Exploration of Unsupervised Deep Learning-Based Gear Fault Detection for Wind Turbine Gearboxes

Gearboxes are critical mechanical components in various modern constructions, including wind turbines, making their real-time monitoring and the prevention of major failures essential. Machine learning (ML) offers a precise and robust method for early-stage failure detection and efficient gear monitoring during operation, with computational efficiency that allows for use on edge devices. This article presents a method for detecting surface damage on gear teeth using unsupervised machine learning. Using only experimentally measured vibrational signals from a healthy gearbox as a training set, novel neural network architectures, including convolutional and recurrent autoencoders, were employed and compared with a classical dense autoencoder. The study confirmed the effectiveness of these methods in gear transmission diagnostics and demonstrated the potential for achieving high-quality classification metrics using unsupervised learning.