Transformer network enhanced by dual convolutional neural network and cross-attention for wheelset bearing fault diagnosis

Advances in deep learning methods have demonstrated remarkable progress in wheelset fault diagnosis. However, current deep neural networks suffer from design flaws, including low accuracy, high computational complexity, limitations in frequency-domain analysis, and inefficient long time-series feature encoding. To address these challenges, this study proposes a Transformer network model based on dual convolutional neural networks and cross-attention enhancement (Trans-DCC) for wheelset bearing fault diagnosis. The model incorporates a dual feature fusion mechanism in the first layer of the Transformer encoder, utilizing dual CNNs to extract low-level time-frequency features while reducing subsequent attention computation complexity. Additionally, a cross-attention mechanism is integrated into the last encoder layer, combining multi-head attention with time- and frequency-domain features from a feedforward connection layer. Attention weights are computed to prioritize critical features before enhancement fusion. Finally, fully connected layers and a softmax classifier are employed for fault classification. Experimental evaluation on a train wheelset bearing dataset confirms the model’s effectiveness, demonstrating high diagnostic accuracy. The proposed Trans-DCC model overcomes key limitations of existing methods by enhancing feature extraction and fusion, offering a robust solution for wheelset bearing fault diagnosis.