Wind Turbine Rotor Fault Prediction Based on Convolutional Neural Network

Condition monitoring and early deterioration trend prediction are important to reduce downtime and economic loss. In this paper, the status of wind turbine before downtime are classified into Normal, Risky and High-risk based on the time to the moment of failure. Vibration signal analysis techniques are applied in de-noising and feature engineering, suitable feature combination is selected for prediction models. We visualize the selected features as two-dimensional line charts, and the convolutional neural network (CNN) algorithm is used to predictive modeling. To verify the effectiveness of proposed method, the prediction scores of several classical machine learning algorithms are compared with the prediction scores of CNN. At the same time, considering that most evaluation indexes are static, which ignore the dynamic of online prediction. In this study, a time-series prediction ability evaluation index (TPAEI) is proposed to select a prediction model with small prediction error and good consistency. The effectiveness of the proposed method is verified by wind turbine rotor data sets.