Abstract:

To eliminate redundant features in original fault feature space, a novel feature selection algorithm based on support vector data description (SVDD) and modified genetic algorithm is proposed. Firstly, it constructs a complete acoustical fault feature space through theoretic analysis and experiments. According to established criterion of feature separation and SVDD classifier prediction accuracy, prior knowledge is extracted from training data set and used as initialization to improve the efficiency of genetic algorithm. Inner and intraclass distance criterion and classifier prediction accuracy are introduced to establish fitness function and evaluate the degree of importance of every gene, thus the optimized feature subset is obtained. Experiments with unbalancefault data set simulated on rotor vibration testbed show that the proposed algorithm can improve the diagnosis accuracy.

Key words: fault diagnosis, feature selection, genetic algorithm, support vector data description