SCGAN-TrAdaBoost: A Skip-connection Generative Adversarial Network with Transfer AdaBoost for Bearing Fault Diagnosis with Limited Data

Abstract Intelligent fault diagnosis is essential for ensuring the reliability and safety of mechanical systems, particularly in key parts like bearings. A significant challenge lies in the limited data, which hinders the development of robust data-driven diagnostic models. While generative adversarial networks (GANs) hold considerable promise for generating synthetic data, they often struggle with instability and produce data lacking in fidelity. Improving the fidelity of synthetic data generated by GANs invariably necessitates increased computational resources. To address this, a new approach called SCGAN-TrAdaBoost is proposed, combining skip-connection GAN (SCGAN) with GAN-TrAdaBoost. SCGAN incorporates lightweight skip connection (LSC) modules in the generator to enhance the fidelity of synthetic data by bridging multi-scale features. Furthermore, while the integration of the LSC module is an incremental addition, its inclusion of 1×1 convolutional layers ensures minimal impact on computational cost due to their lightweight nature. GAN-TrAdaBoost further improves the use of synthetic data by adaptively adjusting its influence during training, minimizing the negative effects of low-fidelity data. Experiments on two bearing fault datasets demonstrate that SCGAN-TrAdaBoost achieves superior diagnostic accuracy compared to state-of-the-art methods.