基于RIS的元素分组面状全连接网络
Element Grouping Faceted Fully Connected Network Based on RIS
侯顺虎 1方胜良 2曾庆尧 1王孟涛1
作者信息
- 1. 航天工程大学 研究生院,北京 101416
- 2. 航天工程大学 航天信息学院,北京 101416
- 折叠
摘要
针对神经网络全连接层在训练中参数量多、所占内存多、易产生过拟合问题,从智能超表面(reconfigurable intelligence surface,RIS)结构特征出发,提出了一种基于RIS的元素分组面状全连接神经网络(RIS-based element grouping areal fully connected neural network,RGFCNN).借鉴RIS的结构特征,在传统全连接神经网络上进行优化.设计了透射面注意力机制用于数据有效特征提取,相比于传统的全连接网络,该网络没有对数据进行一维排列,而是提出了一种运用于神经网络构建的元素分组策略,直接对二维面状数据进行分组全连接处理,各组处理输出进行数据串联.实验结果表明:在公开的具有IQ数据特征的通信信号数据集上,RGFCNN在信噪比大于0 dB时具有更好的识别精度,而训练参数是原来的大约1/6.
Abstract
In view of the over-fitting problem that caused by multiple parameters and high memory usage of the full connection layer of neural network in training,a RIS-based element grouping areal fully connected neural network(RGFCNN)is proposed for the first time based on the structural characteristics of reconfigurable intelligence surface(RIS).Based on the structural characteristics of RIS,the network is optimized on traditional FCNN.A novel transmission surface attention mechanism is designed for the effective feature extraction of data.Compared with the traditional FCNNs,the proposed network does not arrange the data in one-dimensional manner.Instead,a element grouping strategy is proposed for the neural network construction,which directly groups the two-dimensional surface data,carries out the fully connected processing on each group,and concatenates the output of each group.The experimental results show that,on the public available communication signal datasets with IQ data features,RGFCNN has better recognition accuracy when SNR is greater than 0 dB,and the training parameters are approximately 1/6 of the original.
关键词
智能超表面/全连接神经网络/元素分组策略/IQ信号/调制识别Key words
reconfigurable intelligence surface(RIS)/fully-connected neural network(FCNN)/element grouping strategy/IQ signal/modulation recognition引用本文复制引用
出版年
2024
NETL
NSTL
万方数据