Abstract: As a critical component, bearings are difficult to excavate fault features due to the weak or coupled characteristics of multi-state fault features, which brings great challenges to the health monitoring of mechanical equipment. Therefore, a fault diagnosis method based on the sparse resolution of feature atoms is proposed to realize sparse feature resolution and diagnosis of multi-state faults. First, a feature atom filter construction method is proposed. The wavelet basis atoms are extracted based on the maximum kurtosis criterion of the reconstructed sub-band of the tunable Q-factor wavelet transform. The basis atoms are given periodic a priori features to match the multi-state fault components in the signal. Secondly, a multi-channel convolutional sparse multi-state fault feature resolution model is constructed. The sparse coefficient components are resolved based on convolutional sparse coding with multiple feature atoms and alternating direction multiplier method optimization. Reconstruct each fault feature component to realize the sparse resolution of multi-state fault components. Meanwhile, a sparsity parameterization method based on the combination of sparse signal kurtosis and energy is proposed. Finally, the results of the envelope spectrum analysis of the reconstructed components are used to determine the feature components in the signal for multi-state fault diagnosis.. The proposed method is experimentally verified through simulations, experiments and industrial pump signals, and has the advantage of better sparse resolution compared to the convolutional dictionary learning method.

Key words: multi-state faults, convolutional sparse parse, feature decoupling, feature atoms, fault diagnosis