SVM Classifier
The training input data is pre-processed to make the
training of LS-SVM network more efficient. (Zero mean
and unit variance).

RBF kernel is selected for LS-SVM classifier.(More


diagnosis accuracy).

The more frequent faulty condition is placed in the


upper layer of the FDD system to improve the
performance of FDD.

Number of LS-SVM layers =Number of fault modes


Demerits
One layer of LS-SVM network for diagnosing one fault
mode.

With the increase of fault modes, number of layers will


also increase linearly.

Centralized air-conditioning system : Possible faults will


be in the order of hundred.

Increases Computational burden.


Hexadecimal residual code
approach
Hexadecimal residual code
approach

Hexadecimal Codes are generated using these flags.

Hexadecimal number is compared with the


hexadecimal number stored in the fault library and
this will give fault status.
Merits
Can be used for single as well as multiple faults.
Multiple fault pattern is a linear combination of the
single fault patterns.(Except when patterns cancel each
other)

Single fault : Either component or sensor is faulty at a


time.
Multiple fault : Both sensor and component can be
faulty simultaneously.
Demerits

Fixed Threshold : Difficult to get good diagnosis


accuracy. Residual may be more than the threshold, due
to reasons other than the fault .

Selection of optimum threshold is very challenging

Hybrid Classification
Approach
Hybrid Classification Approach
Mixing LS-SVM and the hexadecimal code approach.

The residuals are fed to parallel layers of LSSVM


classifier.

Each layer corresponds to a different residual (8 layers ).

When a residual will corresponds to a fault, then output


of the LSSVM layer will be 1 otherwise it will be 0
Pre- LSSVM 0
/
r1-r8
processor r1-r8
network 1

Pre- LSSVM
r1-r8 Out
processor network Decision put
logic

Hexadecimal
Code
generation

Hexadeci
mal
decision
table

r1-r8
Pre- LSSVM
processor network
Merits
Combines the merits of both the methods.

No threshold selection is required.

Number of LS-SVM layers does not depend upon the number


of fault modes.

Can be used for systems with higher number of fault modes.


Results and Discussions
12 different fault modes and a fault-free mode is diagnosed.

428 Testing data.

LS-SVM based classifier has better diagnosis accuracy and at the same
time more robust against noises than the neural network based
classifier.

Performance of LS-SVM classifier


Results and Discussions
The hexadecimal residual code approach has poor diagnosis
accuracy in comparison to LSSVM classifier approach.

Variation in the threshold affects diagnosis accuracy badly.

Performance of hexadecimal code approach.


Results and Discussions
The hybrid approach has excellent diagnosis
accuracy in absence of noise as well as presence
of noise.

Performance of hybrid classifier


Conclusion
The diagnostic accuracy of FDD scheme using hybrid
classification approach is found excellent in presence as
well as absence of noise.

This approach can also be used for diagnosing faults in


processes other than AC system.

The limitation of these FDD schemes is that the


diagnosis accuracy depends upon accuracy of the fault
free model.
Thank You

You might also like