ABSTRACT A two‐stage learning framework based on a long‐range structure‐enhanced masked autoencoder is developed for few‐shot radar specific emitter identification (SEI). To improve the modeling of long‐range fingerprint structures and the discriminative stability of embeddings under limited labeled samples, a self‐supervised masked reconstruction task is first constructed for 1D axially integrated bispectrum (AIB) sequences to learn structural priors from unlabeled data. A ResLKA‐TS encoder, consisting of a residual backbone, large‐kernel attention, and soft‐threshold shrinkage, is then employed to enhance the representation of distributed fingerprint structures. In the few‐shot fine‐tuning stage, center loss is further introduced to improve intra‐class compactness and inter‐class separability in the embedding space. Experiments on a measured dataset collected from eight ADALM‐PLUTO devices of the same model show that, at an SNR of 20 dB, identification accuracies of 88.00% and 91.87% are achieved under 10‐shot and 15‐shot settings, respectively, exceeding those of asymmetric masked autoencoder (AMAE) by 5.12 and 4.69 percentage points. The method also maintains advantages under low‐shot and different SNR conditions.
Huang et al. (Thu,) studied this question.