Fault Diagnosis Method for Wind Turbine Based on Feature Optimization and Deep Learning Integration

To address the high dimensionality and strong redundancy of wind turbine operational monitoring data, as well as the insufficient robustness ability of a single deep network under imbalanced fault sample distributions, this paper proposes a feature selection and deep ensemble learning method for wind turbine fault diagnosis. First, SCADA monitoring data are cleaned and normalized. Mutual information is used to evaluate the statistical dependence between monitoring variables and fault categories, while Pearson correlation analysis is employed to remove highly correlated redundant variables. Then, multiple rounds of Recursive Feature Elimination (RFE) are conducted on the pre-screened candidate feature set, and stable fault-sensitive features are selected according to their selection frequencies under different random data splits. The optimized feature subset is subsequently fed into Stacked Autoencoder (SAE) weak classifiers to extract high-level nonlinear fault features. AdaBoost is further introduced to update sample weights iteratively, enabling the model to focus more on minority-class faults and hard-to-classify samples. Finally, multiple SAE weak classifiers are weighted and integrated into a strong classifier for wind turbine fault diagnosis. The proposed method combines redundancy suppression, stable feature selection, deep feature learning, and ensemble classification, providing an effective approach for wind turbine fault warning and condition-based maintenance.