Researches On Modulation Recognition Algorithms Based On Deep Learning

With the development of communication technology,wireless communication is becoming more and more important in modern society.Modulation is a key feature of wireless communication,and the research on modulation recognition is very important in both civil and military field.In military field,those who has capabilities of modulation recognition can easily intercept important information and take the lead in military action.In civil field,modulation recognition technology can help relevant departments to detect the state of spectrum and ensure the security of wireless communication.The traditional modulation recognition method has a great dependency on prior knowledge and takes up a lot of human resources.With the development of wireless communication,it is hard to catch up with the new trend in wireless communication.Driven by the above facts,automatic modulation classification technology has become an important research field in wireless communication.Automatic modulation classification technology can be divided into two categories,likelihood-based and feature-based.Among them,the feature-based automatic modulation classification technology is more common in practical applications because of it's easy to implementation and has an ideal accuracy when the parameters are set properly.Recently,the rising of deep learning has greatly promoted the research and development of feature-based automatic modulation classification technology.Modulation recognition technologies using deep learning model often have a high accuracy and is easy to implement.This thesis mainly about long-term memory network and convolutional neural network,which are two commonly used artificial neural network in the application of modulation recognition technology.A modulation recognition algorithm based on the NIN model and an attention-based LSTM network are designed to improve the performance of the traditional convolution neural network and the long-term memory network that had been used in modulation recognition.In the end,the performance of the algorithms designed in this thesis are simulated and verified in the data set.The specific contents are as follows:1.In order to verify whether traditional machine learning algorithms are suitable for those wireless communication modulation recognition tasks which have a large number of classification targets,this paper selects the appropriate wireless communication modulation signal dataset,and carries out simulation experiments with four classical machine learning algorithms on the dataset.Results show that the traditional machine learning algorithm can't complete the wireless communication modulation tasks that have multiple classification targets effectively.Compared with the traditional machine learning algorithm,the accuracy of algorithms which based on artificial neural network are higher.2.Aiming at the problem that the use of image features in convolutional neural networks for modulation recognition may increase the system complexity and waste system resources in practical applications,learning the advantages of the proposed NIN model,a modulation recognition algorithm based on NIN model is designed.Comparing with the traditional convolutional neural network model,this algorithm has advantages in recognition accuracy,model generalization performance and model training efficiency.3.Aiming at the problem that the time performance of the traditional used LSTM network for modulation recognition will be greatly reduced as the amount of data increases,an attention-based LSTM network is designed.Experimental results show that the time performance of this algorithm is much better than that of the traditional LSTM network used for modulation recognition,and this algorithm has some advantages in overall accuracy.