Interpretable Knowledge Transfer with a Novel Structure Discrepancy Metrics for Unsupervised Bearing Fault Diagnosis from Simulation to Reality

Abstract This paper introduces a novel interpretable transfer learning framework, based on Wavelet Temporal Logic Network (WTLN) encoder, which leverages temporal logic for interpretable knowledge transfer in rolling element bearings fault diagnosis. Unlike traditional approaches, WTLN integrates Weighted Signal Temporal Logic (wSTL) with dynamic learning capabilities to robustly encode and interpret signal behaviors under varying conditions. The model uniquely combines a Temporal Logic Discrepancy Metric (TLDM) for enhanced loss function definition to address the challenges of transfer learning and knowledge structure alignment in fault diagnosis. Extensive experiments conducted with both simulated and real-world datasets show that the model is highly effective in diagnosing bearing faults under various operational conditions. By combining the WTLN with TLDM, the model achieves not only high diagnostic accuracy but also provides interpretability. This interpretability is vital for real-world applications, where it is crucial to understand the decision-making process.