Gearbox Fault Detection Using Continuous Wavelet Transform and Vision Transformer (ViT)

Managing the substantial volume of data generated by gearbox operations and the inherent complexities underscores the strategic advantage of employing deep learning methods. This study addresses fault detection in the gearbox using a Vision Transformer (ViT), whose images were prepared by continuous wavelet transformation. The dataset from the Open Energy Data Initiative (OEDI) serves as the foundation for our analysis. Data recorded by four vibration sensors, situated in the time domain and various operating loads, transform into two-dimensional images which, after some manipulations, are fed into a ViT, allowing us to tackle the classification challenge with 'Healthy' and 'Damaged' classes. Remarkably, With the test dataset, we were able to obtain an exceptional accuracy of %99.1554, along with a decrease in the BCE loss function to 0.040100. These results demonstrate the ViT's remarkable efficacy in detecting faults using images derived from signals recorded in the time domain.