An evolving fuzzy classification technology for bearing fault diagnosis

Abstract

Reliable fault detection and diagnosis in rolling element bearings is essential for ensuring the safety and stability of rotating machinery, as bearings are commonly used in rotating machines, and bearing failures account for almost 50% of all machine malfunctions. However, reliable bearing fault detection and diagnosis remain a challenging task as bearing signals are typically non-stationary and strongly modulated by other vibration signals especially in real-world industrial applications. The objective of this work is to develop an evolving monitoring system for bearing fault diagnosis. It consists of two main stages: signal processing for representative feature extraction and pattern classification for bearing fault diagnosis. In signal processing, an adaptive Chirp-mode Decomposition-based Squared Envelope (CDSE) technique is proposed for bearing fault detection. The CDSE takes two processing operations: 1) The adaptive chirp mode decomposition (ACMD) is employed to extract the most informative amplitude-frequency modulated components from vibration signals. 2) The squared-envelope analysis is performed to extract periodic impact features to predict bearing faults. In diagnostic pattern classification, an evolving fuzzy belief (EFB) technique is developed to integrate the merits of several fault detection techniques, including the CDSE, to provide an interpretable assessment of bearing health conditions. The EFB integrates feature-weighted Manhattan distance, prototype-based learning, and belief-rule inference for pattern classification. Feature weights are adaptively determined through Pearson-correlation analysis to highlight informative spectral and statistical features while reducing the influence of irrelevant dimensions. The rule base is continually adapted to each data chunk so as to provide transparent and interpretable diagnostic decision-making. The effectiveness of the proposed CDSE technique and EFB classifier is validated by simulation and experimental tests. Test results show that the CDSE technique demonstrates improved robustness under background noise and redundant modes reduction, leading to more stable and reliable characteristic frequency extraction. It outperforms other related methods across different bearing datasets and fault conditions. On the other hand, the EFB classifier can achieve higher or comparable classification accuracy relative to other related evolving fuzzy systems.