High-Efficiency Transformer-Based Network for Radar Interference Recognition

Abstract

Interference recognition and adaptively optimizing transmitted signal are two fundamental functions of cognitive radar system. Due to the changing radar transmission waveform, the interference signal from the jammer changes accordingly. Traditional deep convolutional neural network (DCNN)-based interference recognition methods need to increase the network model scale largely since the number of interference signals is expanded significantly. In this article, a novel transformer-based network (TBNet) structure is proposed. First, instead of the time- frequency transform (TFT)-based preprocessing of the received signal used in DCNN, $1 \times 1$ convolution is directly used to perform data preprocession after rearrangement and combination of the original time series signal. Second, stacked transformer blocks are adopted to carry out feature extraction, avoiding the conventional multiple convolution layers. Our proposed TBNet can significantly decrease parameter amount and inference time without affecting recognition performance. Results show that TBNet has 50x less parameter amount and 20x less inference time than benchmark models. Additionally, the proposed method has an average $4.52\%$ higher recognition accuracy under our parameter settings.