Continuous Learning For Modulation Classifications

In the 21st century,machine learning and artificial intelligence are gradually combined with wireless communication technologies,together with continuous expansion of big data and the constant improvement of computing power,which have led to innovation in the field of communications.Nowadays,the world is evolving to a digital economy.Big data and high-performance computing are becoming affordable commodities.With the increasingly complex electromagnetic environment,radio spectrum has become one of the most valuable and scarce resources in the world.After in-depth research on the automatic classification of modulation signals,this thesis simulates wireless communication environments with numerous types of signals,aiming to break through the traditional isolated modulation classification using deep learning,in order to solve the problem of catastrophic forgetting and realize continuous automatic modulation recognition.This thesis analyzes a variety of continuous learning algorithms in depth,and improves upon conventional algorithms to achieve better continuous learning of modulated signals.In addition,neural networks can accurately learn the intricate data structure using deep learning,to optimize the training and learning of neural networks.Therefore,this thesis proposes three different continuous learning algorithms for wireless communication signals.Firstly,this thesis defines the background of wireless networks for continuous automatic recognition of modulation signals.Secondly,this thesis describes the large-scale data set containing two kinds of analog modulation signals and eight kinds of digital modulation signals.Finally,this thesis proposes four different neural network models to reproduce the catastrophic forgetting problem,and proposes three continuous learning different algorithms.Gradient episodic memory algorithm uses the nature of gradient descent to restrict the direction of gradient descent of two tasks,to ensure that both tasks show the trend of optimization losses.Gradient episodic memory combined with replay algorithm not only restricts the descending direction of gradient,but also enhances the memory of neural network model again.The elastic weight consolidation algorithm uses Fisher information matrix to estimate the distribution of model parameters that carrying memory,to selectively slow down the forgetting.Numerous simulation results show that the first two continuous learning algorithms improve the continuous automatic recognition accuracy of modulation signals by 35% and 50%,respectively,and the improvement effect of the third continuous learning algorithm is affected by the value of lambda,which expand the sustainable development of modulation classification.