Angular resampling-assisted multi-stage parameter transfer learning method for fault diagnosis from stable to time-varying operating conditions

Conventional deep learning-based fault diagnosis methods perform impressively when dealing with data from uniform distributions. However, achieving accurate fault diagnosis across diverse operating conditions remains a significant challenge. Although fine-tuning-based transfer learning methods have achieved some success in cross-condition diagnosis, their widespread adoption is impeded by limitations such as limited feature extraction capabilities, inefficient training strategies, and high computational resource requirements. To address the issues above, an innovative transfer diagnosis method is proposed in this paper, which is angular resampling-assisted multi-stage parameter transfer learning (AR-MSPTL) method, designed specifically for transfer diagnosis from stable to time-varying operating conditions. The proposed AR-MSPTL begins with an angular resampling-based order spectrum analysis method that transforms non-stationary time-domain vibration signals into quasi-stationary angular-domain signals, thereby eliminating the impact of speed variations and enhancing fault features. Then, a novel adaptive parametric self-regularizing activation function, APMish, is proposed to improve the feature extractor by capturing the generalized features. Subsequently, a multi-stage parameter transfer learning (MSPTL) strategy is proposed to facilitate deep generalization feature learning. This new training strategy leverages multi-stage trained parameters to bolster the generalization capabilities and boost the transfer diagnostic performance of the model. Finally, extensive experimental evaluations on challenging transmission system datasets validate the efficacy of our AR-MSPTL method, demonstrating that it effectively transfers prior fault knowledge across varying operating conditions while achieving exceptional diagnostic accuracy that significantly advances mainstream transfer learning diagnosis methods.