Gearbox fault diagnosis using a dual-channel fuzzy convolutional neural network with explainable analysis

Fault diagnosis is crucial for preventing serious failures in rotating machinery. However, existing methods are often limited by insufficient feature extraction quality and the “black-box” nature of models, making it difficult to balance accuracy and interpretability. This paper innovatively proposes an intelligent diagnostic framework integrating interpretability analysis, which combines the Layer-wise Relevance Propagation (LRP) algorithm with the dual-channel fuzzy convolutional neural network (TFCNN) to construct a fault diagnosis system that balances accuracy and transparency. Specifically, the fuzzy layer of TFCNN enhances the ability to model noisy and uncertain features through the design of Gaussian membership functions and fuzzy rule layers, thereby improving feature robustness. The framework realises rapid identification of fault modes through the synergistic optimisation between the time-frequency image enhancement technology based on the Improved Type-II Fuzzy Set (ITFE-II) and fuzzy deep learning; the LRP algorithm quantitatively analyzes the decision-making basis of the model, overcoming the limitations of traditional black-box models. Verified by self-built experimental platforms and public datasets, and compared with various advanced methods, the results show that the proposed method has excellent overall performance, providing a highly reliable and interpretable solution for the intelligent maintenance of wind power equipment.