internal fault classification using artificial neural network

8/7/2019 Internal fault classification using Artificial Neural Network

1/17

INTRODUCTION

An artificial neural network (ANN), usually called neural network (NN), is

a mathematical model or computational model that is inspired by the structure

and/or functional aspects of biological neural networks. A neural networkconsists of an interconnected group of artificial neurons, and it processes

information using a connectionist approach to computation. In most cases an

ANN is an adaptive system that changes its structure based on external or

internal information that flows through the network during the learning phase.

Modern neural networks are non-linear statistical data modelling tools. They are

usually used to model complex relationships between inputs and outputs or

to find patterns in data.

The layers network through the mathematics of the system algorithms. The

network function is defined as a composition of other functions , which

can further be defined as a composition of other functions. This can beconveniently represented as a network structure, with arrows depicting the

dependencies between variables. A widely used type of composition is

the nonlinear weighted sum, where , where (commonly

referred to as the activation function) is some predefined function, such as

the hyperbolic tangent. It will be convenient for the following to refer to a

collection of functions as simply a vector .

These are essentially simple mathematical models defining a function ora distribution over or both and , but sometimes models are also intimately

associated with a particular learning algorithm or learning rule. A common use

of the phrase ANN model really means the definition of a class of such

functions (where members of the class are obtained by varying parameters,

connection weights, or specifics of the architecture such as the number of

neurons or their connectivity).


2/17

The i ternal fault are diffi ult t be penetrated because they are much more

complex. To find outthe internal faults, one must know the law ofthe internal

faults attributing to the relation of their infrared thermograph characteristics.

The internal faults of any electrical equipment can be di ided into loose

connection or contact of internal conductors and inferiority in insulation andother faults. The internal fault doesnt have protection system and internal fault

can be detected by using infrared thermograph camera. People and structure

being exposed to the danger of internal fault, it results in the need for internal

fault prediction system.

In many cases the true source ofthe temperature rise is not visible eitherto the

human eye or the infrared camera, making these measurements indirect. The

decision to conductthe corrective actions rest s on the maintenance team, taking

into consideration other factors, such as equipment's criticality, safety and the

availability of parts.Artificial neural network (ANN) is a branch of artificialintelligence (AI). This

system is based on the operation of biological neural network, in other words, it

is an emulation of biological neural system. In engineering, neural networks

serve two important functions i.e. pattern classifiers and nonlinear adaptive

filters. ANN has several advantages. Some of the advantages of the ANN are

thatit can perform tasks that a linear program cannot and it does not need to be

reprogrammed. Besides, ANN can be implemented in any application without

any problem. But, ANN also comes with several disadvantages such as it needs

training to operate and require high processing time for large network. It alsoneeds to be emulated due to different architecture with microprocessor.

This Paper presents ANN-based technique to categorize the fault system based

on temperature and RGB colour image data of the fault picture. In order to

determine the best ANN model, several feed-forward back-propagation ANN

designs were constructed and each network performance was evaluated by using

cross-validation technique. The ANN model with the best performance is

selected as the best design. As a result, the developed fault category system is

generalized well when presented with new sets of input data. Finally,

categorization the faultinto four categories could be done successfully.


3/17

METH L GY

In general, ANN-based technique is used to categorize the fault system based on

temperature and RGB colour image data of the fault picture. In order to

determine the best ANN model, several feed-forward back-propagation ANNdesigns were constructed and each network performance was evaluated by using

cross-validation technique. The ANN model with the best performance is

selected as the best design. As a result, the developed fault category system i s

generalized well when presented with new sets of input data. The research

design for categorizing faultto four categories system can be grouped into three

main stages. They are data collection, data analysis using image processing and

development ofthe ANN.

A. Data CollectionIn this project, fourtypes internal faults are chosen and implemented as the data

base. These internal faults are low, intermediate, medium and high. The faults

data setis distributed as follows: 84 for low fault, 84 forintermediate fault, 84

for medium fault and 84 for high fault samples.

