Abstract

Classification of radar signals has been one of the indispensable parts in today’s surveillance, electronic warfare, remote sensing, target recognition, and intelligent sensing. A reliable signal classification makes radar receiver detect signal categories as well as find out patterns in various conditions. The classical machine learning models often rely on manually engineered features or operate in rather static environments. In this study, we propose Adaptive Multi-Scale Noise Robust Radar Classification Network (AMNR-RCN). It uses adaptive signal pre-processing, multi-scale time-frequency feature extraction, channel-wise attention, and confidence-guided ensemble optimization. At first, raw radar signals are normalized in terms of amplitude and pre-processed by using noise information. Attention module highlights discriminative features whereas confidence-based fusion module fuses the predictions of different classifiers based on their confidence scores. When evaluating our AMNR-RCN on different SNR values between -15 to 17 dB, the model gets 74.6% accuracy, 73.9% macro-precision, 73.2% macro-recall, 73.5% macro-F1, and 84.6% macro-AUC at -15 dB. At 17 dB, the results are 98.6% accuracy, 98.4% macro-precision, 98.2% macro-recall, 98.3% macro-F1, and 99.0% macro-AUC.

Authors

G. Vidhya
Vellore Institute of Technology, Vellore, India

Keywords

Radar Signal Classification, Noise Robustness, Multi-Scale Feature Learning, Attention Mechanism, Confidence-Guided Ensemble