Cross-domain fault diagnosis for bearings via multi-branch convolutional neural network and mixed discrepancy measurement

To address the challenges of cross-domain fault diagnosis for industrial rotating machinery under strong noise interference and time-varying operating conditions, this paper proposes a novel method based on Multi-Branch Convolutional Neural Network (MBCNN) and a mixed discrepancy measurement strategy. The proposed MBCNN model is designed as a feature extractor, integrating parallel branches for moving average smoothing, Gaussian filtering, and original vibration signals, which effectively enhances time-frequency feature representation while suppressing noise interference. Furthermore, to improve feature distribution alignment between source and target domains, a Mean-Variance Mixed Discrepancy (MVMD) based domain adaptation module is introduced. This module jointly leverages Maximum Mean Square Discrepancy (MMSD) and Variance Discrepancy Representation (VDR) to perform multi-granularity statistical alignment. Extensive experiments on the CWRU bearing dataset, including noise robustness and cross-condition transfer tasks, Showcase the robust noise immunity and multi-domain diagnostic prowess of the proposed approach, offering a practical and dependable approach for intelligent fault diagnosis in complex industrial environments.