Optimal feature selection for a weighted k-nearest neighbors for compound fault classification in wind turbine gearbox

The k-nearest neighbors is renowned for its adaptability and ease of use, making it favored for data-driven turbine component fault diagnostics. However, the basic k-NN model is constrained by the curse of dimensionality, rendering it ineffective at capturing the dynamics in high-dimensional turbine gearbox vibration signals. To address this problem, this study advanced a framework for choosing the best features for a k-NN fault diagnostic model while also leveraging the benefits of “weighting” to further improve its performance. The framework was validated using vibration signals from a wind turbine gearbox under four different fault conditions. The study first extracted statistical frequency and time domain characteristics from the vibration dataset for feeding the model. The most discriminative features—including the mean, mean frequency, standard deviation frequency, maximum frequency, and Shannon entropy—were selected from the feature space using the proposed strategy. Results of the study indicate that by incorporating weights into the k-NN model, classification accuracy improved by 0.005 %. The Manhattan distance metric outperformed all other metrics in classifying the various gearbox health states. The optimal k-value was determined to be 20. Overall, the optimal k-NN model achieved an average classification accuracy of 95.95 % across all performance metrics, with accuracy at 95.97 %, recall at 95.97 %, precision at 95.93 %, and F1 score at 95.93 %. The fault diagnostic model is recommended for deployment in wind turbine gearbox condition monitoring.