Wind Turbine Fault Detection an‎d Isolation Robust Against Data Imbalance

4/21/2026 12:00:00 AM

In order to improve the reliability of wind turbines, prevent serious accidents, reduce maintenance costs an‎d increase efficiency, effective fault diagnosis an‎d isolation of these wind energy converters is of particular importance. Due to the difficulties of system modeling, nonlinear effects, uncertainties an‎d availability of sensor data of wind turbine SCADA system, data-driven fault diagnosis methods for wind turbines have gained increasing interest. In this thesis, using wind turbine SCADA data, KNN classifier an‎d its ensemble model, fault diagnosis an‎d isolation an‎d fault prognosis are effectively performed, respectively. First, the operational data are labeled using wind turbine SCADA status an‎d warning datasets, an‎d the labeled dataset, after invalid data elimination, feature selec‎tion an‎d stan‎dardization, is used for training an‎d validation of the proposed model. The model parameters are selec‎ted using the 10-fold validation method, an‎d the classifier training is performed employing the holdout method with an 80-20 ratio. Using the time window idea, the fault prognosis problem is also simplified into a fault diagnosis problem. A robustness criterion of the KNN classifier against data imbalance is presented based on the distances of the data classes’ observations, demonstrating that data imbalance, which is common in real data sets, does not affect the performance of the proposed model, so there is no need for data balancing methods in this algorithm, an‎d the performance metrics are not affected by false alarms. Therefore, the proposed method has provided significant accuracy in fault diagnosis an‎d prognosis compared to previous research on this dataset. Although the collected dataset is used to train the proposed model offline, this model has the ability to assign labels to new observations online. In addition, the proposed model determines the label of new observations only based on the values of its current features’ values. Existing SCADA data is used to train an‎d validate the proposed algorithm, so there is no additional cost for data collection an‎d recording. Many of the faults that are addressed in this thesis, including status an‎d warning faults, have not been considered in the previous research on this dataset. Also, the dependence between different faults based on physical facts an‎d the overlap of these fault occurrence is analyzed.

توربين بادي , تشخيص و جداسازي داده محور عيب , پيش بيني عيب , داده نامتوازن , داده اسكادا , دسته بند KNN , دسته بند تركيبي KNN

Wind turbine , Data-driven fault detection an‎d isolaton , Fault prognosis , Imbalanced data , SCADA data , KNN classifier , Subspace KNN classifier

Ali Fazli

Javad Poshtan