Radar signal deinterleaving is used to separate interleaved pulse streams in the electronic support measure (ESM) systems. The histogram methods based on the difference in time of arrival (TOA) are more mature in engineering applications than other methods because of their more effortless implementation. In the increasingly complex electronic battlefield, pulse repetition interval (PRI) jitter, and pulse missing are inevitable in the interleaved pulse stream. The TOA differences of pulse trains with PRI jitter distributes in multiple adjacent PRI bins, which results in the significant reduction of the corresponding histogram value. Furthermore, the histogram values of real PRIs vary with pulse missing, which indicates that the histogram lacks robustness to pulse missing. First-order difference curve based on sorted TOA difference sequence, denoted as FDC-DTOA, is proposed in this paper to overcome the shortcomings mentioned above. In this method, multi-order TOA differences are calculated and then sorted in ascending order. Then, FDC-DTOA is obtained after a first-order backward difference operation. Finally, an adaptive threshold is applied to FDC-DTOA to extract candidate PRIs. The simulation results show that the proposed method has remarkable robustness to pulse missing and excellent performance in deinterleaving the pulse stream with PRI jitter.

No abstract

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.

No abstract