Comparative Study of Deep Learning Models Versus Machine Learning Models for Wind Turbine Intelligent Health Diagnosis Systems

Wind turbines technology is one of the current solutions and a promising one for renewable green energy. Due to its typical remote installations either on land or off-shore, there is a great interest in autonomously monitoring their system's health. The gearbox of the turbine is the primary source of major failures that render the system unsafe, inefficient and/or dysfunctional. Although there have been number of attempts in addressing this problem, number of gaps exist including the lack of detecting early signs of abnormal signals, deep learning (DL) solutions which are rare in this specific area of interest, the impact of number of features in convolution neural networks (CNNs) which was never studied, and the lack of comparative studies between DL versus conventional machine learning (ML) models for this problem of interest. This work presents an investigation that will help close the gap in literature by addressing all four aforementioned issues. Our approach is to propose a CNN model and test it against 6 conventional ML models of supervised learning models (multilayered perceptron, discriminant analysis classification and K-nearest neighbors) and unsupervised learning models (self-organizing maps, K-means clustering and Gaussian mixture model). For model testing, real data were acquired from the US National Renewable Energy Laboratory which includes sensor readings for number of critical parts of the gearbox for healthy and abnormal scenarios. Significant experimentation was conducted on all the 7 models, and observations and results discussion are presented. Among our interesting findings, the auto-learning ability of the CNN model was superior to the most powerful feature extraction technique in digital signal processing, the FFT.