Research on gear fault diagnosis of gearboxes under few-shot conditions based on DSCGAN

Abstract

The gearbox of quay crane is in heavy load, high speed and strong impact working condition for a long time, and the risk of gear fault is high. But their actual available fault samples are limited, resulting in the prominent problem of small sample and class imbalance in fault diagnosis. In this paper, a fault diagnosis method based on frequency-domain vibration image (FDVI) and dual-source conditional generative adversarial network (DSCGAN) is proposed. Firstly, the vibration signal is converted into FDVI to fully characterize the fault identities in frequency domain. The simulation data with teeth break fault is obtained from the dynamic simulation model of gearbox, and combined with the actual normal operation data to realize the dual-source fusion of the simulation fault characteristics within the actual background signals through DSCGAN, which generates high-quality and diversified pseudo-fault sample images. The augmented samples are used to train the CNN classifier to improve the diagnostic performance in small sample conditions. Experiments on the dataset of Tsinghua University (MCC5-THU) and quay crane scaled experimental platform show that the fault identification accuracy of the two datasets is improved from 90.12% to 99.84% and from 67.26% to 96.53% respectively after the augmentation of DSCGAN samples, which is significantly better than comparative models of DCGAN, DDPM, WGAN-GP, and VAEGAN. Ablation, hyperparameter sensitivity, and computational complexity analyses on the MCC5-THU dataset further verify the contribution of key modules, the robustness of training settings, and the practical feasibility of DSCGAN. The results show that the proposed method can effectively alleviate the problem of gear fault sample shortage and class imbalance, and improve the accuracy and robustness of the fault diagnosis of the quay crane gearbox.