An adaptive feature extraction technique via bispectrum-driving graph domain for bearing fault diagnosis.

Abstract

The rapid development of the mechanical industry has made fault feature extraction based on vibration signals increasingly challenging. The emergence of graph signal processing theory offers a new perspective for signal feature extraction. This study aims to provide an adaptive feature extraction technique via bispectrum-driven graph domain for bearing fault diagnosis. Initially, the bispectrum is established as the core of graph signal construction to pre-demodulate the modulation components in the signal, thereby enhancing the accuracy and interpretability of graph signal processing. Subsequently, an optimal node identification technique is developed to find effective components in the graph signal eigenvalues. Finally, the use of optimal eigenvalues to extract feature information from graph signals is demonstrated through rigorous mathematical derivation, achieving accurate and reliable fault diagnosis. The experimental work presented here illustrates the practical effectiveness of the method for bearing fault signals.