基于DropBlock双模态混合神经网络的无线通信调制识别

doi:10.11959/j.issn.1000-0801.2022099

电信科学 ›› 2022, Vol. 38 ›› Issue (5): 75-86.doi: 10.11959/j.issn.1000-0801.2022099

基于DropBlock双模态混合神经网络的无线通信调制识别

高岩¹, 石坚¹^,², 马圣雨¹^,², 马柏林³, 乐光学²

¹ 河南理工大学计算机科学与技术学院，河南焦作 454003
² 嘉兴学院信息科学与工程学院，浙江嘉兴 314000
³ 嘉兴学院数据科学学院，浙江嘉兴 314000

修回日期:2022-05-15 出版日期:2022-05-20 发布日期:2022-05-01
作者简介:高岩（1963-），男，博士，河南理工大学计算机科学与技术学院教授，主要研究方向为智能信息处理与智能控制等
石坚（1996- ），男，河南理工大学硕士生，主要研究方向为深度学习与智能通信
马圣雨（1997- ），女，河南理工大学硕士生，主要研究方向为深度学习与智能通信
马柏林（1961- ），男，博士，嘉兴学院数据科学学院教授，主要研究方向为调和分析和小波分析、智能计算
乐光学（1963- ），男，博士，嘉兴学院信息科学与工程学院教授，主要研究方向为多云融合与协同服务、边缘计算与一体化通信网络、深度学习与智能通信
基金资助:
国家自然科学基金资助项目(U19B2015)

DropBlock based bimodal hybrid neural network for wireless communication modulation recognition

Yan GAO¹, Jian SHI¹^,², Shengyu MA¹^,², Bolin MA³, Guangxue YUE²

¹ College of Computer Science and Technology, Henan Polytechnic University, Jiaozuo 454003, China
² College of Information Science and Engineering, Jiaxing University, Jiaxing 314000, China
³ College of Data Science, Jiaxing University, Jiaxing 314000, China

Revised:2022-05-15 Online:2022-05-20 Published:2022-05-01
Supported by:
The National Natural Science Foundation of China(U19B2015)

摘要/Abstract

摘要：

自动调制识别作为信号检测和解调的中间步骤，在无线通信系统中起着至关重要的作用。针对现有自动调制识别方法识别精度低的问题，提出了一种双模态混合神经网络（bimodal hybrid neural network， BHNN），该网络利用多个模态中包含的互补增益信息来丰富特征维度。将改进的残差网络与双向门控循环单元并行连接，构建双模态混合神经网络模型，分别提取信号的空间特征与时序特征。引入DropBlock正则化算法，有效抑制网络训练过程中过拟合、梯度消失和梯度爆炸等对识别精度的影响。以双模态数据输入，充分利用信号的空间与时序特征，通过并行连接减少网络深度，加速模型收敛，提高调制信号的识别精度。为验证模型的有效性，采用两种公开数据集对模型进行仿真实验，结果表明，BHNN在两种数据集上识别精度高、稳定性强，在高信噪比下识别精度分别可达89%和93.63%。

关键词: 调制识别, 双模态混合网络, DropBlock正则化, ResNet, BiGRU

Abstract:

As an intermediate step of signal detection and demodulation, automatic modulation recognition played a momentous role in wireless communication system.Aiming at the low recognition accuracy of existing automatic modulation recognition methods, a bimodal hybrid neural network (BHNN) was proposed, which utilized complementary gain information contained in multiple modes to enrich feature dimensions.The improved residual network was connected in parallel with the bidirectional gated loop unit to construct a bimodal hybrid neural network model, and the spatial and temporal features of the signal were extracted respectively.The DropBlock regularization algorithm was introduced to effectively suppress the influence of over fitting, gradient disappearance and gradient explosion on the recognition accuracy in the process of network training.Using bimodal data input, the spatial and temporal characteristics of signals were fully utilized, and the network depth was reduced through parallel connection.The model convergence was accelerated, and the recognition accuracy of modulated signals was improved.In order to verify the effectiveness of the model, two public datasets were used to simulate the model.The results show that BHNN has high recognition accuracy and strong stability on the two datasets, and the recognition accuracy can reach 89% and 93.63% respectively under high signal-to-noise ratio.

Key words: modulation recognition, bimodal hybrid neural network, DropBlock regularization, ResNet, BiGRU

中图分类号:

TP393

高岩, 石坚, 马圣雨, 马柏林, 乐光学. 基于DropBlock双模态混合神经网络的无线通信调制识别[J]. 电信科学, 2022, 38(5): 75-86.

Yan GAO, Jian SHI, Shengyu MA, Bolin MA, Guangxue YUE. DropBlock based bimodal hybrid neural network for wireless communication modulation recognition[J]. Telecommunications Science, 2022, 38(5): 75-86.

