Rolling bearing fault diagnosis based on multi-source domain adaptive residual network

Here, aiming at the problem of weaker applicability of traditional unsupervised domain adaptation methods in multi-working condition rolling bearing fault diagnosis scenarios, a multi-source domain adaptative residual network (MDARN) was proposed. By aligning relevant subdomains from multiple source domains, MDARN could improve its fault diagnosis performance under multiple working conditions. Firstly, ResNeXt residual network was used to fully extract transferable features from source domain and target domain. Then, the local maximum mean difference (LMMD) criterion was introduced to align relevant subdomains in target domain based on subdomains of two source domains to reduce distribution differences among relevant subdomains and global domain. Finally, the proposed method was experimentally verified using the bearing dataset of Case Western Reserve University, US and the actual bearing vibration dataset generated by MFS mechanical comprehensive fault test bench. The results showed that the average fault diagnosis accuracy of this method under multi-working condition reaches 99. 76%; compared with existing representative methods, the proposed method has better fault diagnosis effect. © 2024 Chinese Vibration Engineering Society. All rights reserved.