Dual-Path Mixed-Domain Residual Threshold Networks for Bearing Fault Diagnosis

• Published in: IEEE transactions on industrial electronics (1982) Vol. 69; no. 12; pp. 13462 - 13472
• Main Authors: Chen, Yongyi, Zhang, Dan, Zhang, Hui, Wang, Qing-Guo
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.12.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Asynchronous motors; Convolution; Convolutional neural networks; Deep learning; Dilated convolution; Domains; Fault diagnosis; Feature extraction; Interference; Machine learning; mixed-domain mechanism; Neural networks; Roller bearings; rolling bearings; soft threshold; Task analysis; Vibrations; White noise
• DOI: 10.1109/TIE.2022.3144572
• ISSN: 0278-0046

Intelligent bearing fault diagnosis based on deep learning is one of the hotspots in mechanical equipment monitoring applications. However, traditional deep learning-based methods have a weak antinoise ability and poor generalization performance in a noisy environment. This article presents a new simple and effective deep attention mechanism network, namely, dual-path mixed-domain residual threshold network (DP-MRTN), which aims to improve the accuracy of the rolling bearing fault diagnosis in a noisy environment. The DP-MRTN combines the channel attention mechanism, spatial attention mechanism, and residual structure. The soft threshold function is used as the nonlinear transformation layer, and the dilated convolution is introduced to establish a dual-path neural network so as to select the important features in the signal without resorting to any signal denoising algorithm. The performance of the DP-MRTN is validated against those state-of-the-art results on the real three-phase asynchronous motor experiment platform in Zhejiang University of Technology. We have achieved 99.97<inline-formula><tex-math notation="LaTeX">\%</tex-math></inline-formula> (<inline-formula><tex-math notation="LaTeX"> \pm 0.09\%</tex-math></inline-formula>) accuracy on Gaussian white noise, 99.87<inline-formula><tex-math notation="LaTeX">\%</tex-math></inline-formula> (<inline-formula><tex-math notation="LaTeX"> \pm 0.12\%</tex-math></inline-formula>) accuracy on Laplacian noise, and 99.98<inline-formula><tex-math notation="LaTeX">\%</tex-math></inline-formula> (<inline-formula><tex-math notation="LaTeX"> \pm 0.02\%</tex-math></inline-formula>) accuracy on real noise. The results show that the proposed method can significantly improve the accuracy of fault diagnosis in a noisy environment compared with the traditional deep learning method.