Fault detection in wind turbines: a supervised learning approach with multilayer perceptron neural network

Abstract

The expansion of energies aligning with sustainable development requirements is continually progressing. A particular focus on wind turbines (WT), characterized as a highly intricate system, has brought attention to the costs associated with operation and maintenance. This has prompted a quest for increasingly efficient maintenance procedures. This article proposes that deviations from expected behavior be detected and classified as failure using operational data from permanent magnet direct drive wind turbines, obtained through the supervisory control and data acquisition system. To achieve real-time fault detection and gather information on the state of faults, the CFS subset evaluator method was employed to extract the most relevant attributes from the dataset. Experimental results demonstrate that the strategy of selecting the most pertinent attributes for the Multilayer Perceptron algorithm (MLP), comprising four layers, in fault classification, resulted in a high recognition rate for detecting and classifying faults in wind turbines. Computational results validating the models are presented.