A machine learning framework for multi-class wind turbine blade fault classification using statistical and wavelet features

Abstract Reliable fault diagnosis is essential for improving the safety, reliability, and operational efficiency of wind turbines. This study presents a machine learning (ML) framework for multi-class classification of wind turbine blade faults using vibration signals. The proposed approach combines statistical features and wavelet packet energy features to characterize the dynamic response of the blades under five health conditions: healthy, crack, erosion, twist, and unbalance. A total of 26 features, comprising 10 statistical and 16 wavelet packet energy features, were extracted from the vibration signals and used to train and evaluate seven conventional ML classifiers, including support vector machine (SVM), random forest, K-nearest neighbors (KNN), logistic regression, decision tree, gradient boosting, and Gaussian Naïve Bayes (GNB). Model performance was evaluated using stratified five-fold cross-validation. Among the evaluated models, the random forest classifier achieved the highest and most consistent performance, with an accuracy of 99.05 ± 1.17\%. Feature importance analysis further showed that both statistical descriptors and wavelet packet energy features contribute significantly to fault discrimination, highlighting the complementary nature of time-domain and time-frequency information. The results indicate that the proposed feature-based conventional ML framework provides an accurate, computationally efficient, and interpretable solution for multi-class wind turbine blade fault diagnosis, making it well suited for practical structural health monitoring (SHM) applications.