The DRTS-Net Framework: Enhanced Probabilistic Health Perception for Wind Turbine Bearings Under Complex Operating Conditions With Improved Interpretability

Health perception in mechanical systems is a crucial aspect of prognostics and health management (PHM). Deep learning models have been successfully applied to health perception in wind turbine bearings. However, due to the time-varying, redundant, and noisy characteristics of time-series data representing bearing wear, existing models face significant challenges, particularly in terms of generalization and interpretability. Moreover, although most current models have provided health perceptions, they overlook the strong stochastic nature of bearing degradation data, and thus fail to deliver reliable predictions of uncertainty information. To address these technical challenges, this study introduces a novel Deep residual temporal-spatial framework (DRTS-Net) for probabilistic health perception of wind turbine bearing wear states. This article creatively designs spatiotemporal information fusion module for extracting features from bearing degradation data. Inspired by improvements in residual network structures, the DRTS-Net integrates an attention mechanism and a soft thresholding mechanism from the deep residual shrinkage network to optimize the feature extraction process. This integration aims to mitigate the interference from minor redundant features in the degradation data. The framework incorporates Gaussian process regression in the output layer to yield the probability distribution of single-health perception values for bearing features. A comparative study was conducted using the same bearing health state dataset against current mainstream deep learning methods for bearing health assessment. Results demonstrate that the proposed DRTS-Net achieves superior performance in high-precision health perception, high-reliability probabilistic health perception.