Improving the performance of Wavelet based Machine Fault Diagnosis System using Locality Constrained Linear Coding

Abstract

Support Vector Machine (SVM) is a popular machine learning algorithm used widely in the field of machine fault diagnosis. In this paper, we experiment with SVM kernels to diagnose the inter turn short circuit faults in a 3kVA synchronous generator. We extract wavelet features from the current signals captured from the synchronous generator. From the experiments, it is observed that the performance of baseline system is not satisfactory because of the inherent non linear characteristic of the features. Feature transformation techniques such as Principal Component Analysis (PCA) and Locality-constrained Linear Coding (LLC) are experimented to improve the performance of the baseline system. Although PCA allows for choosing dimensions with maximum variance, the dimension reduction always contributes to underperformance. On the other hand, LLC uses a codebook of basis vectors to map the features onto higher dimensional space where a computationally efficient linear kernel can be used. Experiments and results reveal that LLC outperforms PCA by improving the baseline system with an overall accuracy of 25.87 %, 21.47 %, and 21.79 % for the R, Y, and B phase faults respectively.