Residual Attention-Aided U-Net GAN and Multi-Instance Multilabel Classifier for Automatic Waveform Recognition of Overlapping LPI Radar Signals

Abstract

Automatic waveform recognition of overlapping low probability of intercept (LPI) radar signals is an important and challenging task in electronic reconnaissance of the increasingly complicated spectrum environment. In this article, an overlapping LPI waveform recognition processing framework incorporating residual attention-aided U-net generative adversarial network (GAN) and multiinstance multilabel (MIML) classifier is proposed. This framework includes five cascade modules and can achieve satisfactory recognition performance by training with single type of signals only. First, the training signals are transformed into time–frequency images. Then, a residual attention U-net GAN (RAUGAN) with residual learning is employed to reconstruct signal images from noise-contaminated images and with the supervision of the high-quality ones. After that, an instance generation module with asymmetric convolutions generates instance representations, which are then fed into the subsequent residual attention MIML classifier (RAMIML). Finally, an adaptive threshold calibration module is implemented to obtain appropriate thresholds for the final decision. Besides, two loss functions are elaborately designed for the RAUGAN and RAMIML, respectively. Extensive experimental results validate the recognition performance of the proposed framework with a recognition accuracy of 80$\%$ at signal-to-noise ratio $>-18$ dB and show higher robustness on the power ratios and generative performance compared with other state-of-the-art methods.