| 摘要: |
| 深度学习(Deep Learning，DL是实现自动调制识别(Automatic Modulation Identification，AMI技术的有效方法，但通常难以同时兼顾识别精度和效率。为此，提出了一种增强多尺度特征融合的轻量化AMI方法。首先，设计了轻量化的多尺度特征融合模块，通过跨尺度卷积结构高效提取调制信号的多尺度特征，以增强模型对不同信号特征的表征能力；其次，构建了自适应特征增强模块，结合深度可分离卷积与注意机制，能够自适应地学习关键特征的通道权重，突出重要信号特征的同时减少无关特征的干扰；最后，设计了差异平衡分类器，通过聚焦细微调制模式的识别，从而实现高效分类。实验结果表明，所提方法在识别精度上平均提高了5.91%，参数量减少约8.5×105，单次迭代时间缩短0.062 4 s，与对比的先进模型相比具备更高的精度、更快的速度和更少的参数量。 |
| 关键词: 自动调制识别 深度学习 多尺度特征融合 |
| DOI：10.20079/j.issn.1001-893x.240613002 |
|
| 基金项目:国家自然科学基金区域创新发展联合基金重点资助项目（U21A20114 |
|
| Lightweight Modulated Signal Recognition Based on Enhanced Multi-scale Feature Fusion |
| WU Changcheng,SUN Xiaochuan,YU Jike,LI Yingqi |
| (College of Artificial Intelligence,North China University of Science and Technology,Tangshan 063210,China) |
| Abstract: |
| Deep learning(DL is an effective method for achieving automatic modulation identification(AMI technology.However,DL methods generally struggle to balance recognition accuracy and efficiency simultaneously.To address this,a lightweight AMI method based on enhanced multi-scale feature fusion is proposed.First,a lightweight multi-scale feature fusion module is designed,which efficiently extracts multi-scale features of modulation signals through a cross-scale convolutional structure,enhancing the ﹎odel搒 ability to represent different signal features.Next,an adaptive feature enhancement module is constructed,combining depthwise separable convolution and attention mechanisms to adaptively learn channel weights of key features,highlighting important signal features while reducing interference from irrelevant ones.Finally,a differential balance classifier is designed to focus on recognizing subtle modulation patterns,enabling efficient classification.Experimental results show that the proposed method improves recognition accuracy by an average of 5.91%,reduces the number of parameters by approximately 8.5×105,and decreases iteration time per sample by 0.062 4 seconds.Compared with the advanced models,it achieves higher accuracy,faster speed,and fewer parameters. |
| Key words: automatic modulation identification deep learning multi-scale feature fusion |
