Abstract: In the field of electronic countermeasures, a situation was faced where there is a large volume of data with few samples, and radar and communication signals are modeled separately. This has led to fragmentation in capabilities and low data utilization rates. To address this issue, an unsupervised electromagnetic pre-training model scheme was designed based on the CNN-transformer as the main network structure. The massive amount of unlabeled electromagnetic data collected through reconnaissance was fully leveraged to learn the distribution of signal features, preventing the model from overfitting on a small number of specific samples. As a result, data utilization and the generalization ability of the recognition model are improved. Furthermore, considering the multi-source data scenario in reconnaissance data, which includes both real-numbered radar data and complex-numbered communication data, the CNN-transformer-based pre-training model was innovatively transformed into a complex number representation. A complex CNN-transformer-based electromagnetic pre-training model was designed,achieving integrated joint modeling of radar and communication data. Additionally, by adopting the technical route of complex CNN-transformer pre-training model + backend fine-tuning, multiple electromagnetic perception and recognition tasks were successfully modeled, including radar individual identification, modulation style recognition, and communication individual identification. This effectively resolves the issue of capability fragmentation. Experimental results show that compared to deep learning models based on real-number networks, On the non-homologous test set, the recognition accuracy was improved by an average of more than 10 percentage points, greatly enhancing the model's accuracy, generalizability, and uniformity.

Key words: Electromagnetic Perception, Radar-Communication Integration, Artificial Intelligence, Unsupervised Pretraining

摘要： 目前在电子对抗领域面临大数据、小样本的数据现状以及雷达与通信信号分别独立建模的研究现状，存在能力碎片化、数据利用率低的问题。为解决此问题，本文设计了一套以CNN-transformer为主网络结构的无监督电磁预训练模型方案，充分利用侦察到的海量未标注电磁数据来学习信号特征分布，避免模型在少量特定样本上过拟合，从而提高数据利用率以及提高识别模型的泛化性；然后针对侦察数据中既有以实数信号为主的雷达数据又有以复数信号为主的通信数据的多源数据场景，本文将以CNN-transformer为主结构的电磁预训练模型创新性地进行了复数化改造，设计了一套基于复数CNN-transformer的电磁感知预训练模型，实现了雷达和通信数据的一体化联合建模，同时基于复数CNN-transformer预训练模型+后端微调的技术路线实现了多个电磁感知识别任务的建模，包括雷达个体识别、调制样式识别和通信个体识别等，解决了能力碎片化的问题。实验结果表明，相比基于实数网络的深度学习模型，在非同源测试集上，本方案识别准确率平均提升10个百分点以上，极大提高了模型的准确性、泛化性和统一性。

关键词: 电磁感知, 雷通一体化, 人工智能, 无监督预训练