Evolutionary strategy for classification problems and its application in fault diagnostics

Abstract

Genetic algorithms (GAs) based evolutionary strategy is proposed for classification problems, which includes two aspects: evolutionary selection of the training samples and input features, and evolutionary construction of the neural network classifier. For the first aspect, the GA based k-means-type algorithm (GKMT) is proposed, which combines GA and k-means-type (KMT) to achieve the optimal selection of the training samples and input features simultaneously. By this algorithm, the “singular” samples will be eliminated according to the classification accuracy and the features that facilitate the classification will be enhanced. On the opposite, the useless features will be suppressed and even eliminated. For the second aspect, the hierarchical evolutionary strategy is proposed for the construction and training of the neural network classifier (HENN). This strategy uses the hierarchical chromosome to encode the structure and parameters of the neural network into control genes and parameter genes respectively, designs and trains the network simultaneously. Finally, the experimental study pertained to the fault diagnostics for the rotor-bearing system is given and the results presented show that the proposed evolutionary strategy for the classification problem is feasible and effective.