Wind turbines are complex electromechanical hybrid systems that convert wind energy into electrical power. Their transmission systems include numerous high-precision mechanical components such as gears and bearings. These components are prone to fatigue and wear during prolonged operation, making the transmission system vulnerable to various failures, which significantly impacts both the safety of the units and their power generation efficiency. To ensure the safety of wind turbines, it is necessary to efficiently and accurately detect abnormal states in the transmission system. This paper addresses the limitations of existing research in capturing complex operational states and the low accuracy in component maintenance and white noise rejection. A state monitoring model based on a learnable decomposition and dual-attention network using autoregressive embeddings is proposed. First, the data collected by the supervisory control and data acquisition (SCADA) system is preprocessed, and correlation analysis is performed to select feature parameters such as temperature that are strongly correlated with fault evolution. Next, an autoregressive embedding module is introduced, utilizing dynamic tokens to better capture multidimensional time-series features. This allows for the prediction of relevant temperature variables in the transmission system, achieving dynamic modelling of fault characteristics. Then, a dual-threshold discrimination monitoring network is proposed, combining the residuals and information entropy of relevant variables to determine the dual health thresholds, further eliminates abnormal signals such as white noise and providing accurate warning times. Finally, the effectiveness of the proposed model is verified through two actual wind turbine transmission system failure cases. Compared to traditional SCADA systems, early warning signals can be detected approximately 6 to 10 days in advance when a fault occurs in the wind turbine.