An interpretable waveform segmentation model with embedded fault transient information and its applications

Intelligent diagnosis technologies are essential for enhancing the intelligence level and reliability of industrial equipment. However, most existing approaches are constrained by their “black-box” nature, commonly exhibiting limitations such as the difficulty in intuitively aligning diagnosis evidence with measured signal waveforms and the disconnection between ante hoc and post hoc interpretability analyses. These shortcomings result in opaque decision-making processes and unclear physical meanings, which significantly limits their practical engineering applications. To address these challenges, this paper proposes an interpretable waveform segmentation-based diagnosis framework with embedded fault transient information. Exploiting the sparsity and cyclostationary characteristics of fault-induced impulsive features, a time-domain minimum entropy deconvolution module and a frequency-domain minimum l1/l2-norm deconvolution module are designed to achieve adaptive enhancement of weak transient fault information and effective noise suppression. Furthermore, a dual-branch X-shaped waveform segmentation network is constructed to fuse complementary time-domain and frequency-domain features, enabling structured embedding and explicit segmentation of fault transient information. To improve model robustness under complex operating conditions, a supervised class-level weighted Lovász-Softmax segmentation constraint and an unsupervised dual-branch consistency constraint are introduced, effectively alleviating learning bias caused by signal disturbances. The proposed method provides intuitive diagnosis evidence through structured segmentation results and supports the construction of fault fingerprint maps based on segmented waveforms, facilitating quantitative damage assessment and post hoc visual traceability of the decision-making process. Experimental results demonstrate that the proposed framework achieves superior accuracy, robustness, and interpretability in weak fault identification, quantitative damage evaluation, and adaptability to varying operating conditions, offering an effective and trustworthy solution for intelligent fault diagnosis.