近年来,深度学习(Deep Learning,DL)在调制识别算法上的研究表明了其较传统方法的优越性.通过提取信号的空间或时间特征进行分类通常具有较高准确率,然而针对高阶调制的分类精度却存在一定偏差.为此,研究并实现了一种基于时空联合特征注意力机制的调制识别算法.该算法在提取信号时空联合特征的基础上加入了一种自注意力机制,有效提升了高阶调制的识别准确率,具有很高的分类精度.同时,基于GNU Radio开发框架实现,该算法能够部署于各类软件无线电嵌入式平台,具有通用性和实用性.利用GNU Radio Out of Tree(OOT)模块与开放神经网络交换(Open Neural Network Exchange,ONNX),将模型部署在嵌入式设备中进行DL推理.测试表明该算法在软件定义无线电(Software Defined Radio,SDR)平台下拥有较高识别准确率和推理速度,对高阶调制信号的分类精度高,同时在嵌入式环境下的推理与训练过程具有一致性.
Research and Implementation of Modulation Recognition Based on Spatiotemporal Joint Feature Combined with Attention Mechanism
In recent years,research on Deep Learning(DL)based modulation recognition has shown its superiority over traditional methods.Classification by extracting spatial or temporal features of signals usually has a high accuracy,but there is certain deviation in the classification for high-order modulation.Therefore,a modulation recognition algorithm based on spatiotemporal joint feature combined with attention mechanism is researched and implemented.This algorithm adds a self-attention mechanism on the basis of extracting the spatiotemporal joint features of the signal,effectively improving the recognition accuracy of high-order modulation,and has high classification accuracy.At the same time,based on the GNU Radio development framework,this algorithm can be deployed on various software radio embedded platforms,with universality and practicality.Using the GNU Radio Out of Tree(OOT)module and the Open Neural Network Exchange(ONNX),the model is deployed on embedded devices for deep learning inference.Tests show that the algorithm has high recognition accuracy and reasoning speed on the Software Defined Radio(SDR)platform,and high classification accuracy for high-order modulation signals.At the same time,the reasoning result in the embedded environment is consistent with that in the training process.
万方数据
维普
