Abstract: There are significant differences in the load characteristics and fault evolution laws of wind turbines in three scenarios: onshore plains, mountainous terrain, and offshore environments. A fault diagnosis system with multi-source signal acquisition, time-frequency domain feature extraction, and fusion is constructed for three key components: gearbox, blades, and main bearings. A hybrid model of LSTM-CNN parallel dual branch is designed. The average diagnostic accuracy of the three scenarios can reach 95.2%, which is 14.8% higher than traditional support vector machines. The differentiated maintenance decision system reduces the average number of unplanned shutdowns by 38.6% and compresses operation and maintenance costs by 27.2%, providing systematic technical support for intelligent operation and maintenance of wind turbines under different geographical and climatic conditions.