A fault diagnosis method for wind turbine gearbox based on adaptive probability random forest is proposed to address the issue of noise pollution in SCADA data of wind turbine gearbox. Firstly, SMOTE oversampling is used to balance sample categories, and then CART is trained and classified by constructing multiple balanced subsets. The sample error rate represents the weight of sample ambiguity, and the label uncertainty is determined. Monte Carlo simulation is used to calculate the mean distribution of features, which is fused with each sample instance to obtain the uncertainty of sample features. Utilizing adaptive labels and sample uncertainties as inputs to probabilistic random forest can enhance the ability to manage feature noise and label noise, thereby improving the robustness of fault diagnosis. Conduct an experimental evaluation using the SCADA dataset of wind turbine gearbox. The results show that this model outperforms other methods in terms of false alarm rate, false alarm rate, and F1 rating metrics when dealing with missing values, Gaussian noise, and label noise in the dataset, as compared to other methods. This method is of great significance for improving the accuracy and robustness of wind turbine gearbox fault diagnosis.