关键词: 多任务学习, 调制识别, 深度神经网络, 联合损失, 不确定性加权

Abstract:

To address the challenge of degraded radar modulation recognition performance under low signal-to-noise ratio （SNR） conditions, a denoising-classification learning method is proposed based on multi-task collaborative optimization. The method constructs a joint framework integrating a denoising sub-network and a classification sub-network, facilitating deep coupling between time-frequency feature denoising and signal classification. The denoising sub-network employs a deep autoencoder to learn the intrinsic features of signals in the time-frequency domain, suppressing noise interference via a hybrid loss function. The classification sub-network utilizes an attention-based residual network to achieve high-robustness recognition. A collaborative optimization mechanism, employing an uncertainty-weighted joint loss function, promotes the adaptive interaction and balance between denoised features and classification features. Simulation results demonstrate that the proposed method significantly enhances the recognition accuracy for 12 types of radar modulations at low SNRs ranging from −20 dB to −8 dB, validating the effectiveness of the multi-task collaborative optimization strategy.

Key words: multi-task learning, modulation recognition, deep neural network, joint loss, uncertainty weighting