Few-Shot Diffusion Domain Generalization for Diagnosing Joint Reducer Faults in Industrial Robots

Abstract

Reliable fault diagnosis of joint reducers is essential for ensuring the operational efficiency of industrial robots. However, obtaining sufficient data corresponding to different fault types is a considerable challenge, as normal operation data are more easily obtainable than fault ones for real applications. To address this issue, a few-shot diffusion domain generalization (FSDDG) approach is proposed for diagnosing industrial robot joint reducers with few faulty samples for model training. An improved U-Net based on a one-dimensional convolutional neural network is first designed to improve computational efficiency as well as decrease computational burden. The improved U-Net is integrated into a diffusion model for data generation from few faulty samples. A domain generalization model is then trained on the source domain composed of generated fault data and collected normal data. This enables semi-supervised transfer of source domain knowledge to the target domain, which consists of a large amount of normal data and a small amount of faulty data. Experimental validations on an industrial robot test rig and a public dataset demonstrate the superior performance of the present FSDDG compared to state-of-the-art approaches, highlighting its robustness and reliability for the joint reducer fault diagnosis. This work provides a robust foundation for advancing fault diagnosis in industrial robots, ensuring their reliable and efficient operations.