Fault Feature of Gearbox Vibration Signals Based on Morphological Filter Dynamic Convolution Autoencoder

Abstract

Vibration signals are widely used in fault diagnosis of rotating machinery. Although deep neural networks (DNNs) have been widely used in feature learning of vibration signals, they have not shown good performance in the extraction of impulsive features and noise filtering. In this study, a new DNN, morphological filter dynamic convolutional autoencoder (MF-DCAE), is proposed for fault feature extraction from vibration signals. First, a multiscale morphological filter layer embedded with a convolutional autoencoder (CAE) is proposed to extract fault features of vibration signals. Kurtosis is used to fuse the feature information extracted by morphological operators with different scales. Dynamic convolution is used to adjust the weight of each convolution kernel for adaptive feature extraction. Finally, residual learning is used to connect the encoder and decoder to obtain good feature learning of vibration signal. The effectiveness of MF-DCAE is verified on two gearbox testbeds. The results show that MF-DCAE can perform both noise reduction and feature learning on vibration signals in an unsupervised learning manner. The comparison results show that the feature extraction performance of MF-DCAE is better than that of the state-of-the-art DNNs.