基于注意力机制的门控密集卷积网络调制识别算法

A MODULATION RECOGNITION ALGORITHM BASED ON ATTENTION MECHANISM AND GATED DENSE CONVOLUTIONAL NETWORKS

自动调制识别(AMR)是非合作通信系统中的重要组成部分,也是一个通信领域的研究难点。针对该难点,利用深度学习,将密集卷积网络(DenseNet)、门控循环单元(GRU)和注意力机制(Attention)三者结合,提出一种基于注意力机制的门控密集卷积网络(AGDCN)的调制识别算法。该算法提取了信号的空间特征和时序特征,将两者相结合解决了信号识别率低的问题。同时,在网络中加入注意力机制,对GRU训练过程进行权重的自适应调整,有效地加强关键特征的学习。通过实验验证了AGDCN模型性能优于其他神经网络算法,在信噪比超过2 dB时,对11种调制类型的识别率可以达到90%。

Automatic modulation recognition(AMR)is an important part of non cooperative communication system,and it is also a research difficulty in the field of communication.In order to solve the above problems,this paper proposes an attention based gated dense convolutional network(AGDCN)modulation recognition algorithm by combining dense convolutional network(DenseNet),gated recurrent unit(GRU)and attention mechanism.The algorithm extracted the spatial and temporal features of the signal,and combined them to solve the problem of low recognition rate.Attention mechanism was added to the network to adaptively adjust the weight of GRU training process and effectively strengthen the learning of key features.Experiments show that the proposed AGDCN model outperforms other mainstream neural network algorithms.Specifically,when SNR exceeds 2 dB,the recognition rate of 11 modulation types can reach 90%.