A vision transformer based multi-sensor fusion approach for fault diagnosis of key bogie components

Fault diagnosis of bogie core components under complex operating conditions remains a challenging task due to the large number of fault categories and the strong coupling among multi-source sensor signals. To address this issue, this paper proposes a large-scale fault diagnosis method for bogie systems based on multi-sensor information fusion and Vision Transformer (ViT). First, continuous wavelet transform (CWT) is employed to convert multi-sensor vibration signals into time–frequency representations. These representations are then segmented into patches and jointly encoded as input tokens for the ViT model, enabling effective feature extraction and global dependency modeling across multiple sensors.The proposed method is validated on the Beijing Jiaotong University Railway Bogie Dataset (BJTU-RAO), involving a challenging 51-class fault diagnosis task under multiple operating conditions. Experimental results demonstrate that the ViT-based model achieves fast convergence, high diagnostic accuracy, and stable performance, with an average accuracy above 95%. Comparative studies further indicate that single-sensor-based diagnosis is insufficient for large-scale fault classification, yielding significantly lower accuracy and poor stability. In addition, the proposed approach achieves better overall diagnostic accuracy and robustness than multi-channel CNN-based fusion models. A comprehensive evaluation in terms of model size, parameter count, and computational complexity confirms that the ViT model exhibits more favorable lightweight characteristics while maintaining superior diagnostic performance. These results demonstrate the effectiveness and practicality of the proposed method for large-scale bogie fault diagnosis and potential deployment in real-world railway systems.