信息工程大学学报2025，Vol.26Issue(6)：652-659,8.DOI:10.3969/j.issn.1671-0673.2025.06.004
改进ResNet-CNN-BiLSTM混合模型的信号调制识别方法
Modulation Recognition Method Based on Improved ResNet-CNN-BiLSTM Hybrid Model
TAN Xiongwei 1ZHANG Hongmin 1CAO Lei 1LIU Xiangyu 1MA Shuang2
作者信息
- 1. Information Engineering University,Zhengzhou 450001,China
- 2. Unit 77526,Lhasa 850032,China
- 折叠
摘要
Abstract
Aiming at the poor robustness of traditional modulation recognition methods in complex electromagnetic environments,as well as the high computational complexity and slow training conver-gence of the existing ResNet-CNN-BiLSTM model,an improved model with multi-dimensional collab-orative optimization is proposed to achieve the coordination of high accuracy,lightweight design and real-time performance for signal modulation recognition.Multi-scale kernels and depthwise separable convolution are used to adapt to time-frequency features,adaptive temporal pooling is adopted to en-sure the integrity of temporal features,and input downsampling combined with gating merging is ap-plied to simplify the BiLSTM structure.Hierarchical regularization and SNR-aware dynamic learning rate are introduced to enhance noise robustness,and the multi-dimensional confidence calculation method is improved to boost the reliability of decision-making.Experimental results show that the opti-mized model achieves a validation accuracy of 98.17%,reduces the parameter count to 1.5×106(a 51.6%reduction),shortens the training time to 48 minutes and 20 seconds,and achieves an inference speed of 8 ms per sample.The misjudgment rate in low-SNR environments is reduced by 4.2%～6.8%,providing an efficient solution for computing-constrained scenarios.关键词
混合模型/多维度协同优化/自适应时序池化/信号调制识别Key words
hybrid model/multi-dimensional collaborative optimization/adaptive temporal pooling/signal modulation recognition分类
信息技术与安全科学引用本文复制引用
TAN Xiongwei,ZHANG Hongmin,CAO Lei,LIU Xiangyu,MA Shuang..改进ResNet-CNN-BiLSTM混合模型的信号调制识别方法[J].信息工程大学学报,2025,26(6):652-659,8.
