DSTF-Net: A novel framework for intelligent diagnosis of insulated bearings in wind turbines with multi-source data and its interpretability

Deep Learning has gained widespread attention in the field of mechanical equipment fault diagnosis. However, the structural complexity of insulated bearings in high-power variable-frequency wind turbines, coupled with random load variations and time-varying operating conditions, introduces nonlinear signals with weakly coupled features that are easily disrupted by noise. Relying on single-source data is insufficient for extracting comprehensive fault characteristics. Additionally, insulated bearings are subjected to long-term metaphysical coupling effects, including thermal, electrical, fluid, and mechanical interactions, resulting in highly complex fault information. Current models struggle to accurately extract the temporal features of these faults. Moreover, the internal operational mechanisms and potential behaviors of insulated bearing fault recognition remain poorly understood, limiting the ability to explain the causal relationships behind electrical pitting fault diagnosis results. To address these challenges, this paper proposes a novel framework, DSTF-Net, aimed at dynamic spatial-temporal feature fusion using multi-source data. The framework is specifically designed to address the real-world engineering problem of identifying insulation bearing faults in high-power inverter wind turbines, and significant progress has been made in fault diagnosis. Specifically, the key contributions of this study include the following. Firstly, a new temporal information feature fusion unit, which is Weight Reduction Recurrent Unit (WDRU), was introduced, accompanied by an updated backpropagation formula. Second, a novel dynamic spatial-temporal residual structure module was designed, incorporating the WDRU method to enhance the DSTF-Net framework's ability to fuse fault features of insulated bearings. Third, t-SNE was integrated to provide a visual explanation of the fault feature extraction process, significantly improving the interpretability and trustworthiness of the DSTF-Net. Last, the proposed framework was validated against eight existing methods using real-world insulated bearing datasets. Results demonstrated that DSTF-Net achieved fault identification accuracy improvements of 4.2 %, 7.5 %, 15.2 %, 20.3 %, 22.9 %, 26.7 %, and 29.7 %, respectively. This highlights the superior accuracy and generalization ability of DSTF-Net. In conclusion, the proposed framework significantly expands the application of DL in fault diagnosis tasks, offering notable engine