The ultimate goal of a condition monitoring system for wind turbines (WTs) is to predict upcoming failures; this could be achieved using artificial intelligence techniques. In this paper, a model for detecting excessive temperature anomalies in key components of WTs, ie gearbox, generator and transformer, is proposed. This model consists of integrated modules that continuously interact following the never-ending learning paradigm based on artificial neural networks addressing the challenge of limited pre-classified data and understanding of the concept to be learned for a system with continuous change in its operating conditions: (i) the normal behaviour (NB) module estimates the temperature of the WT key components; (ii) the expected time to failure (ETTF) module calculates the deviation between the estimated normal temperature and the real-time measurement data to predict the upcoming failure of WT key components a few hours before a failure occurs; and (iii) in the anomaly detection (AD) module, the temperature deviation time-series signal is divided into normal or abnormal clusters. The proposed methodology has been applied on data from a real wind farm in Germany. The results show that the system could correctly detect a large number of upcoming WT failures; this implies the effectiveness and generalisation of the proposed model in terms of classification accuracy.
展开▼