Semantic-aware experience replay for continual air traffic control instruction intent recognition

The intent interpretation of controller-pilot communications is crucial to robust Air Traffic Control (ATC) operations. As spatio-temporal flight environments dynamically evolve, conventional intent recognition models suffer from catastrophic forgetting when deployed on streaming data. Existing continual learning techniques like experience replay offer some mitigation. However, these approaches treat historical experience with equal significance, which is inefficient for knowledge retention and adaptation in fine-grained semantic scenes. This paper proposes a semantic-aware experience replay framework for continual intent recognition to prioritize critical knowledge transitions. Specifically, we first introduce a prototype-based buffer that selects and stores representative samples within the feature space, thereby effectively preserving historical data distributions. Subsequently, a curriculum-based selection strategy is designed to schedule rehearsal samples in an easy-to-hard manner, ensuring that the model first consolidates reliable foundational knowledge. This process is further augmented with knowledge distillation to enforce representation consistency between successive models. Finally, comprehensive experiments on regular and emergency-involved ATC instruction datasets demonstrate that the proposed method achieves superior performance compared with the state-of-the-art baselines. The method in this work indicates the potential for safety–critical control operation applications.