A rapid recognition method for radar active jamming based on the Hybrid 3_CNN-Transformer model

As the electromagnetic environment becomes increasingly complex, radar encounters more intricate jamming patterns. Accurate and real-time jamming recognition is crucial for effective radar anti-jamming decisions. For scenarios with limited data, a rapid active jamming recognition method based on the Hybrid 3_CNN-Transformer model is proposed. To thoroughly evaluate and validate this method, we simulated nine jamming types and used pulse compression for preprocessing. Combining the strengths of CNNs and Transformers, we constructed the Hybrid 3_CNN-Transformer model, which employs three CNNs to extract local features from the complex domain, real part, and imaginary part of the jamming signals. After concatenating these features and performing position encoding, the model utilises Transformers to capture global features, enhancing recognition accuracy and reducing training time. To enhance computational efficiency and reduce storage, we applied an L1 norm-based unstructured pruning algorithm for model compression, achieving an 82% pruning rate and cutting inference time to 33 ms. Experiments show that the Hybrid 3_CNN-Transformer model significantly boosts recognition accuracy and speed over other models. On a small dataset with nine jamming types, it achieved 95.1% accuracy after 25 epochs and 100% after 90 epochs. This approach enhances training speed and accuracy, allowing rapid and reliable jamming recognition in resource-limited environments.