Wind turbine failure prediction and health assessment based on adaptive maximum mean discrepancy

Abstract

Wind turbines are being designed with increasingly sophisticated electrical and mechanical components, leading to more complicated maintenance procedures and higher failure costs. Much of the research in this area has provided reliable early warning of wind turbine failures in an effort to reduce downtime and maintenance costs. Recently, data-driven approaches have been introduced into the health status assessment of wind turbines. However, obtaining high-precision failure predictions and interpretable health status assessments is still challenging. In this research, we propose a unified framework for predicting failures in and assessing the health of wind turbines. First, we empirically grouped wind turbine failures into four categories. Low frequency failures involving extensive downtime are critical, due to their tendency to gradually deteriorate and the severe consequences they introduce; thus, these were adopted here for use in the predictive maintenance strategy. The adaptive maximum mean discrepancy algorithm, a novel similarity metric for distributions, was used as a means of capturing the dynamic weights of sensor attributes and predicting potential failures. A convolutional neural network was then employed to identify false alarms and improve the precision of the model. This approach will significantly improve the prediction rate and reflect the relationship between failures and features/sensors, which is essential to further improve the performance of wind turbines. Experiments showed that our health assessment method had a higher failure prediction rate and a lower false alarm rate than the traditional method. The process also provides the gradual degradation trend for wind turbines, as influenced by specific failures. An empirical study of one-year of SCADA data from six wind turbines was further used to demonstrate the interpretability of the approach.