Resilient fast nonlinear blind deconvolution with uniform initialization for the bearing fault diagnosis

The extraction of defective bearing feature under sudden load variations and mutual interference among components is a challenging task. The key is to overcome the strong background noise and random shock disturbances. Fast nonlinear blind deconvolution (FNBD) with superior noise adaptability is considered as a powerful tool to tackle the challenge. However, the reliability of FNBD is reduced by misdiagnosis under random shock interference and computational instability. In addition, extraction performance of FNBD is affected by the setting of complex parameters. To address above issues and broaden the applicability of FNBD, resilient fast nonlinear blind deconvolution (RFNBD) is proposed. First, the impact of filter initialization on the extraction accuracy and stability of FNBD is studied. The results indicate that the FNBD converges to components in the signal that are close to the center frequency of the initial filter, and the robustness of FNBD is limited by the original initialization mode. Based on this, a novel initialization pattern is proposed to improve the robustness under random shock interference and computational stability. Subsequently, the inferior filter elimination strategy is introduced to enhance the extraction efficiency and intelligence of RFNBD. Finally, the superior robustness under variable parameters and extraction performance under strong interference of RFNBD is demonstrated by simulation and experiment. In the XJTU-SY datasets, the proposed RFNBD extracted fault characteristic frequency and its first four harmonics from 16,384 sampling points 11 min earlier than the traditional method.