Presentations

Abstract

INTRODUCTION Maintenance activities in wind farms, which contribute to a significant portion of O&M costs, are generally reactive and schedule-based. This leaves considerable room for improvement by transitioning to a predictive approach to prognostics and health management. A key requirement for this transition is the ability to predict the remaining useful life (RUL) of wind turbine components. Existing approaches often assume a predetermined, typically linear trend for component degradation and rely on deterministic, supervised Machine Learning (ML) approaches to predict the RUL. However, component degradation is inherently stochastic due to uncertainties in material properties, loading conditions and environmental factors. Therefore, reliable RUL predictions must be probabilistic. On the other hand, the few existing probabilistic approaches generally use simple methods, such as ARIMA and Gaussian Processes, overlooking the enhanced performance of advanced Deep Learning (DL) frameworks. To overcome these limitations, this work proposes an unsupervised DL method based on Convolutional Autoencoder (CAE) to derive the component Health Indicator (HI) from SCADA sensor signals, which is shown to accurately reflect its true degradation trend. Then, a Conformal Prediction method is used to probabilistically forecast the HI’s future trend and predict when it will cross the failure threshold, indicating failure. This approach constructs prediction intervals with coverage guarantees, making the prediction of the RUL highly reliable. Methodology HI CONSTRUCTION The proposed approach assumes that the Health Indicator (HI) is monotonic without assuming a specific trend. A CAE is trained to find a nonlinear decomposition of the sensor signals,  separating the degradation factor from those related to the operational and environmental conditions. The training uses a hybrid of Backpropagation and Particle Swarm Optimisation algorithms to simultaneously maximise the monotonicity of the CAE middle layer output, which becomes the HI, and minimise the reconstruction error. It is shown that this approach is able to effectively extract the true underlying component degradation trend, which can lead to a higher accuracy in the RUL prediction step. RUL PREDICTION Once the training set HIs are built, a portion of them is set aside as the calibration set. An LSTM neural network is trained using the remaining HIs to predict one-step-ahead HI values. The absolute difference between the actual and predicted one-step-ahead HI value is defined as the non-conformity metric and is calculated for the calibration set forecasts, based on which the prediction interval is determined. Subsequently, the trained LSTM model is used to predict the future trend of the test set HIs by iteratively forecasting the one-step-ahead HI values and aggregating the prediction intervals at each step until both the upper and lower intervals cross the failure threshold. RESULTS AND CONCLUSIONS Applying the proposed approach to a SCADA dataset with multiple bearing failures shows that the RUL predictions are more accurate and exhibit a lower variance compared to commonly used simple time series forecasting methods. Furthermore, the prediction intervals provide more precise coverage than those constructed by bootstrapping the training residuals. This demonstrates the effectiveness of this approach in reliably predicting the RUL of wind turbine components.