Research on Radar Signal Recognition Technology Based on Residual Convolutional Neural Network

The research on recognition of radar intra-pulse modulated signals is an important development direction of radar countermeasure technology. In order to recognize the radar intra-pulse modulated signals effectively, many sgnal recognition techniques have been developed. In which, the one that based on residual convolutional neural network is one of the most promising techniques. In this paper, radar signal recognition techniques based on residual convolutional neural network are researched. The influence of model depth and residual block on time-frequency image recognition is verified. Firstly, using the feature extraction and recognition method of radar signal time-frequency image from Choi Williams, the problem of signal recognition is transformed into image recognition. Time-frequency images of 8 kinds of common radar signals are converted to grayscale images by the Choi Williams time-frequency transformation. Secondly, these grayscale images are recognized with six kinds of signal recognition algorithm models (AlexNet, DarkNet-19, GoogLeNet, VGG-16, ResNet-18, MobileNet-v2) and the recognition effect is compared. Thirdly, the residual block of the above six models are modified by increasing or decreasing residual and Inverted residual block, and the experimental results are compared. The results show that the shallow model AlexNet has the best accuracy and speed in recognizing time-frequency images, and the shallow network with an inverted residual block will improve the recognition speed with the cost of reducing the recognition accuracy slightly.