Interpretable fault diagnosis for multi-sensor systems via a lifting wavelet Siamese transfer network with spatial-channel synergistic attention

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Single-sensor signals often provide insufficient information for bearing fault diagnosis, while conventional deep models may overfit when target-domain labels are limited and suffer from weak interpretability. To address these challenges, this paper proposes an interpretable multi-sensor diagnosis framework, termed lifting wavelet transform network-spatial-channel synergistic attention (ILWTN-SCSA)-Multi, which integrates Siamese transfer learning, synergistic attention, and wavelet-based interpretable decomposition into a unified architecture. First, one-dimensional vibration signals are converted into time–frequency representations using the short-time Fourier transform, and a VGG19 backbone pre-trained on ImageNet is adopted to provide robust initialization. Subsequently, a weight-sharing Siamese architecture is employed to enforce consistent feature encoding across multiple sensor channels within a shared representation space. To further enhance cross-sensor feature learning, a SCSA module is introduced to capture spatial semantic priors and adaptively recalibrate channel responses. In addition, an interpretable ILWTN is incorporated to perform adaptive multi-scale decomposition through lifting-based operators, together with high-frequency sparsity regularization and low-frequency mean-consistency constraints, thereby improving both fault-sensitive representation learning and physical interpretability. Extensive experiments on the public HUST dataset and a self-constructed five-sensor dataset demonstrate that the proposed method achieves average diagnostic accuracies of 99.70% and 100%, respectively, with standard deviations of 0.31% and 0.00%. The model contains only approximately 2.14 million parameters and outperforms several representative deep learning baselines in terms of accuracy, robustness, and parameter efficiency. These results indicate that ILWTN-SCSA-Multi provides an effective and practically meaningful solution for bearing fault diagnosis under multi-sensor and compound-fault conditions.