A trustworthy lightweight multi-expert wavelet transformer for rotating machinery fault diagnosis

Rotating machinery plays a crucial role in industrial applications, making reliable fault diagnosis essential for operational efficiency and system safety. Current deep learning approaches face challenges in trustworthiness, computational efficiency, and limitations in training with small datasets. To address these issues, this paper proposes a trustworthy lightweight multi-expert wavelet transformer (TLMW-former) for fault diagnosis. The TLMW-former incorporates a multi-wavelet sparse representation denoising (MSRD) layer to effectively suppress background noise, enhancing fault features of signals. A wavelet linear self-attention (WLSA) mechanism is designed to improve global feature mining and model interpretability, while a stride inverted residual module is introduced to enhance local feature extraction and downsampling. Additionally, a multi-expert feedback layer is developed to strengthen decision-making reliability and performance through collaborative expert mechanisms. Experimental evaluations on four datasets demonstrate the superiority of TLMW-former, achieving 96.86%, 88.32%, 92.16%, and 90.88% accuracy, respectively, while requiring significantly fewer computational resources than most baseline models. The results highlight the performance and trustworthiness of the model and suitability for deployment in complex industrial environments.