GAN-Based Fault Diagnosis Method for Aircraft Hydraulic System Under Data Imbalance

Abstract

In actual operating of the aircraft hydraulic system, the frequency of occurrence of the fault status is much lower than that of the normal status. The faults of the aircraft hydraulic system are very expensive to replicate, making it unrealistic to conduct a real fault simulation experiment of the aircraft hydraulic system. Hence, a large number of data samples related to the normal status of the aircraft hydraulic system is available but very few data samples related to the fault status. This causes a severe data imbalance in the fault diagnosis of the aircraft hydraulic system, that is, the data samples for the normal operation are much more numerous than the fault data samples, which will directly affect the accuracy of aircraft fault diagnosis. To solve the data imbalance in the fault diagnosis of the aircraft hydraulic system, this paper used an auxiliary classifier and gradient penalty Wasserstein generative adversarial network (ACWGAN-GP) algorithm, which can stably and accurately generate high-quality simulated fault data using a small number of fault data, and perform accurate fault diagnosis by using the imbalanced fault data. The ACWGAN-GP algorithm expands the fault data set until a balance is achieved between the normal and fault data. It was verified by simulation that the simulated fault data obtained by this algorithm were highly similar to the real fault data. Hence, the generated data can be used similar to the real fault data. In addition, multiple data-driven intelligent fault diagnosis methods were used to verify the validity of the data and improve the accuracy rate of the fault diagnosis methods. It is concluded that the accuracy level of fault diagnosis increases steadily when the number of fault data is gradually increased until it finally reaches a balance with the number of normal data. In addition, the sample generation mode of the ACWGAN-GP model can be applied to fault diagnosis in other fields.