图/表 19

图1

表1

表2

图2

图3

表3

BHNN模型隐藏层详细参数"

隐藏层名称	类型	输出维度
Input1	Featureinput	(None,2,128)
Conv	Conv1D,kernel size:1	(None,32,128)
Norm	Batch Normalization	(None,32,128)
Conv1	Conv1D,kernel size:5	(None,32,128)
DropBlock	DropBlock	(None,32,128)
Conv2	Conv1D,kernel size:5	(None,32,128)
DropBlock1	DropBlock	(None,32,128)
Add	Addition[Norm,DropBlock]	(None,32,128)
Conv3	Conv1D,kernel size:5	(None,32,128)
DropBlock2	DropBlock	(None,32,128)
Conv4	Conv1D,kernel size:5	(None,32,128)
DropBlock3	DropBlock	(None,32,128)
Add1	Addition[Add,DropBlock3]	(None,32,128)
Maxpooling	MaxPooling1D	(None,32,64)
Conv5	Conv1D,kernel size:1	(None,32,64)
Norm1	Batch Normalization	(None,32,64)
Conv6	Conv1D,kernel size:5	(None,32,64)
DropBlock4	DropBlock	(None,32,64)
Conv7	Conv1D,kernel size:5	(None,32,64)
DropBlock5	DropBlock	(None,32,64)
Add2	Addition[Norm,DropBlock5]	(None,32,64)
Conv8	Conv1D,kernel size:5	(None,32,64)
DropBlock6	DropBlock	(None,32,64)
Conv9	Conv1D,kernel size:5	(None,32,64)
DropBlock7	DropBlock	(None,32,64)
Add3	Addition[Add2,DropBlock7]	(None,32,64)
Maxpooling1	MaxPooling1D	(None,32,32)
Conv10	Conv1D,kernel size:1	(None,32,32)
Norm2	Batch Normalization	(None,32,32)
Conv11	Conv1D,kernel size:5	(None,32,32)
DropBlock8	DropBlock	(None,32,32)
Conv12	Conv1D,kernel size:5	(None,32,32)
DropBlock9	DropBlock	(None,32,32)
Add4	Addition[Norm2,DropBlock9]	(None,32,32)
Conv13	Conv1D,kernel size:5	(None,32,32)
DropBlock10	DropBlock	(None,32,32)
Conv14	Conv1D,kernel size:5	(None,32,32)
DropBlock11	DropBlock	(None,32,32)
Add5	Addition[Add4,DropBlock11]	(None,32,32)
Maxpooling2	MaxPooling1D	(None,32,16)
Flatten1	Flatten	(None,512)
Input2	Featureinput	(None,128,2)
Bidirectional	BiGRU	(None,128,256)
Flatten2	Flatten	(None,32768)
Concatenate	Concatenate[Flatten1, Flatten2]	(None, 33280)
Dense	Dense,activation:softmax	(None,10)
Reshape	Reshape	(None,10)
Output	Regression output	Regression output

