Resilient fast convolutional sparse filtering with convergence directivity for the bearing fault diagnosis

Abstract

The fast nonlinear convolutional sparse filtering (FNCSF), which detects fault features 42 h earlier than traditional methods in the intelligent maintenance system dataset and takes only 0.12 s, is widely considered a powerful tool for early fault diagnosis. However, random shocks caused by the structure or external interference pose challenges to the extraction accuracy of FNCSF. In addition, the extraction reliability of FNCSF is affected by computational instability and complex parameter settings. To address the above defects, resilient fast convolutional sparse filtering (RFCSF) is proposed in this study. First, the collected vibration signal is normalized by Z-score to eliminate the scale variability of the nonlinear transformation. Then, the frequency components of the signal are evenly divided by the initialized filters to indicate the convergence direction of fault and improve random shock resistance and extraction stability. Next, the nonlinear $l_{\frac{1}{2}}$ norm, which can effectively distinguish fault features from irrelevant disturbances, is regarded as the target of blind deconvolution. In the process of deconvolution, the filter converging to random shocks and noise is adaptively eliminated to improve the extraction efficiency and intelligence. Finally, the filtered signal is envelope demodulated to obtain fault information. In the simulation analysis, features under noise with −16 dB and random shocks with 50 m/s² are accurately extracted by RFCSF with parameters robustness. Several experimental cases demonstrate that RFCSF with these advantages is a promising feature extraction tool.