Research On Automatic Modulation Classification Algorithm Based On Deep Learning Under Limited Resources

Automatic Modulation Classification(AMC)based on Deep Learning(DL)is a key technology in civil and military fields.Traditional modulation classification technology can be classified into two categories: Likelihood Based(LB)modulation classification technology and Feature Based(FB)modulation classification technology.The computational complexity of traditional LB modulation classification technology is higher.While the traditional FB modulation classification technology requires manual design of signal features and formulation of classification rules,which has a complicated process and unsatisfactory classification performance under complex channel environment,DL method does not require manual design of features.In recent years,many scholars have tried to apply DL to AMC,which has led to AMC’s performance breakthrough.The main work of this paper is as follows:Firstly,the traditional FB modulation classification technique and DL based AMC technique are studied in this paper.In order to further break through the limitations of traditional methods on performance,an improved Residual Network(Res Net)structure is proposed in this paper.The simulation results show that Res Net designed in this paper has advantages in classification performance compared with both traditional neural network and traditional FB classification technology.Secondly,existing DL-based AMC methods usually have high space and computational complexity,require a large amount of time and rely on a large number of Graphics Processing Units(GPUs),which makes it difficult to use DL-based AMC technology in the case of limited time and computing resources.To solve this problem,the AMC method based on Res Net lightweight design is further studied in this paper.A more lightweight MLRes Net structure was designed by means of depth separable convolution,global average pooling,ECA attention mechanism and multi-level feature fusion.The simulation results show that compared with MCLDNN and other networks,the lightweight structure MLRes Net designed in this paper can effectively reduce the number of network parameters and use time,and also has better classification performance.Finally,since the number of labelled samples is very limited in actual communication scenarios,two semi-supervised AMC methods based on MLRes Net model are proposed in combination with the semi-supervised learning theory to solve this problem.Compared with the modulation recognition of unlabelled samples by using MLRes Net directly,The semi-supervised AMC method based on MLRes Net model can improve the overall classification accuracy of the model,and at the same time,it can deal with the changes of the data set more quickly,so as to improve the model’s adaptability to the data set.