基于稠密残差注意力网络的调制方式识别

Modulation Recognition Based on Dense Residual Attention Network

摘要:

针对低信噪比和复杂信道条件下通信信号调制方式识别特征提取准确性难以保证及深度残差收缩网络特征学习能力不足的问题，提出基于多特征的稠密残差注意力网络。通过提取I/Q信号的单路与两路联合特征，利用多特征稠密瓶颈层替代传统卷积层构造DRSM，实现深浅层特征的高效融合与抗噪增强。结合混合注意力机制建立跨维度特征强化机制，提升关键信号特征的捕获精度。基于RadioML2016.10A数据集的实验表明，该方法在信噪比大于4dB时11类信号总体识别准确率较传统深度残差收缩网络提高约5%，验证了所提方法在复杂信道条件下的有效性与鲁棒性。此外，针对传统模型在16QAM和64QAM信号识别中的不足，稠密残差注意力网络表现出较好的识别能力。

Abstract:

To address the challenges of ensuring feature extraction accuracy for communication signal modulation recognition under low signal-to-noise ratio (SNR) and complex channel conditions, as well as the insufficient feature learning capability of deep residual shrinkage networks, this study proposes a multi-feature dense residual attention network. By extracting single-branch and dual-branch joint features from I/Q signals, the method replaces traditional convolutional layers with multi-feature dense bottleneck layers to construct a Dense Residual Shrinkage Module (DRSM). This achieves efficient fusion of shallow and deep-layer features while enhancing noise resistance. A hybrid attention mechanism is integrated to establish a cross-dimensional feature enhancement mechanism, improving the capture precision of critical signal characteristics. Experiments on the RadioML2016.10A dataset demonstrate that the proposed method achieves approximately 5% higher overall recognition accuracy for 11 modulation types compared to traditional deep residual shrinkage networks when SNR exceeds 4 dB, validating its effectiveness and robustness in complex channel environments. Furthermore, the dense residual attention network exhibits superior recognition capability for 16QAM and 64QAM signals, addressing the limitations of conventional models in distinguishing these modulation schemes.