Rolling bearing fault diagnosis based on multi-scale attention and improved joint distribution adaptation network

To address the issues of distribution differences caused by complex working conditions in industrial scenarios and the scarcity of labeled data, a novel multi-scale attention and improved joint distribution adaptation network (MA-IJDA) is developed for rolling bearing fault diagnosis. Firstly, a multi-branch convolutional neural network (MACNN) integrated with a channel attention mechanism is constructed as a shared feature extractor. This network sufficiently captures multi-scale transferable fault features through differentiated receptive fields and adaptively enhances key information. Secondly, by fusing softmax confidence and feature space clustering results, a dual-path pseudo-label generation strategy is developed to boost target domain pseudo-labeling reliability. On this basis, an improved joint distribution adaptation (IJDA) mechanism is developed, which employs a joint mean–variance discrepancy (JMVD) metric to synchronously align marginal and conditional distributions, thereby enhancing inter-class separability and cross-domain discriminative capability in the feature space. Extensive cross-working condition experiments on CWRU and JNU bearing datasets verify that MA-IJDA achieves superior diagnostic performance across varying loads and rotational speeds, confirming its exceptional transferability and generalization in complex industrial scenarios.