The condition monitoring of wind turbine (WT) main bearings is critical in enhancing wind farm economic efficiency and advancing wind power development. However, as complex detection models grow, their parameter counts increase substantially. This proliferation of parameters strains computational resources, reduces model interpretability, and complicates parameter tuning. To address these challenges, this paper proposes an early fault detection model for WT main bearings based on neural circuit policies. Firstly, the key features are extracted from Supervisory Control And Data Acquisition (SCADA) data using a hybrid architecture combining convolutional neural networks and long short-term memory networks. Secondly, these features are then processed by neural circuit policies containing only 19 neurons to complete model training. Finally, validation across four WT datasets demonstrates the model’s effectiveness. Results indicate that the proposed model achieves, and in some cases surpasses, the performance of benchmark models with 2–3 times more parameters while requiring minimal computational resources. During healthy operation, the model reduces RMSE by an average of 6.4%. More critically, it provides timely early warnings for bearing abnormalities with zero false alarms.