Aerial Target Intention Recognition Based on Incremental Learning

ABSTRACT In aerial target intention recognition, deep learning models face a critical challenge known as catastrophic forgetting—the tendency to lose the ability to recognise historical intentions when learning new ones dynamically. To address this, we propose RSRKD, an incremental learning framework that integrates representative sample replay and knowledge distillation, employing Mamba as the backbone. Our method features a dual mechanism for knowledge preservation and transfer. First, to achieve explicit knowledge preservation, a sample selection mechanism integrating gradient norm and feature space distribution is devised, ensuring the identification and retention of the most informative samples from historical classes throughout incremental training. Second, a joint loss function incorporating both classification and distillation losses is employed to constrain output consistency and promote implicit knowledge transfer between historical and updated models. Experimental results demonstrate the effectiveness of RSRKD: it maintains a high recognition accuracy of 99.25% whilst limiting the forgetting rate of historical data to just 1.22%. Compared to traditional joint training, RSRKD reduces training time by 50.28% and storage space by 93.32%, validating its efficiency in dynamic air combat scenarios.