Fault detection framework in wind turbine pitch systems using machine learning: Development, validation, and results

This work develops a methodology for detecting faults in wind turbines through supervised machine-learning classification methods focusing in two modes of operation. The Normal mode will be related to the regular operation of the wind turbines and in the case of the abnormal mode, the system will be cataloged as Pitch system failure. This methodology uses real data through the supervisory control and data acquisition (SCADA) system and the recording of events and actions (LOG alarm system). Sixteen classifications artificial intelligence (AI) models were tested, such as Random Forest Classifier (RF), Gradient Boosting Classifier (GBC), Extra Trees Classifier (ET), Extreme Gradient Boosting (XGBOOST), Light Gradient Boosting Machine (LIGHTGBM), Categorical Boosting Classifier (CATBOOST), Adaptative Boosting Classifier (ADA), Decision Tree Classifier (DT), Linear Discriminant Analysis (LDA), K-Nearest Neighbors Classifier (KNN), Logistic Regression (LR), Naive Bayes (NB), Quadratic Discriminant Analysis (QDA), Dummy Classifier (DC), Super vector machine (SVM), Ridge Classifier (RC). The training process model was carried out by structuring the dataset in 75% for training and validation and 25% for the final test, considering cross-validation and optimization of the hyperparameters. The Random Forest Classifier (RF) and Extra Trees Classifier (ET) classifiers models presented the best performances, and the models with the lowest performances were K-Nearest Neighbors Classifier (KNN) and Linear Discriminant Analysis (LDA) ones. The average hit effectiveness for both Healthy and Faulty operating modes was approximately 80% across most of the models developed.