表3

表4

表5

图4

图5

图6

图7

表6

图8

表7

图9

图10

图11

图12

参考文献 21

[1]	WOUTERS J , PATRINOS P , KLOOSTERMAN F ,et al. Multi-pattern recognition through maximization of signal-to-peak-interference ratio with application to neural spike sorting[J]. IEEE Transactions on Signal Processing, 2020(68): 6240-6254.
[2]	LIU F , ZHOU Y , LIU Y A . A deep neural network method for automatic modulation recognition in OFDM with index modulation[C]// Proceedings of 2019 IEEE 89th Vehicular Technology Conference. Piscataway:IEEE Press, 2019: 1-5.
[3]	WANG Y , LIU M , YANG J ,et al. Data-driven deep learning for automatic modulation recognition in cognitive radios[J]. IEEE Transactions on Vehicular Technology, 2019,68(4): 4074-4077.
[4]	TUNZE G B , HUYNH-THE T , LEE J M ,et al. Sparsely connected CNN for efficient automatic modulation recognition[J]. IEEE Transactions on Vehicular Technology, 2020,69(12): 15557-15568.
[5]	LIN C , YAN W J , ZHANG L M ,et al. A real-time modulation recognition system based on software-defined radio and multi-skip residual neural network[J]. IEEE Access, 2020(8): 221235-221245.
[6]	CHEN S Y , ZHANG Y , HE Z W ,et al. A novel attention cooperative framework for automatic modulation recognition[J]. IEEE Access, 2020(8): 15673-15686.
[7]	PIJACKOVA K , GOTTHANS T . Radio modulation classification using deep learning architectures[C]// Proceedings of 2021 31st International Conference Radioelektronika (RADIOELEKTRONIKA). Piscataway:IEEE Press, 2021: 1-5.
[8]	ZHAO X D , ZHOU X H , XIONG J ,et al. Automatic modulation recognition based on multi-dimensional feature extraction[C]// Proceedings of 2020 International Conference on Wireless Communications and Signal Processing (WCSP). Piscataway:IEEE Press, 2020: 823-828.
[9]	赵雄文, 郭春霞, 李景春 . 基于高阶累积量和循环谱的信号调制方式混合识别算法[J]. 电子与信息学报, 2016,38(3): 674-680.
	ZHAO X W , GUO C X , LI J C . Mixed recognition algorithm for signal modulation schemes by high-order cumulants and cyclic spectrum[J]. Journal of Electronics ＆ Information Technology, 2016,38(3): 674-680.
[10]	孙姝君, 彭盛亮, 姚育东 ,等. 基于深度学习的调制识别综述[J]. 电信科学, 2021,37(5): 82-90.
	SUN S J , PENG S L , YAO Y D ,et al. A survey of deep learning based modulation recognition[J]. Telecommunications Science, 2021,37(5): 82-90.
[11]	O’SHEA T J , CORGAN J , CLANCY T C . Convolutional radio modulation recognition networks[M]// Engineering Applications of Neural Networks. Cham: Springer International Publishing, 2016: 213-226.
[12]	O’SHEA T J , ROY T , CLANCY T C . Over-the-air deep learning based radio signal classification[J]. IEEE Journal of Selected Topics in Signal Processing, 2018,12(1): 168-179.
[13]	GHIASI G , LIN T Y , QUOC V . DropBlock:a regularization method for convolutional networks[C]// Proceedings of the 32nd International Conference on Neural Information Processing Systems (NIPS'18). New York:ACM Press, 2018: 10750-10760.
[14]	O’SHEA T J , WEST N . Radio machine learning dataset generation with gnu radio[C]// Proceedings of the GNU Radio Conference. 2016: 1-6.
[15]	CHO K , MERRIENBOER B V , GULCEHRE C ,et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[J]. Computer Science,arXiv:1406.1078v3, 2014.
[16]	QI P H , ZHOU X Y , ZHENG S L ,et al. Automatic modulation classification based on deep residual networks with multimodal information[J]. IEEE Transactions on Cognitive Communications and Networking, 2021,7(1): 21-33.
[17]	RAMJEE S , JU S , YANG D ,et al. Fast deep learning for automatic modulation classification[J]. arXiv:1901.05850, 2019.
[18]	YANG H G , ZHAO L Z , YUE G X ,et al. IRLNet:a short-time and robust architecture for automatic modulation recognition[J]. IEEE Access, 2021,9(10): 143661-143676.
[19]	SRIVASTAVA N , HINTON G , KRIZHEVSKY A ,et al. Dropout:a simple way to prevent neural networks from overfitting[J]. Journal of Machine Learning Research, 2014,15(1): 1929-1958.
[20]	TOMPSON J , GOROSHIN R , JAIN A ,et al. Efficient object localization using Convolutional Networks[C]// Proceedings of 2015 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE Press, 2015: 648-656.
[21]	LU M , PENG T J , YUE G X ,et al. Dual-channel hybrid neural network for modulation recognition[J]. IEEE Access, 2021(9): 76260-76269.

层	输出维度
输入	2×128
残差块	32×128
残差块	32×64
残差块	32×32
输出1	512

参数	说明
调制类型	数字调制：BPSK、QPSK、8PSK、16QAM、64QAM、CPFSK、PAM4、GFSK
	模拟调制：AM-DSB、WBFM、AM-SSB
数据格式	IQ
数据维度	2×128
采样频率	1 MHz
信噪比范围	[-20 dB，-18 dB，…，16 dB，18 dB]
总样本量	220 000
训练集样本量	154 000
测试集样本量	66 000

参数	说明
调制类型	数字调制：BPSK、QPSK、8PSK、16QAM、64QAM、CPFSK、PAM4、GFSK
	模拟调制：AM-DSB、WBFM
数据格式	IQ
数据维度	2×128
采样频率	1 MHz
信噪比范围	[-20 dB，-18 dB，…，16 dB，18 dB]
总样本量	1 200 000
训练集样本量	840 000
测试集样本量	360 000

算法	正则化因子	正则化惩罚项
L1	l1=0.01	loss = l1×reduce_sum(abs(x))
L2	l2=0.01	loss = l2×reduce_sum(square(x))
L1L2	l1=0.01	loss = l1×reduce_sum(abs(x))
	l2=0.01	loss = l2×reduce_sum(square(x))
L1_L2	l1=0.01	loss = l1×reduce_sum(abs(x))
	l2=0.01	loss = l2×reduce_sum(square(x))

真实情况	预测结果
真实情况	正例	反例
正例	TP（真正例）	FN（假反例）
反例	FP（假正例）	TN（真反例）

基于DropBlock双模态混合神经网络的无线通信调制识别

DropBlock based bimodal hybrid neural network for wireless communication modulation recognition

在线阅读

PDF下载

可视化

摘要/Abstract

引用本文

使用本文

图/表 19

参考文献 21

相关文章 3

Metrics

推荐阅读 0

[1]	孙姝君, 彭盛亮, 姚育东, 杨喜. 基于深度学习的调制识别综述[J]. 电信科学, 2021, 37(5): 82-90.
[2]	王闯，张更新，谢智东，胡婧. 雨衰环境下ACM过程的动态调制识别[J]. 电信科学, 2015, 31(6): 2015155-.
[3]	战金龙,郭永明,卢建军. GLSFBC-OFDM信号调制识别算法 ^*[J]. 电信科学, 2012, 28(6): 63-67.

层	输出维度
输入	128×2
BiGRU	128×256
输出2	32768