Diagnosis of rotor faults through generative adversarial networks enhanced by embedded wavelet scattering transforms

Fault diagnosis methods based on deep learning typically require a substantial amount of data. However, in practical engineering applications, it is often challenging to collect sufficient vibration signals due to various constraints, affecting the performance of these diagnostic methods. This paper proposes a Wavelet Scattering Transform-based Generative Adversarial Network model (WST-GAN). In this model, the wavelet scattering transform is embedded into the discriminator to enhance its discriminative capability. The performance of the generator is improved through adversarial training, which also increases the volume of training samples and enhances the model's generalisation ability. The method includes offset resampling of limited samples to augment the initial sample size and converts signals into matrix images through pixel stacking, which are then fed into the generative adversarial network for training. The approach is applied to a case study using vibration signals from a simulated gas turbine rotor system. The fault detection accuracy of the dataset that underwent data augmentation exhibited a significant improvement compared to conditions with a small sample size, achieving a diagnostic accuracy of at least 97.03%. Experiments are conducted using a dual-disk rotor, and the identification accuracy for nine different fault conditions approached an accuracy of at least 99.47%.