Real-time monitoring of aircraft wing surface pressure is vital for aerodynamic validation and structural health assessment. A major challenge in this field is the accurate reconstruction of high-resolution flow structures using only sparse experimental measurements. Existing methodologies often struggle to balance the computational speed required for real-time applications with the precision needed to capture complex non-linear pressure gradients. Furthermore, the reliance on high-fidelity simulation data for training creates a significant computational bottleneck, while traditional black-box models often lack the interpretability required for critical engineering decisions. This paper introduces a multi-fidelity digital twin framework that integrates Variational Autoencoders (VAE) and Kolmogorov-Arnold Networks (KAN). The proposed method leverages a three-tier data hierarchy: low-fidelity potential flow solutions (L_1), medium-fidelity RANS simulations (L_2), and high-fidelity sparse experimental data (L_3). Instead of traditional linear decomposition techniques, we employ a VAE to extract a non-linear latent representation of the pressure field, ensuring that intricate flow features are preserved in a compressed format. A KAN is then utilized to map airflow conditions, such as Mach number and angle of attack, to this latent space. Unlike standard networks with fixed activation functions, KANs utilize learnable B-splines on edges, offering enhanced approximation accuracy and symbolic transparency. The framework operates through an offline multi-fidelity training phase and an online stage where the model provides instantaneous pressure field predictions. Furthermore, a residual fine-tuning process allows the digital twin to adapt to real-time sensor data by correcting the predicted fields based on physical measurements. Validation on the ONERA M6 and CRM-NFL wings dataset demonstrates that the VAE-KAN approach delivers superior reconstruction performance and robust generalization across varying flight conditions. By incorporating variable fidelity data sources, the framework achieves a high-precision digital twin while significantly reducing the computational burden of generating large-scale simulation samples.