All samples were collected using infrared thermography camera. Table 1 shows

each group of fault with its respective temperature range. From the table we can

see that the internal fault is divided to four categories. These are low,intermediate, medium, and high internal fault. The internal fault can be

described as the different temperature between spot temperature and reference

temperature.

TABLE 1: Internal Fault Category


4/17

Type of internal fault implemented in this research.

Figure 1: Low Internal fault

Figure 2: Intermediate Internal fault


5/17

Figure 3: Medium InternalFault

Figure 4: High InternalFault


6/17

B. Data Analysis using Image Processing

Image processing is the most critical part since the RGB componentis used as

an inputto ANN. Ifitis done in improper manner, the recognition process will

become inaccurate. Impixelregion is one of the commands in image

processing tool box. Impixelregion creates a Pixel Region tool associated with

the image displayed in the current figure, called the target image. The Pixel

Region tool opens a separate figure window containi ng an extreme close-up

view of a small region of pixels in the target image. When the command

impixelregion is used, the pixel region rectangle will be displayed. The Pixel

Region rectangle defines the area of the target image that is displayed in the

Pixel Region tool. The pixel region rectangle can be moved over the target

image using the mouse to view different regions. To get a closer view of the

pixels displayed in the tool, use the zoom buttons on the Pixel Region tool

toolbar or change the size of the Pixel Region rectangle using the mouse.

Figure 6 shows the pixel region rectangle used in the targetimage. By using this

command, the RGB colour data can be obtained and used for the input of the

ANN. The figure also shows the scale colour which is employed to take the

RGB colour data which is based on certain temperature level.


7/17

C. Development of the ANN

Figure 8 shows a three-layer feed-forward back propagation ANNs, developed

in Matlab forthis research. There are three layers presenti.e. inputlayer, hidden

layer and outputlayer. Each layer consists of one or more nodes, represented in

the diagram by the small circles. The lines between the nodes indicate the flow

of information from one node to the next. In this particular type of neural

network, the information flows only from input to output. Figure 9 shows the

block diagram for categorized fault system. In the block diagram, temperature

data and RGB colour data are used as the input and internal fault category as the

target outputto the ANN. The developed ANN receives the input data and also

the targeted outputto produce network output.

Figure 8: ANN Diagram

Figure 8: Categorize Fault System Architecture


8/17

D. Training Process

In order to determine the best design, each network is trained and tested to

estimate its performance and determine which configuration provides the best

results. This can be done by using cross validation, a model evaluation method

that estimates generalization error based on re-sampling. It is a statistical

practice of partitioning a sample of data into subsets where the analysis is

performed on a single subset while the other subsets are retained for subsequent

use in confirming and validating the initial analysis. There are two types of

cross validation methods that can be considered. They are the holdout cross

validation and the K-fold cross validation. The holdout method is the simplest

kind of cross validation. The data set is separated into two sets, called the

training set and the testing set. The holdout cross validation uses the training set

only. The errors it makes are accumulated as before to give the mean absolute

test set error, which is used to evaluate the model . The advantage of this method

is that it is usually preferab le to the residual method and takes no longer to

compute. In K-fold cross-validation, the original sample is randomly partitioned

into Ksubset. Ofthe Ksubset, a single subsetis retained as the validation data

fortesting the model, and the remaining K 1 subset are used as training data.

The cross-validation process is then repeated Ktimes, with each ofthe Ksubset

used exactly once as the validation data. The Kresults from the folds then can

be averaged to produce a single estimation. The advantage ofthis method over

repeated random sub-sampling is that all observations are used for both trainingand validation, and each observation is used for validation exactly once. Figure

9 show the K 1 subsetis used as a validation set.

Figure 9: Partitioning of data for K = 4

E. Testing Process

Testing process is carried out to measure the performance of the trained

network, which can be measured to some extent by the errors on the training

and validation sets and the testing data or by performing a linear regression

analysis between the network response and the corresponding targets.


