The bearing multi-sensor fault diagnosis method based on a multi-branch parallel perception network and feature fusion strategy

Limited information from a single sensor constrains the precision of bearing fault diagnosis. Despite the abundance of multi-sensor data, the high dimensionality and complexity of data fusion make it difficult for existing methods to effectively extract and integrate multi-sensor features. To address these challenges, this paper proposes a novel multi-branch feature cross-fusion bearing fault diagnosis model (MCFormer), leveraging the powerful capabilities of Transformers in feature extraction and global modeling. First, to tackle the heterogeneity of multi-sensor data, a multi-branch structure is introduced to extract local features from each sensor separately, reducing information loss and redundancy. Then, based on the multi-branch feature extraction structure, a feature cross-fusion strategy and a dynamic classifier module are designed to achieve a unified representation of global features, enhancing feature discrimination and classification capabilities. Extensive experimental studies were conducted on two bearing cases, demonstrating that MCFormer achieves excellent diagnostic results on both the Northeast Forestry University (NEFU) bearing dataset and the Huazhong University of Science and Technology (HUST) bearing dataset, achieving diagnostic accuracies of 99.50 % and 98.33 %, respectively, surpassing the best performances of five other methods by 1.17 % and 2.36 %. Finally, ablation experiments confirm the efficacy of both component modules.