Wind turbine fault diagnosis via ensemble learning with adaptive feature engineering

Abstract Fault diagnosis of wind turbines is essential for ensuring the safe and reliable operation of wind farms However, Supervisory Control and Data Acquisition (SCADA) data are usually high-dimensional, redundant, and nonstationary, posing challenges to accurate classification and model generalization. To address these issues, this paper proposes a data-driven fault diagnosis framework. First, the k -nearest neighbors algorithm is applied to detect and remove anomalous samples, improving data quality. Then, a two-stage Adaptive Feature Engineering (AFE) approach is developed: physics-informed features are designed to characterize intra-turbine thermal imbalance and coupling, and Light Gradient Boosting Machine (LightGBM) combined with permutation importance adaptively selects the most discriminative subset. An ensemble classifier integrating LightGBM, long short-term memory, and multilayer perceptron is then constructed, and the whale optimization algorithm is employed to fine-tune its hyperparameters. Experimental results on real SCADA data demonstrate that the proposed method significantly outperforms individual baseline models in both accuracy and robustness. Compared with the baseline LightGBM model, the optimized ensemble improves the accuracy from 0.9279 to 0.9913, with absolute gains of 6.34%, 6.23%, 7.27%, and 6.95% in accuracy, precision, recall, and F1-score, respectively, confirming its effectiveness and engineering applicability.