基于轻量神经网络的无线电调制识别算法

A Radio Modulation Recognition Algorithm Based on Lightweight Neural Network

在信号环境日益复杂、信号调制样式种类多变的情况下,采用深度学习方法实现通信信号的调制识别是一种有效手段。针对当前模型存在着超参数量大、部分信号类型(如正交幅度调制信号)识别率低、识别时间过长等问题,提出了一种基于轻量神经网络的无线电自动调制识别算法。首先通过基于深度可分离卷积的基础单元实现特征提取,并引入通道洗牌操作对不同通道的特征进行重新分配,最终使用注意力机制和Smoothing Maximum Unit(SMU)激活函数加强特征挖掘、复用及学习能力。所提模型能够显著增强空间和通道间的信息交流,有效减少模型超参数量和训练耗时,并进一步解决深层网络中的梯度消失问题。实验结果表明,所提模型的平均识别准确率为90.60%,参数量为75000,训练耗时更短,优于目前流行的调制识别算法,尤其能缓解模型越复杂响应速度越慢的问题,证明了所提模型的有效性及鲁棒性。

In the case of increasingly complex signal environment and variety of signal modulation types,it is an effective means to use deep learning methods to achieve modulation identification of communication signals.For the problems of large amount of hyperparameters,low recognition rate of some signal types such as quadrature amplitude modulation signal,and long recognition time in the current model,an automatic radio modulation recognition algorithm based on lightweight neural network is proposed.Firstly,feature extraction is achieved through the basic unit based on depth-wise separable convolution,and the channel shuffling operation is introduced to redistribute the features of different channels,and finally the attention mechanism and Smoothing Maximum Unit(SMU)activation function are used to strengthen the feature mining,reuse and learning capabilities.The proposed model can significantly enhance the information exchange between spaces and channels,effectively reduce the amount of model hyperparameters and training time,and further solve the problem of gradient disappearance in the deep network.The experimental results show that the average recognition accuracy of the proposed network model is 90.60%,the parameter size is 75000,and the time cost is shorter,which is better than that of the current popular modulation recognition algorithms.Especially,the problem that the more complex of model is,the slower the response speed is,thus illustrating the effectiveness and robustness of the algorithm.