Wavelet-patch embedding and two-stage attention fusion for multi-sensor mechanical fault diagnosis

Transformer-based fault diagnosis models have shown promising performance in rotating machinery monitoring owing to their ability to capture long-range temporal dependencies. However, most existing methods rely on purely data-driven patch embedding strategies that lack physical interpretability and insufficiently exploit the heterogeneous sensitivities of multiple sensors. To address these limitations, an interpretable multi-sensor fault diagnosis framework, termed LCWC-DAMiT, is proposed. The framework introduces a learnable continuous wavelet convolution (LCWC) module to incorporate time–frequency priors into patch embedding. By initializing convolution kernels with Morlet wavelets and optimizing their parameters during training, LCWC enables adaptive extraction of fault-related transient features while preserving physical interpretability. In addition, a prototype-guided dual attention mechanism is developed to model sensor-level importance and multi-scale temporal dependencies, facilitating effective fusion of complementary information from multiple sensors. Experiments are conducted on a gearbox dataset under varying operating conditions and a centrifugal pump dataset collected from 26 different machines. The results demonstrate that the proposed method consistently outperforms representative convolutional neural network, and Transformer-based approaches in terms of diagnostic accuracy and generalization capability. Furthermore, analyses of learned wavelet kernels, attention distributions, and feature perturbations reveal that the model focuses on physically meaningful fault characteristic frequencies and informative sensor locations. These results demonstrate the effectiveness of the proposed framework on the evaluated gearbox and centrifugal pump systems and suggest its potential applicability to a broader range of rotating machinery.