动力学仿真数据驱动域自适应智能故障诊断

High-quality labeled data was a crucial prerequisite for the effectiveness of deep learning-based fault diagnosis methods. However, obtaining a substantial number of industrial labeled fault cases was challenging, which led the model's generalization ability weak. A novel domain adaptive intelligent diagnosis method driven by dynamics simulation data was proposed to address the above issue. This method considered the fundamental disparity between simulation data and actual data, and introduced a feature separation network for domain adaptation in diagnostic modeling. Based upon traditional domain adaptation models, a unique feature extractor that was specific to the target domain was incorporated to explicitly separate environmental noises present in actual data. This enhancement improved fault feature representation and clustering capabilities through other features that remain invariant across domains. Furthermore, a novel training strategy was proposed that leveraged diagnostic results obtained from the shared feature extractor to itcrativcly update the model parameters of the unique feature extractor, thereby enhancing training stability even further. The proposed method was evaluated using the bearing dataset from Case Western Reserve University, demonstrating improved feature extraction and clustering capabilities compared to other transfer methods for comparison, as evidenced by enhanced performance and diagnostic accuracy. Additionally, the hyper-parameter sensitivity was analyzed empirically.