Supervised GAN based on joint gradient penalty for sparse sample fault diagnosis of harmonic reducer

Harmonic reducers enable high-precision torque transmission in humanoid and collaborative robots; however, their failure can cause severe accuracy degradation, unexpected downtime, and even structural damage. Data-driven fault diagnosis of harmonic reducers remains challenging, as labeled fault samples are typically scarce and fault signatures are subtle and easily masked by operational noise. To address these issues, this study proposes a supervised generative adversarial network with a joint gradient penalty mechanism (sGAN-JGP) for generating high-fidelity labeled fault samples under data-scarce conditions. First, a spatio-temporal tensor framework is constructed to extract robust multi-sensor representations that capture intrinsic space-time correlations. Second, the JGP mechanism constrains the discriminator's decision boundaries, effectively mitigating mode collapse and ensuring realistic feature generation. Finally, comprehensive experiments on a harmonic reducer testbed under three operating conditions validate the proposed method. With only 30 labeled samples per health state, sGAN-JGP achieves diagnostic accuracies of 98.48%–99.09%. Compared with representative mainstream generative models, sGAN-JGP outperforms the best-performing CWGAN-GP baseline by up to 2.42% in accuracy, while reducing the Maximum Mean Discrepancy of synthesized features to 0.032. These results demonstrate that sGAN-JGP is a robust fault diagnosis tool for sparse-data scenarios, well suited to real-world applications in collaborative and humanoid robots.