Attention specialized multi-view vision transformer for physics interpretable rolling element bearing fault diagnosis

Rolling element bearing fault diagnosis remains challenging because different fault conditions produce complex and overlapping vibration characteristics. This study proposes a physics-guided Multi-View Vision Transformer (MV-ViT) that converts raw vibration signals into six complementary image representations: Short-Time Fourier Transform (STFT), Continuous Wavelet Transform (CWT), Constant-Q Transform (CQT), Spectral Kurtosis (SK), Teager Energy Map (TEM), and Envelope Spectrum (ES). The six representations are jointly processed using a shared Vision Transformer with multi-head self-attention to learn cross-view relationships among time-frequency, spectral, impulsive, and nonlinear energy characteristics. The proposed framework is evaluated for five-class bearing fault diagnosis, including combined faults. MV-ViT achieves a maximum test accuracy of 99.62%, with a mean accuracy of 98.72 ± 0.70% across five random seeds, demonstrating consistent performance across independent training runs. The class-wise results further show F1-scores above 0.96 for the combined-fault class. Transform-level attention analysis identifies the Envelope Spectrum as the most influential representation, accounting for approximately 42.8% of the measured attention energy, followed by CQT at 20.3%. This observation is verified using leave-one-view ablation, where removing the Envelope Spectrum leads to a decrease in the test accuracy by 1.12% points, going from 99.62% to 98.50%, while removing the CQT, CWT, or TEM view results in a decrease in accuracy of 0.75% points. The ablation study on the architecture of the network also shows that decreasing the depth and embedding dimension of the Transformer does not cause noticeable changes in the accuracy, which implies that representation diversity plays a more important role in diagnosis than increasing the model capacity. In Leave-One-Working-Condition-Out experiments, the accuracy for unseen working conditions is achieved at levels of 97.92%, 98.01%, and 97.23%, respectively. The statistical comparison of the results of five repeated experiments shows no significant differences between the validation and test accuracy ( p = 0.3125). Finally, post-training quantization leads to a decrease in the size of the model from 77.9 MB in FP32 to 19.7 MB in INT8 and 9.9 MB in FP16Q4, preserving high classification accuracy.