Blade performance prediction model coupled autoencoder with multi-source data fusion strategy

The use of artificial intelligence to develop data-driven aerodynamic performance prediction models has recently become a research hotspot, as it offers significant advantages in computational speed and cost compared to traditional wind tunnel experiments and computational fluid dynamics (CFD) simulations. However, loss models for compressor blade profiles based on single-source data often suffer from limited accuracy and poor generalization in small-sample scenarios. This paper proposes a blade performance prediction model coupled autoencoder with multi-source data fusion strategy (CAM), which demonstrates excellent performance on small-sample datasets. The CAM model comprises three submodules: isentropic Mach number prediction submodel, convolutional autoencoder submodel, and total pressure loss prediction submodel, and the tri-modal training strategy is proposed to train CAM model. The CAM model employs a convolutional autoencoder for feature extraction, encoding high-dimensional loading distributions into a two-dimensional latent space. The denoising capability of the convolutional autoencoder effectively mitigates error propagation between submodels. Latent variable 1 primarily affects the relative position of the stagnation point at the blade leading edge and exhibits a quadratic relationship with total pressure loss, while latent variable 2, mainly influenced by blade loading, shows a negative correlation with total pressure loss. The strong correlation between loading distribution and total pressure loss, captured in the two-dimensional latent variables, is a key factor enabling CAM to improve prediction accuracy. On a small-sample dataset with fixed geometry, the CAM model achieves a coefficient of determination (R²) as high as 0.9966. Under variable geometry conditions, the prediction accuracy of the CAM model outperforms the conventional single-source data-driven model by 56.8%.