Comparison of Convolutional Neural Networks Using Gramian Angular Difference Field, Recurrence Plot, and Markov Transition Fields for Rotor Imbalance Detection in Wind Turbines

Abstract This study investigates various data representations derived from wind turbine dynamics, utilizing convolutional neural networks (CNNs) as diagnostic tools to identify rotor mass imbalance. The analysis examines three distinct methods for encoding the dataset into 2D images: Gramian angular difference field (GADF), recurrence plot (RP), and Markov transition fields (MTF). To achieve this, simulations of a 1.5 MW three-bladed wind turbine model were conducted using turbsim, fast, and matlabsimulink. These simulations generated rotor speed data under diverse conditions, including varying wind speeds and induced mass imbalances achieved by adjusting blade density in the software. Comparative analyses across nine classifiers demonstrated that the Gramian angular difference field provided the most promising results. The study presents detailed results and simulations, offering a thorough evaluation of the proposed architecture's effectiveness in detecting and diagnosing rotor mass imbalances in wind turbines.