In real-world industrial scenarios, due to variations in working conditions, the vibration data gathered from mechanical equipment presents different probability distributions. Moreover, the gathered vibration data is usually unlabeled. Therefore, the unlabeled vibration data may be misclassified by the trained diagnostic source models. Deep transfer learning with the adaptation of marginal or conditional distribution can extend the knowledge of the trained models in source domains to a novel but it is related to a target domains. However, the only reduction in marginal distribution discrepancy limits the generalization ability and performance of the diagnostic model. Besides, the most effective diagnostic model that matches both marginal and conditional distributions fail to take full advantage of complex and sensitive features in the data set. To address these issues, a multiscale subdomain adaptive model (MSDAM) for rolling bearings fault diagnosis under cross-working conditions is proposed. Firstly, in order to eliminate signal pre-processing and manual feature extraction dependencies, the original vibration signal is used as the input of the proposed MSDAM. Secondly, a multiscale convolutional neural network is built to capture multiscale common features with fine-grained information and transfer the features of labeled source data and unlabeled target data to the same subspace. Then, the prediction results of the source classifier are used as the soft labels for the unlabeled target data, and the relevant subdomains are divided according to different fault types. Finally,the local maximum mean distance (LMMD) method is employed to align the conditional distributions of the subdomains in the common feature space to effectively reduce the distribution discrepancy of similar fault characteristics under different working conditions. The location selection experiment of the subdomain adaptation layer shows that the subdomain adaptation layer is set before the classification layer, which can better guide the optimization of the entire network and maintain feature distribution matching. The results on the Paderborn bearing dataset show that the average diagnostic accuracy of the multiscale model is improved by 7.98% compared to the single-scale model. Despite sacrificing extra training time, the proposed MSDAM has stronger transfer ability. To further verify the effectiveness of the proposed MSDAM, a total of 12 transfer learning tasks are carried out on the CWRU bearing dataset, Paderborn bearing dataset, and laboratory-built datasets respectively. The experimental results indicate that the proposed MSDAM has higher accuracy in fault diagnosis across working conditions than the deep adaptation methods such as domain adaptive networks(DAN) and domain-adversarial neural networks (DANN). In addition, t-SNE visualization of the subdomain adaptation layer representations indicates that the proposed MSDAM can accurately align the relevant subdomain distributions. The following conclusions can be drawn through the analysis of experimental results: (1) The proposed MSDAM can extract more fine-grained common fault features and align relevant subdomains based on LMMD compared with DAN and DANN. (2) The proposed MSDAM only requires the original vibration signal as input, eliminating the need for signal preprocessing and manual feature extraction. (3) A high cross domain diagnostic accuracy on all three datasets proves the strong generalization ability of the proposed MSDAM. © 2024 Chinese Machine Press. All rights reserved.
