An intelligent diagnosis method for PMSM radial misalignment based on current signal and Swin-BiGRU multimodal fusion

Motor Current Signature Analysis (MCSA) faces considerable challenges in diagnosing mechanical faults in motors, particularly in accurately detecting misalignment in rotating machinery. Traditional vibration-based methods often suffer from high hardware costs and difficulties associated with sensor installation. To address these limitations, this study proposes an intelligent diagnosis method for radial misalignment in Permanent Magnet Synchronous Motors (PMSMs) based on Swin-BiGRU multimodal fusion using current signals. First, Adaptive Variational Mode Decomposition (AVMD) is applied to the collected three-phase current signals to eliminate high-frequency noise. Considering the inherent difficulty of fault feature extraction in current signals, Markov Transition Fields (MTF) are used to transform one-dimensional time-series current data into time-frequency images, thereby highlighting subtle or weak fault signatures. The proposed framework employs a dual-branch network: one branch utilizes the Swin Transformer to extract deep features from MTF images, while the other adopts a Bidirectional Gated Recurrent Unit (BiGRU) with Global Attention (GATT) to model the original current time-series. After feature extraction, a bidirectional Cross-Attention mechanism is introduced to enable efficient interaction and enhancement between the multimodal features, improving both diagnostic accuracy and robustness. To validate the proposed method, ablation studies and comparative experiments were conducted before and after denoising. Experimental results demonstrate that under radial misalignment conditions ranging from 0.5 mm to 1.5 mm, the proposed method achieves an average diagnostic accuracy of 99.375%. Even without the AVMD denoising step, the method maintains a high accuracy of 98.125%, outperforming other benchmark methods. In industrial settings, this method can be integrated into automated equipment.