An intelligent discrimination scheme for radar compound jamming signals built upon a deep residual framework

Abstract

Electronic countermeasure technology continues to advance rapidly, and the electromagnetic conditions confronting radar receivers have grown correspondingly intricate. A particularly difficult problem for present-day radar recognition and anti-jamming studies is compound jamming, in which suppression jamming and deception jamming act at the same time. Under such jamming, recognition schemes that rely on hand-crafted features tend to represent the signal poorly, resist noise weakly, and produce accuracy that fluctuates, so they fall short of what intelligent recognition demands in realistic combat conditions. This paper therefore presents a Deep Residual Network (ResNet) scheme for the intelligent identification of radar compound jamming. The procedure has three steps: the received one-dimensional compound jamming signal is first mapped to a two-dimensional time-frequency image through the Short-Time Fourier Transform (STFT), which makes the joint time-frequency characteristics fully accessible; an improved deep residual architecture is then built, where residual connections relieve the gradient vanishing that troubles deep networks and a batch normalization layer is placed in every residual block to keep training stable; finally, a Softmax classifier assigns each sample to one of twelve representative radar compound jamming categories. The simulations indicate that, at a Jamming-to-Noise Ratio (JNR) of 0 dB, the method recognizes signals with an overall accuracy of 95.273%, and even at an extremely low JNR of −10 dB it still reaches 82.461%—markedly ahead of support vector machine, ordinary convolutional neural network, and VGG-16 baselines. These outcomes confirm that the approach is both effective and robust for recognizing radar compound jamming in complicated electromagnetic surroundings.