Neural network adaptive methods excel at managing dynamical uncertainties and are widely used in hypersonic vehicle control. However, traditional approaches are limited by their dependence on restricted online information for training, which hampers learning efficiency and control performance. This paper presents a neural network adaptive control method enhanced by meta-learning-based feature extraction. It initially categorizes unknown dynamics into two types through offline meta-learning: deterministic features, consistent across flights, and non-deterministic features, influenced by stochastic elements like atmospheric density variations. These features are integrated into an online adaptive controller, with deterministic features processed by an offline-trained neural network and non-deterministic features identified by linear parameters. This framework employs a composite adaptive approach using least squares modulation to recognize non-deterministic features. Theoretical analysis and simulations show that incorporating deterministic features reduces the complexity of online learning of unknown dynamics, improving both learning efficiency and control performance with limited online information.