Bearing Fault Detection: Feature Selection Algorithm Efficacy Study

Abstract

Early detection of bearing failures is crucial, requiring a comprehensive diagnostic scheme with relevant features and an effective classifier. Computationally intelligent algorithms (CIA) contribute to enhancing the feature selection process. In this study, 99 classification datasets with classes ranging from 4 to 48 were derived from vibration recordings of Case Western Reserve University. 18 features are extracted in the time domain, of which 12 are statistical features and 6 are time-based frequency features (TBFF). Extracted features are ranked by mutual information, multi-cluster feature selection, and Laplacian scores. The best feature subsets are explored by four CIAs: ant colony optimization, simulated annealing, particle swarm optimization, and wheel-based differential evolution(WBDE) algorithms. The top 4 features of every approach are evaluated with 99 datasets. It was discovered that the feature subset recognized by most of the methodologies is the energy of the 4th derivative, waveform length ratio, sparseness, and mean. This combination results in 92.4% (with 4 features) and 88.94%(with 3 features) accuracy for classifying 48 class data, whereas the deep learning model resulted in 85.6% (with 18 features). Remarkably, it is noted that the feature subset identified by WBDE is identical to the feature subset arrived by voting. Finally, the results are compared to similar research in the literature, and improvements in classification accuracy are revealed.