Data-driven intelligent diagnosis of bearing faults using sparse feature extraction and SVM

Abstract

Rolling bearings are core supporting parts of rotating machinery. Practical fault diagnosis faces challenges including insufficient fault samples, difficult feature extraction and low identification efficiency. This paper proposes a data-driven method for time-frequency feature extraction and fault diagnosis. Firstly, short-time Fourier transform converts vibration signals into time-frequency maps to retain non-stationary signal characteristics. Secondly, an L2 regularized sparse filtering algorithm is established to extract sparse and highly distinguishable features through unsupervised learning. Finally, a soft margin support vector machine classifier is employed for fault classification, and the improved genetic algorithm is adopted to optimize model parameters. Validated on the CWRU bearing dataset, the proposed method achieves an overall diagnostic accuracy of 99.6% with limited training samples, outperforming traditional machine learning and some deep learning algorithms. It can realize high-precision fault detection under small sample scenarios and has promising engineering application value.