A Small Sample Generation Method Based on a Periodic-RealNVP and its Application in the turbine cooling air flow prediction for aero-engines

In the data-driven soft sensing modeling of aero-engine operation, the quantity of samples significantly impacts the accuracy of the developed model. To address the challenge posed by a limited number of training samples, a data generation method is proposed based on a flow model, utilizing a unique masking mechanism and an improved affine coupling layer. Specifically, this method simplifies the computation by analyzing the correlations among variables to provide prior knowledge for affine coupling. Furthermore, the proposed algorithm reconstructs the dataset by establishing a mapping between the true distribution and the latent distribution through iterative transformations of the affine coupling layer, which can capture periodic characteristics. By applying this method to the aero-engine operation data from Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset, its effectiveness in achieving good generalization performance and modeling accuracy has been validated where the accuracy of small sample soft sensing has an improvement of 20~30% than the other state-of-the-art (SOTA) methods.