Numerical simulation driven Cycle-GAN for domain generalization-based fault diagnosis of axial piston pumps

Deep learning models have been extensively employed for the fault diagnosis of axial piston pumps but they require abundant labeled fault data for model training. Since axial piston pumps mostly operate under healthy condition, it becomes difficult or even impossible to collect fault data in real applications. Although the fault data can be generated by numerical simulation, they often deviate significantly from real ones, leading to poor diagnostic accuracy. This paper proposes a domain generalization fault diagnosis method without requiring real fault data during model training. First, a computational fluid dynamics (CFD) model is constructed to simulate the discharge pressure data of an axial piston pump under healthy and faulty conditions. Second, a modified version of cycle-consistent generative adversarial network (Cycle-GAN) is designed to generate fake healthy and faulty data using these simulated data and the real healthy data. Finally, the simulated data, fake data and real healthy data are combined to train a classifier for the generalization from simulation to real scenarios. The proposed method achieves an average diagnostic accuracy of 93% in multiple experimental tasks, which is at least 17.8% higher than the comparative methods.