Joint optimization of statistical and deep representation features for bearing fault diagnosis based on random subspace with coupled LASSO

Abstract

In bearing fault diagnosis, statistical features and deep representation features reflect the signal characteristics from different perspectives and demonstrate tremendous diagnostic potential. Nevertheless, previous studies have paid little attention to the heterogeneousity between statistical and deep representation features. Besides, directly combining these two kinds of features may also lead to redundancy and conflict, which may negatively affect the diagnostic performance. To address this issue, an enhanced random subspace method with coupled LASSO (RS-CL) is proposed in this paper to jointly optimize statistical and deep representation features. In the feature extraction stage, statistical features are constructed from the time-domain, frequency-domain and time-frequency domain, while deep representation features are extracted by bidirectional long short-term memory. In the model construction stage, an enhanced RS-CL method is developed to generate more efficient and diverse base classifiers. To verify the performance of the proposed RS-CL method, experiments were conducted on a bearing fault diagnosis data set provided by the University of Paderborn. The experimental results verify the effectiveness and feasibility of the proposed method.