Research on fault diagnosis of synchronous generators based on dual path SeNet and improved ViT

Abstract To tackle the challenges in fault diagnosis of synchronous generators, which often rely heavily on expert knowledge and result in low diagnostic accuracy, and to address the overfitting issues associated with Vision Transformer (ViT)-based models due to their large parameter size when applied to limited data samples, this paper introduces an innovative fault diagnosis approach that integrates Dual-Path SeNet with an enhanced ViT architecture. Initially, continuous wavelet transform (CWT) is applied to convert raw one-dimensional electrical signals into two-dimensional time-frequency images, effectively capturing both temporal and frequency characteristics. These images are then processed by the Dual-Path SeNet for deep and multi-scale feature extraction. Subsequently, depthwise separable convolutional neural networks are used to transform the extracted feature maps into long vectors, significantly reducing the parameter size associated with patch processing and linear mapping operations in the original ViT. These vectors are then fed into a multi-head attention mechanism to extract temporal features. To further reduce the parameter size, sequence pooling is employed to replace the classification head in the original ViT, enhancing the model’s lightweight design. Experimental results on three fault types demonstrate that the proposed method reduces parameter size by over 50% and achieves a diagnostic accuracy of 99.07%, highlighting its effectiveness in synchronous generator fault diagnosis.