Data synthesis via supervision-enhanced CDCGAN for aero-engine bearing fault diagnosis under noise and small samples

Abstract In practical engineering scenarios, it is challenging to acquire sufficient fault data for aero-engine bearings, and they are often exposed to a noisy environment, both of which increase the difficulty of fault diagnosis. This paper proposes an improved conditional deep convolutional generative adversarial network framework with enhanced supervision and robust performance, which solves the problem of low accuracy in small-sample fault diagnosis for aero-engine bearings under noisy conditions. First of all, a confidence mechanism that has the ability to filter out low-fidelity data is designed to enhance the degree of supervised learning. Accordingly, high-quality synthetic data are generated. Then, to establish the balance between the discrimination and classification, two dedicated discriminators, the authenticity discriminator and label discriminator, are proposed to take on the tasks of authenticity judgment and label prediction individually. As a result, the training process of the model becomes more stable and converges faster because of no task interference. Furthermore, with the aim of improving fault diagnosis performance under noise, the composite loss function of the generator is constructed based on the feature error and reconstruction error losses. Consequently, the discrepancies between the generated data and the original data are constrained from the perspective of features and pixels. Finally, experimental results on two datasets demonstrate that the proposed model performs well with limited data, achieving accuracies of 97.62% and 96.18% using only 6% of the training set. Meanwhile, it achieves the maximum accuracy improvements of 9.96% and 11.87% under different noise distributions, demonstrating strong noise resilience.