Abstract: The operation data of wind turbines has the characteristics of multiple sources, fast changes in operating conditions, and low proportion of fault samples. Conventional diagnostic models are easily affected by insufficient labels and sample imbalance in the identification of minority faults and early anomaly capture. The article analyzes the modeling foundation of multi-source data intelligent diagnosis for wind turbines and constructs an intelligent diagnosis framework that integrates self supervised learning and imbalanced sample optimization. The research results indicate that the framework can enhance the model's ability to characterize the normal state, degradation trend, and abnormal fluctuations of the unit under insufficiently labeled data conditions, improve the tendency of missed diagnosis caused by uneven distribution of fault categories, and have important significance for intelligent operation and maintenance, fault warning, and condition based maintenance decision-making of wind farms.