Few-shot Incremental Identification of Specific Emitter Based on Ancillary Prototypes

This paper proposes a recognition method based on ancillary prototypes to address the incremental identification of specific emitter under few-shot conditions. Firstly, we design a preprocessing method based on adaptive wavelet decomposition to process all received signals to reduce the impact of signal propagation environment. Then, we conduct conventional classification training on known class and save the resulting model. Next, based on the known class model, we perform ancillary prototypes task combining new tasks and previous tasks to increase the distance of each class in the feature embedding space, enhancing the model's fitting ability to few-shot. During this period, we design a sliding window slicing data augmentation method to alleviate the impact of overfitting caused by few-shot, and introduce a carefully designed data storage strategy to reduce the model's computational cost. Finally, we test all classes appearing in the prototype training task. We validate our algorithm on self-collected datasets and publicly available datasets, achieving recognition rates of 96.98% and 95.34% respectively at the maximum classes. In noise immunity testing and complexity comparison, our algorithm outperforms other state-of-the-art incremental recognition algorithms. Furthermore, we validate the rationality of our algorithm through ablation experiment.