Feature extraction based on PSO-FC optimizing KPCA and wear fault identification of planetary gear

The feature extraction problem of coupled vibration signals with multiple fault modes of planetary gear has not been solved effectively. At present, kernel principal component analysis (KPCA) is usually used for nonlinear feature extraction, but the blind setting of kernel function parameters greatly affects the performance of KPCA algorithm. For the optimization of kernel parameters, it is necessary to study theoretical modeling to improve KPCA performance. In this paper, employing a Fisher criterion (FC) discriminant function in pattern recognition, the optimization mathematical model of the kernel parameter was presented and the improved particle swarm optimization algorithm (PSO) was applied to search for the optimum value, and the performance of the Kernel principal component analysis for nonlinear problems was improved. The optimized KPCA was applied for feature extraction of different wear fault modes of a planetary gear, and the feature dimensions were reduced from 27 to 10. The feature parameters with 92.9 % contribution rates were retained and sample sets were formed to feed the support vector machine (SVM) for final classification and identification. The intelligently optimized KPCA based on the PSO-FC has improved the structural distribution of data in the feature space and showed a good scale clustering effect in planetary gear wear state recognition. The accuracy of the SVM classification was improved by 17.5 %.