Improving the monitoring indicators of a variable speed wind turbine using support vector regression

Abstract

For over a decade, most wind turbines have worked by adapting their rotation speed to that of the wind. This operating method, now widely used, allows optimal tip speed ratio to be achieved whatever the weather conditions, and in fact produces much better output than stall controlled turbines, particularly in calm weather conditions. However, this improvement means that monitoring systems are required to adapt to constant macroscopic variations in load and speed. In addition, these non-stationary operating conditions make it difficult to undertake machine diagnostics over the long term, due to the fact that the operating conditions in which successive indicators are obtained will almost never be the same. The scientific community has, in many respects, proved the usefulness of regression analysis of these indicators in relation to properly selected variables. The focus of this paper is on regression methods based on machine learning tools, which are becoming more and more popular. The difficulty lies in designing a robust self-adaptive method for estimating the statistical behaviour of an indicator in relation to operating conditions. Indeed, the concern is that indicators may obey disparate and unpredictable multivariate laws: there are many complications which make it difficult to use linear regression tools. Kernel machines, used in this paper as a robust and efficient way of normalising indicators, have proved to be capable of greatly improving a monitoring system’s diagnostic capabilities. The demonstration is based on a practical example: monitoring a bearing defect by analysing the instantaneous angular speed of the wind turbine shaft line. As this defect can only be detected under certain operating conditions – a priori unknown – the chosen example will be particularly effective in highlighting the usefulness of such an approach.