Physics-informed dual guidance method using physical envelope harmonic distribution and transfer learning for few-shot gear fault classification

Abstract

Recent years have seen artificial intelligence algorithms gain considerable popularity in gear fault classification, yet their performance remains hindered by the scarcity of labeled fault data, often leading to suboptimal classification results. Several state-of-the-art studies have demonstrated that incorporating physical information can improve classification accuracy. However, the differences between simulated and measured signals pose a significant challenge in enhancing the performance of physics-informed methods. In order to fill this gap, this paper introduces a novel physics-informed dual guidance (PI-DG) method using physical envelope harmonic distribution (PEHD) and transfer learning (TL) for few-shot gear fault classification. Within the proposed method, we introduce a new concept of PEHD, which is defined as the distribution feature of the shaft frequency and its harmonics in envelope spectrum. A physics-informed parameter optimization model (PI-POM) is developed to minimize the difference between the simulation and measured signals in terms of PEHD, enabling the accurate identification of detailed parameters within the dynamic model. Subsequently, a TL guidance framework is established for the fine-tuning and adaptation of a Long Short-Term Memory-aided Different Kolmogorov-Arnold network (LSTM-DKAN), with the aim of improving classification accuracy. Validation on a constructed back-to-back gear test rig with induced crack and spalling faults demonstrates the PI-DG method's effectiveness in reducing physics-simulation discrepancies, exhibiting superior classification performance especially in few-shot cases.