Prototype-based method for incremental radar emitter identification

With the widespread use of radars, different types of radar emitters are being used in the real electromagnetic environment. Radar emitter identification (REI) is an important technique in spectrum management. Methods based on deep learning have been successful in REI. However, they are difficult to be updated with signals from newly available categories. Additionally, two issues that must be considered in a real REI task. First, signals from new unknown radar emitters may appear in the identification stage. Hence the model used in REI must have an open-set recognition capability. Second, obtaining numerous labelled samples in time are difficult in the incremental stage of the model. Thus, it is important to keep the model performance stable under conditions of small samples. To solve these problems, a one-dimensional neural network is designed combined with ArcFace loss in the initial training stage to reserve more embedding space for future new classes, thereby facilitating the update of the model and open-set recognition. An unbiased cosine similarity classifier was adopted, and the historical categories were memorised by their prototypes. When a new category is added, the prototype was calculated and the classifier weights were updated. The proposed method can identify unknown classes and add them to the model when given new labels, thereby resulting in an increase of the model's identifiable types. Extensive experiments were performed. The results show that the proposed model is highly efficient and requires small sample sizes, thereby making it suitable for REI.