A Pseudolabel Method With Semantic Drift for Specific Emitter Identification

Specific emitter identification (SEI) is a critical area of research in the study of electronic countermeasures that focuses on extracting subtle radar features to determine carrier identity attributes. However, in complex scenarios, semantic drift between emitter samples significantly impacts identification performance in the target domain. Therefore, an algorithm for SEI is proposed under conditions of semantic drift utilizing a dual-feature approach involving bispectral slices and an attention mechanism to extract radar fingerprint features. The proposed algorithm trains the network with source-domain samples to obtain the identification model and then predicts target-domain samples to obtain class probabilities. Considering the interference of pseudolabel noise, the proposed algorithm innovatively employs a multiple pseudolabel serial filtering mechanism (MPF) to filter labels. This includes a positive–negative pseudolabeling strategy for the target domain, label uncertainty prediction, and hard sample filtering. Finally, a positive–negative learning process is performed by combining source-domain data with MPF-filtered pseudolabeled data. Through multiple iterations, the algorithm achieves SEI under the condition of semantic drift. The experimental results demonstrate that the proposed algorithm exhibited stable performance and improved identification accuracy from approximately 50% to around 90%. Thus, it meets the requirements of practical scenarios.