Deep autoencoder for condition monitoring of wind turbines - Detecting and diagnosing anomalies

Abstract

Over the last decade, energy production from wind turbines has grown by 400%, prompted by public investments and climate change awareness as well as advances in technology. With subsidies offered to wind farms dwindling, the owners and operators of wind farms are forced to cut operational cost to stay profitable. This has lead to a renewed focus on predictive and preventive maintenance, targeting not only the traditionally well monitored large components in the wind turbine, but also including smaller, more easily replaced components. Advances in the use of machine learning to model complex system, combined with the growing access of data have allowed advanced methods for condition monitoring and anomaly detection to be developed. This has been applied to the field of condition monitoring for wind turbines in various research projects, utilizing data from the Supervisory Control and Data Acquisition (SCADA) system available in modern wind turbines. Most of these systems have modeled just one component at a time and are therefore not able to provide a complete condition monitoring system. Using autoencoders for condition monitoring of wind turbines enables the whole wind turbine to be modeled as one system, by learning the internal connections between the SCADA signals. This has previously been studied, but most studies focus on the anomaly detection step, and leave out the important part of diagnosing which signals are most affected by the anomaly. By finding these signals, the source of the fault can be found, which is important to allow for recommendation on where to do maintenance. This thesis investigates the application of deep autoencoders to detect and diagnose developing faults. The autoencoder has been used to produce a residual, taken as the error between the input to the autoencoder and its reconstructed signal. For the fault detection, the Mahalanobis distance has been used on the residual. For fault diagnosis, the residual for each signal has been standardized and analyzed to examine which signals are mostly affected by the fault. This was tested on eight known faults found in five different components: gearbox, cooling system, hydraulic system, yaw encoder issue and generator slipring. The proposed condition monitoring system was successful in detecting and diagnosing all faults but one. This thesis also presents an approach to understanding what the autoencoder has learned, with the use of simulated faults. The study provides a good method for discovering what connections between the SCADA signals the autoencoder has learned as well as information about how the residual is affected when one signal is experiencing a fault.