In English

Condition monitoring system for wind turbines - based on deep autoencoders

Niklas Renström
Göteborg : Chalmers tekniska högskola, 2019. 89 s.
[Examensarbete på avancerad nivå]

Wind turbines consist of many mechanical, electrical and hydraulic components. Failures in any of these can lead to high financial loss, both as actual repair costs and from lost production time. Often, failures do not occur instantaneously but rather as a consequence of sequential degrading. Therefore, many failures can be detected in advance using so-called condition monitoring systems. Through their supervisory control and data acquisition system, modern wind turbines store information about their operating state. Among other things, signals such as produced power, wind speed, and various component temperatures are recorded. In this thesis, a condition monitoring system that leverages this data is developed. The system is based on deep autoencoders, a type of neural network that learns to reconstruct its input data. By training an autoencoder on data from a healthy wind turbine it can learn the dependencies between different SCADA signals under normal conditions. If it then gives a poor reconstruction for new data, it is likely that something has changed in the internal dynamics of the wind turbine which could indicate a degraded component. Previously, many similar systems have been developed. These have shown good results and detected multiple component failures up to months in advance. However, they usually only monitor one component at a time and are therefore not able to provide a complete condition monitoring system. Autoencoders, which do not suffer this problem, have also been investigated but not at a larger scale. In this thesis, a relatively large, labeled dataset was utilized. With this data, the efficiency of condition monitoring systems based on autoencoders was tested on a variety of real faults. Moreover, the influence of various properties of the autoencoder was investigated. The results of the investigation showed that an autoencoder based condition monitoring system is capable of detecting a variety of failures in wind turbines. Finally, suggestions for future developments are discussed in the thesis.

Nyckelord: Wind turbine, condition monitoring system, anomaly detection, preventive maintenance, SCADA, machine learning, unsupervised learning, neural network, autoencoder, hyperparameter selection

Publikationen registrerades 2019-04-24. Den ändrades senast 2019-04-29

CPL ID: 256653

Detta är en tjänst från Chalmers bibliotek