| Summary: | The need for renewable energy has been growing in recent years for the reasons we all know, wind power is no exception. Wind turbines are complex and expensive structures and the need for maintenance exists. Conditioning Monitoring Systems that make use of supervised machine learning techniques have been recently studied and the results are quite promising. Though, such systems still require the physical presence of professionals but with the advantage of gaining insight of the operating state of the machine in use, to decide upon maintenance interventions beforehand. The wind turbine failure is not an abrupt process but a gradual one. The main goal of this dissertation is: to compare semi-supervised methods to at tack the problem of automatic recognition of anomalies in wind turbines; to develop an approach combining the Mahalanobis Taguchi System (MTS) with two popular fuzzy partitional clustering algorithms like the fuzzy c-means and archetypal analysis, for the purpose of anomaly detection; and finally to develop an experimental protocol to com paratively study the two types of algorithms. In this work, the algorithms Local Outlier Factor (LOF), Connectivity-based Outlier Factor (COF), Cluster-based Local Outlier Factor (CBLOF), Histogram-based Outlier Score (HBOS), k-nearest-neighbours (k-NN), Subspace Outlier Detection (SOD), Fuzzy c-means (FCM), Archetypal Analysis (AA) and Local Minimum Spanning Tree (LoMST) were explored. The data used consisted of SCADA data sets regarding turbine sensorial data, 8 to tal, from a wind farm in the North of Portugal. Each data set comprises between 1070 and 1096 data cases and characterized by 5 features, for the years 2011, 2012 and 2013. The analysis of the results using 7 different validity measures show that, the CBLOF al gorithm got the best results in the semi-supervised approach while LoMST won in the unsupervised scenario. The extension of both FCM and AA got promissing results.
|
|---|