Remaining Useful Life Prediction of Wind Turbine Main-Bearing Based on LSTM Optimized Network

The remaining useful life (RUL) prediction is a crucial aspect of predictive maintenance for equipment. However, main bearings operate in complex, high-frequency vibration-prone nacelle conditions, identifying fault characteristics and accurately predicting the degradation process pose significant challenges. To address this, this article presents the first-ever research on fault diagnosis and prognosis based on historical vibration data from in-service wind turbine units with confirmed damages in a certain offshore wind farm. Proposing a tree seed algorithm optimized long short-term memory (TSA-LSTM) predictive model founded on the sideband energy ratio (SER) principle. Considering the distinctive structural configuration of wind turbines, fault indexes are extracted from the modulation component composed of the main bearing defect frequency occurring on both sides of the gearbox characteristic frequency based on SER. The time sensitivity of LSTM and the excellent global search ability of TSA were combined to complete modeling. This approach effectively captures the degradation process and achieves accurate fault identification. Through comparison, the superior predictive performance and robustness of the proposed optimization algorithm are verified, with mean absolute percentage error (MAPE) lower than 0.228 and root mean square error (RMSE) lower than 0.014. Additionally, the exponential fitting can accurately describe the whole process of the gradual accumulation of faults over time until failure, facilitating RUL prediction. The results demonstrate that SER-based features exhibit higher sensitivity to early-stage faults and better fit the degradation trend, providing a promising solution for wind turbine main bearing fault prognosis.