Unsupervised learning-driven intelligent fault diagnosis algorithm for high-end bearing

Abstract

<p indent="0mm">Spectral kurtosis is an effective method for rolling bearing fault diagnosis. However, it is highly sensitive to common accidental impact features in engineering, which frequently leads to feature extraction failure. To solve this problem, a robust evaluation method based on unsupervised learning is proposed to calculate the kurtosis of time series under the interference of multiple sporadic impulses (MSIs). First, dynamic signal samples are divided into equal intervals according to the characteristics of high energy concentration in the time domain. Second, each block is mapped to statistical parameter feature space (SPFS). Third, based on the significant statistical difference between MSI and other components in SPFS, an isolated 2-means clustering algorithm is proposed that realizes the self-location of clustering center and self-adaptive adjustment of clustering number and identifies and suppresses blocks receiving interference in the signal step by step. Finally, kurtosis information fusion is performed for the remaining samples, excluding MSI interference. A new method for high-end bearing fault diagnosis is proposed, which is combined with the robust kurtosis evaluation method and the multi-scale decomposition method. Through simulations and actual bearing fault diagnosis, it is verified that the method can accurately extract abnormal fault features under the adverse effects of multicomponent coupling and multiple accidental impact interference and can consider intelligence, high robustness, and high computational efficiency in applications.