PIGAN: Physics-Informed Generative Adversarial Network for Fault Data Generation with Dynamic Model Priors

Despite rapid progress in intelligent gear fault diagnosis, it still faces two bottlenecks, namely the scarcity of labeled fault data and the computational burden of digital-twin-based approaches. To address this, we propose a physics-informed generative adversarial network (PIGAN) that integrates physical priors with data-driven learning. A Fourier feature mapping layer encodes multi-scale frequency content to better capture gear-meshing harmonics and transient impacts, while a physics-informed neural network imposes a gear dynamic model as a hard constraint to enforce governing laws. An adversarial framework further learns latent distributions from a few samples, thereby enhancing generalization and sample diversity. Experiments on a two-stage gear transmission system show that PIGAN generates physically consistent vibration data with high time-and frequency-domain fidelity, proving the effectiveness of the proposed method.