In real-world industrial environments, achieving interpretability in cross-domain fault diagnosis is essential for ensuring transparency in the decision-making process. Traditional deep learning-based approaches, however, lack sufficient exploration of interpretable invariant feature learning for cross-domain fault diagnosis. To address this gap, we propose a novel method, Learnable TQWT-Enhanced Unrolling Net (LTQU-Net). Specifically, the network first employs a Learnable TQWT Subband Alignment and Fusion module to extract physically meaningful time–frequency features, enabling the sparse representations obtained by TQWT to be trainable in an end-to-end manner. Furthermore, by adopting algorithm unrolling, the sparse coding process is unfolded into a deep network, where the inherent interpretability of iterative algorithms is transferred to the dictionary learning process. In addition, we design a set of task-specific loss functions to further enhance the performance of cross-domain fault diagnosis. Experimental results on publicly available datasets demonstrate that LTQU-Net achieves significantly higher accuracy than existing methods while ensuring interpretability. This work provides a transparent and effective solution for intelligent fault diagnosis under cross-domain conditions.

algorithm unrolling,TQWT,interpretable,cross-domain fault diagnosis