基于卷积神经网络的多进制相位调制信号识别算法

MULTI-PHASE MODULATION SIGNAL RECOGNITION ALGORITHM BASED ON CONVOLUTIONAL NEURAL NETWORK

由于多进制相位调制子类信号相似度高,传统的信号识别方法和机器学习算法难以实现特征的自动提取和准确的分类.针对此问题,提出一种基于时频图和深度卷积神经网络的识别算法.将实测信号通过短时傅里叶变换转换成时频图作为实验数据,并设计一个33层的卷积神经网络ReSENet对特征进行自动提取和调制识别.该网络融合了经典模型ResNext和SENet的优点,能通过深度学习和特征重定向学习到数据中复杂抽象的特征.为进一步提高ReSENet的性能,分别从梯度下降算法、激活函数等方面对模型进行优化.与现有方法相比,该算法在对多进制相位调制信号识别上有更优的分类表现.实验结果显示,最终的识别准确率达到99.9%,验证了该算法的有效性.

Due to the high similarity of multi-phase modulated signals,traditional signal recognition methods and machine learning algorithms are difficult to achieve automatic feature extraction and accurate classification.To overcome the problems mentioned above,we proposed a recognition algorithm based on time-frequency graphs and deep convolutional neural network.The measured signals were converted into time-frequency graphs by short-time Fourier transform as experimental data.Then,a 33-layer convolutional neural network ReSENet was designed for automatic feature extraction and modulation recognition.Combining the advantages of the classic models ResNext and SENet,ReSENet was capable of learning complex and abstract features of data through deep learning and feature redirection.In order to further improve the performance of ReSENet,the model was optimized from the aspects of gradient descent algorithm,activation function,etc.Compared with the existing methods,the proposed algorithm has better performance for multi-phase modulation signal recognition.Experiments conducted on the measured data have verified the effectiveness of the proposed algorithm with an accuracy of 99.9%.