基于脉冲积分的CNN分类性能提升方法用于低截获概率雷达信号识别

In electronic warfare, technologies that accurately detect the adversary's location and signal characteristics to execute precise strikes or effectively jam signals are crucial assets. To meet these capabilities, low probability of intercept (LPI) radars are employed, which transmit low-power signals and utilize various modulation techniques to minimize the probability of signal detection. Previous studies presented an approach for classifying LPI radar waveforms using a convolutional neural network (CNN) with time-frequency images (TFIs). Even though they have shown higher identification performance than that of the conventional fast Fourier transform (FFT) techniques, they did not consider the inherent characteristics of radar systems, such as repetitive pulse transmissions and the interceptor's limited knowledge of the start of pulses. Motivated by the limitations, this paper proposes a framework that converts multiple intercepted signals into images and extracts multiple features through a CNN to improve the classification performance with a more realistic LPI dataset constructed under timing errors. To validate the performance of the proposed framework, we consider various CNNs with different combinations of feature extraction methods and backbone networks. We provide a quantitative analysis of the advantage of pulse integration. Furthermore, the experimental results clearly demonstrated that the proposed framework achieved a 18.2% accuracy improvement at $-10$ dB and a 26% improvement at $-12$ dB, compared to the conventional scheme without pulse integration.