Abstract: To address issues of feature redundancy, low prediction accuracy, weak generalization ability, and high false alarm rates in existing gearbox fault warning methods, a fault warning method based on neural basis expansion analysis was proposed. The method combined Sigmoid growth curve with a density-based spatial clustering algorithm to extract normal monitoring data from raw data. An improved grey relational analysis algorithm was used to assign weights to different data points, variable information was deeply explored. Feature variables related to the oil temperature of the gearbox were extracted from the data collected by the high-dimensional monitoring and data acquisition system. Local nonlinear projection was used to decompose the target signal into basis functions to predict the state variables of the gearbox accurately, fault thresholds were set by sliding window and confidence interval to reduce the false alarm rate. The method was verified based on actual wind field data. Results show that the fault prediction method based on the NBEATSx model significantly improves the prediction accuracy in the state prediction of gearboxes, reduces the false alarm rate compared to traditional models, and has the ability to predict faults several hours in advance, thereby effectively ensuring the stable operation of wind turbine units.