Adaptive sparse representation based on circular-structure dictionary learning and its application in wheelset-bearing fault detection

Wheelset bearings are among the crucial elements of bogie frames used in high-speed trains. Wheelset-bearing fault detection can actively reduce or preclude safety-related accidents and realize condition-based maintenance in high-speed train service. Therefore, it is of great significance to automatically detect wheelset-bearing faults. Sparse representations based on circular-structure dictionary learning (SRCSDL) provide an excellent framework for extracting fault impact trains (FITs) induced by wheelset-bearing faults. However, the performance of SRCSDL on extracting FITs heavily relies on the selection of method-related parameters. A systematic method for selecting such parameters has not been reported in the literature. A novel fault detection method, adaptive SRCSDL (ASRCSDL), is therefore proposed in this paper. The effects of the selection of each SRCSDL parameter on extracting FITs are investigated. It was found that three parameters (the length of single set signals, the number of signal sets, and convergence error) can be fixed according to the characteristics of the SRCSDL algorithm. To adaptively tune the remaining three parameters, main frequency analysis is used to select the number of kernel functions, the number of maximum extreme values is employed to determine the length of the kernel function, and envelope spectra kurtosis-guided self-tuning algorithms are proposed to tune the target sparsity of SRCSDL. The proposed method is then validated using the simulated signals and bench and real-line tests.