A Priori Fusion Fully Interpretable Transfer Learning Method for Rotating Machinery Fault Diagnosis

• Novel A Priori Fusion Framework : The proposed method introduces an innovative a priori fusion mechanism, incorporating wavelet packet decomposition's prior knowledge into the feature extraction module of intelligent fault diagnosis. It adopts the Riemann entropy prior as a bandwidth selection criterion in the fault type classification module. The framework ensures that fault diagnosis methods can both utilize data-driven patterns and integrate expert knowledge, thereby enhancing diagnostic accuracy and achieving generalization capabilities across different operating conditions or mechanical types. • Fully Interpretable Architecture : Different from the existing fault diagnosis methods which are only local interpretable in feature extraction, the proposed method is interpretable in both feature extraction and fault discrimination. • Small Sample Transfer Learning Strategy: To address the challenges of limited fault diagnosis label data and low diagnostic accuracy, this study proposes a small sample transfer strategy. The strategy involves two transfer phases: First, transferring the ResNet18 model trained on large-scale 2D image datasets to the fault diagnosis model using laboratory data, significantly reducing training time. Second, transferring the model trained on laboratory data to actual fault data for fine-tuning, which not only improves diagnostic accuracy but also drastically reduces the reliance on labeled fault data. • Comprehensive Experimental Validation : The proposed method was validated using datasets from CWRU and bench test data, covering various operational conditions (inner race, outer race, rolling elements, and compound failures). Comparative results with state-of-the-art transfer learning methods demonstrated its diagnostic accuracy (average precision>98%). Furthermore, Grad-CAM and correlation coefficients were employed to demonstrate the method interpretability advantages in both qualitative and quantitative aspects, confirming its practical value in industrial fault diagnosis. This study addresses the problems of poor interpretability, difficulty obtaining the actual fault label data, and low diagnosis accuracy of the current intelligent fault diagnosis method for rotating machinery. In this study, a priori fusion fully interpretable transfer learning method is proposed. In the feature extraction stage, a fully interpretable multi-scale wavelet module is proposed, which is composed of a multi-scale wavelet fusion layer and an interpretable dynamic threshold activation layer. Multi-scale interpretable extraction of fault features is realized. In the fault decision stage, a frequency band attention mechanism is embedded in the residual layer of the ResNet18 framework, which enables decision-making to be interpretable. The experimental data and bench test results show that the fault diagnosis accuracy of the proposed method is 98.23%. In the experimental analysis of interpretability, the main frequency matching degree with tolerance is proposed as the index of the quantitative interpretability. The results show that the interpretability of the proposed method is nearly 2.5 times higher than that of the comparison method, and the noise resistance is obviously better than that of the comparison method.