9/17

The regression coefficient, R with values close to one indicates that there is a

strong correlation between the targeted outputs and network outputs while the

values that are close to zero indicates otherwise. In order to measure network

performance in terms of its R-value, the best network is trained once again

using both training and validation sets as the whole training data. Itsperformance is then assessed using the testing data. A fully trained network

should be able to categorize fault into four categories from this set of unseen

data and is evaluated by measuring the R-value. Figure 10 shows the flowchart

for ANN algorithm.

The data were selected from RGB colour and temperature internal fault record.

336 data internal fault is selected, 168 were utilized for training network and

168 were employed for the testing process. The developed ANN receives the

input data and also the targeted output to produce network input. The feed-

forward back-propagation ANN were developed in Matlab. Networkconfigurations such as the number of neurons and transfer functions were

determined heuristically. For each network design, the learning rate was chosen

to be 0.5 while the momentum is constant at 0.9 (typical value). The value of

learning rate and momentum is kept constant in the holdout and K-fold

validation methods. After finishing the validation method, the number of neuron

and transfer function is selected depend ing on the R-value such thatitis close to

one.

In the holdout cross validation method, the training data is subdivided into 2

sets of data, training and validation sets. For this study, the K-fold crossvalidation have 168 training data and this data is again subdivided into 4 sets.

There are 3 sets for training and 1 set for validation. However the number of

fold depends on the size the data. When the number of fold is large, the resultis

more accurate.

The comparison between holdout and K-fold validation method depends on the

value R-value. When the R-value is close to one, it indicates that there is a

strong correlation between target output and network output. On the other hand,

ifthe R-value is close to zero itindicates otherwise. From table 2, the best

R-value for holdoutis 0.98683 as shown in figure 11. The best value, 0.9879 is

obtained from an average of K-fold as shown in figure 12 to figure 15. It can

also be seen that the network trained in using K-fold cross validation gives

better results than that using holdout method. By comparison, between holdout

and K-fold cross validation method, the R-value obtained by the network

trained using K-fold cross validation method was higherthan that using holdout


10/17

method. Based on these, it can be concluded that K-fold cross validation method

gives better than result than holdout cross validation method in term R-value.

The K-fold cross validation method can obtain the best R-value when it uses

fixed learning rate and momentum constant of 0.5 and 0.9 respectively

Figure 10: ANN Algorithm

Figure 10: ANN Algorithm


11/17

TABLE 2: Comparison Between Holdout & K-Fold Validation Method


12/17

Figure 11: The best R-value for Holdout Cross Validation Method

Figure 12: The R-value for partitioning of data for K=1


13/17

Figure 13: The R-value for partitioning g of data for K=2



14/17


The R-value is changed when the value learning rate and momentum is varied.

The K-fold cross validation method is use to getthe best of value learning rate

and momentum. From table 3, the best value oflearning rate and momentum are

0.1 and 0.5. The R-value obtain by this networkis 0.99381 as shown in Figure

16.

Figure 16: The best of ANN design


15/17

TABLE 3: Learning Rate and Momentum

From the previous observations, it can be concluded that the best ANN to

categorize internal fault can be divided to four categories as shown in table 4.The accuracy oftraining and testing depend on the percentage of the R-value.


16/17

TABLE 4: Properties to Develop a Network to Categorize Internal Fault into

Four Categories:


17/17

CONCL S ON

This research mainly presents a contribution in the field ofimage processing to

analysis the internal fault picture and to get the RGB colour data. This

information is extracted and then can be used to produce intelligent model

system for internal fault classification. The feed-forward back propagation

network developed in the ANN model categorizes the internal fault to 4

categories. The development of these networks required a certain degree of

experience because the network architecture is such thatthe number of neurons

and transfer functions needed to be determined heuristically. From the

comparison of cross validation method between holdout and K-fold, it wasconcluded that, the K-fold cross validation method produces better result for

ANN in terms R-value. Therefore, based on the result, to categorize internal

faultto 4 categories is generalized well when presented with new sets ofinput

data.

internal fault classification using artificial neural network

Documents