Classification of Acoustic Emission Signals from Rail Cracks by Binary Particle Swarm Optimization Combined with SVM

In order to classify the acoustic emission (AE) signals of rail cracks more accurately and quickly, a significant problem to be solved is how to select the optimal feature set of AE signals. Two aspects are included in this problem, one is feature sets dimensionality reduction, the other is feature selection. In this paper, the method is proposed to find effective features and the best dimension of feature set, which based on an adaptive weight binary particle swarm optimization algorithm combined with support vector machine (BPSO-SVM) method. The optimization ability in the later stage of the algorithm has been enhanced to avoid falling into local minima. The error rate is taken as the optimization objective which is generated by SVM to obtain the weight of features. These weights are counted in several experiments to get the number of feature votes. Effective features are filtered by an adaptive threshold and form the optimal feature set. Experimental results show that the improved method is able to effectively reduce the feature set dimension, improve the classification accuracy, and boost the real-time performance in rail AE signal detection. The stability and validity of this method are verified by several experiments, which provide guidance for the AE signal classification and feature selection problems.