What question did this study set out to answer?

The aim is to enhance rolling bearing fault diagnosis accuracy while providing interpretable results linked to physical excitation structures.

April 17, 2026VibrationOpen Access

Multi-Wavelet Fusion Transformer with Token-to-Spectrum Traceback for Physically Interpretable Bearing Fault Diagnosis

View Full Paper

Ask AI

Bookmark

Authors

Demonstrates accurate and interpretable bearing fault diagnosis using a multi-wavelet approach, suggesting a robust framework for reliability.

The aim is to enhance rolling bearing fault diagnosis accuracy while providing interpretable results linked to physical excitation structures.
Developed a multi-wavelet TF fusion network named MBT-XAI.
Implemented a Token-to-Spectrum Traceback mechanism for interpretability.
Utilized three wavelets — Morlet, Mexican Hat, and Complex Morlet — to create multi-view TF representations.
Adopted cross-channel attention within a Transformer backbone to fuse TF representations.
Achieved 98.13% accuracy on CWRU and 96.23% on IMUST dataset at SNR = 0 dB, outperforming the strongest baseline by approximately 2.5%.
Maintained accuracy of 95.44% and 93.45% under AWGN at SNR = -2 and -4 dB for CWRU.
Demonstrated high alignment between frequency regions and theoretical fault-characteristic bands with IoU = 80.21%.

Fan et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark

April 17, 2026VibrationOpen Access

View Full Paper

Ask AI

Bookmark

Authors

Demonstrates accurate and interpretable bearing fault diagnosis using a multi-wavelet approach, suggesting a robust framework for reliability.

The aim is to enhance rolling bearing fault diagnosis accuracy while providing interpretable results linked to physical excitation structures.
Developed a multi-wavelet TF fusion network named MBT-XAI.
Implemented a Token-to-Spectrum Traceback mechanism for interpretability.
Utilized three wavelets — Morlet, Mexican Hat, and Complex Morlet — to create multi-view TF representations.
Adopted cross-channel attention within a Transformer backbone to fuse TF representations.
Achieved 98.13% accuracy on CWRU and 96.23% on IMUST dataset at SNR = 0 dB, outperforming the strongest baseline by approximately 2.5%.
Maintained accuracy of 95.44% and 93.45% under AWGN at SNR = -2 and -4 dB for CWRU.
Demonstrated high alignment between frequency regions and theoretical fault-characteristic bands with IoU = 80.21%.

Fan et al. (2026) studied this question.

View Full Paper

Ask AI

Bookmark