Eccentricity fault diagnosis of permanent magnet synchronous generators based on 2D recursive fusion graph and CBAM-ConvNeXt-FPN

To detect the eccentricity fault of permanent magnet synchronous generator (PMSG) and ensure the stable operation of the equipment, a new method based on multi-channel signal 2D recursive fusion combined with Convolutional Block Attention Module (CBAM)-ConvNeXt-FPN is proposed to detect complex eccentricity faults of PMSG in this work. First, the 1D three-phase current signals are mapped into 2D images and fused across multiple channels to enhance the richness and complementarity of fault feature information. Hereafter, a novel multi-scale fusion classification deep learning framework, named CBAM-ConvNeXt-FPN, is proposed to capture both local and global information from feature maps, conduct deep fault feature extraction, and improve feature representation capability, achieving precise classification of complex eccentric faults across fault types. Experimental results from the test platform demonstrate that the diagnostic accuracy of the CBAM-ConvNeXt-FPN model reaches 99.08%, outperforming comparison models such as Swin Transformer. The robustness of the model was verified through noise experiments and load fluctuation tests. The results show that under noisy environments or load fluctuation conditions, the accuracy of this model remains within a stable range and is higher than that of other models. These results confirm the superiority and effectiveness of the proposed approach.