Efficient genetic-wrapper algorithm based data mining for feature subset selection in a power quality pattern recognition application

Abstract EN

Power quality monitors handle and store several gigabytes of data within a week and hence automatic detection, recognition and analysis of power disturbances require robust data mining techniques.

Literature reveals that much work has been done to evolve several feature extraction and subsequent classification techniques for accurate power disturbance pattern recognition.

However the features extracted have been rarely evaluated for their usefulness.

The objective of this work is to emphasize that feature selection is an important issue in power quality disturbance classification and that genetic algorithms can select good subsets of features.

In this paper, a wrapper based approach that integrates multiobjective genetic algorithms and the target learning algorithm is presented in order to evolve optimal subsets of discriminatory features for robust pattern classification.

The wavelet transform and the S-transform are utilized to produce representative feature vectors that can accurately capture the unique and salient characteristics of each disturbance.

In the training phase the multiobjective genetic algorithms is used to find a subset of relevant attributes that minimizes both classification error rate and size of the classifier discovered by the classification algorithm, using the Pareto dominance approach.

Two different classifiers were compared in this study using genetic feature subset selection: decision tree, a feed forward neural network.

Moreover two different MOGAs namely elitism-based MOGA and Non-dominated sorting genetic algorithm have been employed separately in the training phase.

Experimental results reveal that both of these proposed variants of MOGA combined with classifiers namely decision trees / FFNN yield improved classification performance and reduced classification time as compared to standard classifiers namely decision trees decision tree or standard feed-forward networks.

Moreover NSGA performs better than the elitism based approach in terms of classification time.