Fault Diagnosis of Permanent Magnet Synchronous Motors Based on FedProx Federated Learning

To address the generalization degradation of permanent magnet synchronous motor (PMSM) inter-turn short-circuit fault diagnosis under multiple operating conditions, as well as the performance drop in federated training caused by decentralized data and non-IID distributions, this paper proposes a multimodal federated fault diagnosis method. Using current and vibration signals as inputs, a dual-branch 1 D residual encoder is employed to extract modality features, and a gated mechanism is used for adaptive fusion. In addition, FedProx is introduced to mitigate client drift, and a validation-performance-based aggregation weighting scheme is adopted to improve robustness under non-IID settings. Experiments are conducted on data with multiple speeds, grounding resistance levels, and no-load/load conditions for three-task recognition. Results show that, compared with FedAvg with equal-weight fusion, the proposed method improves grounding-resistance level and load-state recognition by 12.41 and 6.94 percentage points, respectively, while reducing the overall loss; compared with FedProx with equal-weight fusion, it achieves better overall performance, validating the effectiveness of gated fusion and indicating that aggregation-weight design has a significant impact on performance.