IMPROVED DEEP LEARNING APPROACH IN WIND TURBINE DAMAGE DETECTION

Abstract

The growing number of wind farms is creating an increased demand for their trouble-free operation. Damage to their components can be catastrophic. A particular component that can be subject to damage during long-term operation and is difficult to diagnose is the mechanical gearbox located in the turbine. Traditional approaches to the subject of damage detection require, in the final stage, the involvement of an expert. Therefore, the article proposes a method based on the Deep Learning solution. A transfer learning method and a pre-trained Inception V3 network were used. A gearbox with in three states of healthy, worn out but still working and damaged was analyzed. Signal spectrograms were created from accelerometric measurements and then used as input for the neural network. Various approaches to creating spectrogram images were tested. The InceptionV3 network was taught on images generated in grayscale, and RGB and HSV. Channel reduction in the form of using grayscale improved the speed of the algorithm at the expense of precision. The use of HSV scale, on the other hand, made it more precise in detecting a worn out state.