Machines’ Intelligent Fault Diagnosis Based on Hierarchical Refined Composite Generalized Multiscale Fluctuation Dispersion Entropy

Abstract

Vibration data from mechanical equipment contain extensive information distributed across multiple dimensions. Single-scale analysis fails to comprehensively reflect its damage characteristics, thereby reducing fault diagnosis accuracy. This study proposes a novel signal vibration feature extraction method called hierarchical refined composite generalized multiscale fluctuation dispersion entropy (HRCGMFDE). This method simultaneously extracts fault information from both low-frequency and high-frequency components of the data, addressing the drawback of high-frequency information loss in refined composite generalized multiscale fluctuation dispersion entropy (RCMFDE). Comparative results on two simulated signals demonstrate the method’s advantages of high stability and more accurate complexity measurement. Furthermore, low-frequency and high-frequency components of the data are comprehensively extracted using dual-tree complex wavelet packet transform (DTCWPT), and high-dimensional features are downscaled using t-distributed stochastic neighbor embedding (t-SNE) to obtain low-dimensional sensitive fault features. Subsequently, a Random Forest (RF) classifier is employed for fault identification. Finally, the effectiveness of the proposed method is validated using three typical mechanical datasets. Results confirm the method’s capability to effectively determine the fault states of bearings, gearboxes, and centrifugal pumps, showcasing significant advantages over comparative methods.