Feature selection for artificial neural network model-based condition monitoring of wind turbines

Cross, Philip and Ma, Xiandong (2013) Feature selection for artificial neural network model-based condition monitoring of wind turbines. In: Proceedings of the 10th International Conference on Condition Monitoring and Machinery Failure Prevention Technologies (CM 2013 & MFPT 2013) :. BINDT, POL. ISBN 978-1-901892-37-6

Abstract

Over the next decade, electrical power generated from sustainable sources will become a significant part of the total generating capacity of many European countries. Of the energy sources available, wind power is considered one of the most viable resources. Wind farms are being increasingly constructed offshore to take advantage of stronger and more reliable winds; however, routine inspection and maintenance is more difficult in comparison to onshore wind farms. Further, the turbines are subject to a particularly harsh operating environment. In this regard, condition monitoring systems play an increasingly important role in the operation of offshore wind farms, enabling optimal scheduling of maintenance activities, and minimising the risk of unexpected failure of turbines. This paper proposes a model-based condition monitoring system employing multi-input, multi-output artificial neural networks, optimised using principal component analysis. The data for the models have been obtained from a wind farm SCADA (Supervisory Control and Data Acquisition) system. The artificial neural network models are used to predict the measured outputs for the turbines in the wind farm based on known inputs. The model predictions are compared with the actual output to identify potential faults, and are used to obtain adaptive thresholds for the selected output signals. In turn, these thresholds will form the basis of a system that provides an early warning of component failure and an indication of its severity. Multivariate artificial neural networks can perform poorly in comparison to univariate networks, caused by ‘over-fitting’ of the models due to the high correlation between many of the input parameters. In this paper, principal component analysis is used to minimise correlation, improving model fit and the accuracy of the models. In addition, the components that contribute least to the variation in the output signal are removed, reducing the amount of data processed by the neural networks and hence improving performance.