A Structural Re-Parameterization Network for Bearing Fault Diagnosis Under Variable Working Conditions

Abstract Deep convolutional networks are well known for their strong performance in feature characterization of massive data. Performing intelligent fault diagnosis (IFD) by training an end-to-end model could facilitate the acquisition of health state information of key machinery components while avoiding the need for human intervention and expert knowledge. However, for traditional machine learning, most methods are proposed and implemented based on the assumption that known data collected from the source domain and unknown data collected from the target domain share the same feature distribution, which can hardly be satisfied due to limitations in real industrial applications, as well as variable working conditions. To address this problem, a structural re-parameterization network is proposed, which is used to automatically extract deeper features and minimize the domain discrepancy between the source and target domain by using both metric learning and adversarial learning strategies. To be specific, an initial multi-branched network including shortcut structure is first established. This network enhances feature extraction by fusing features from multiple branches and aligning data when performing IFD under variable working conditions. After completing the training phase, the network can be re-parameterized during the inference stage, transitioning into a single-branched network while maintaining identical results, which is much more computationally efficient. Experiments have shown the effectiveness of the proposed method.