An Improved Multilayer Convolutional Sparse Coding Model for Fault Feature Extraction of Rolling Element Bearings Under Varying Speed Conditions

Rolling element bearings are critical components in rotating machinery, and their faults often lead to machinery failure. Accurate fault diagnosis under varying speed conditions remains a challenging task, especially when dealing with weak, nonperiodic fault features embedded in noise and in the absence of angle encoder. This article proposes an improved multilayer convolutional sparse coding (ML-CSC) model incorporating an overlap-crop (OC) operation for effective fault feature extraction under such conditions. The ML-CSC framework enhances the capability to extract weak transients by stacking multiple convolutional layers, while the OC operation mitigates boundary-induced transient effects and enables efficient parallel processing. The proposed method is validated using both simulated and real-world vibration signals under varying speed conditions. Compared with the existing approaches—including vanilla convolutional sparse coding (CSC), CSC-DFT, tunable Q-factor wavelet transform (TQWT), and WMSDL—the OC CSC consistently yields higher signal-to-noise ratio (SNR) improvements and better reconstruction accuracy. The use of multilayer dictionaries proves effective in capturing complex fault dynamics, with two- and three-layer models offering enhanced performance without significantly increasing model complexity. Future work will explore real-time implementation and extensions to more complex rotating machinery.