| In the process of wireless communication technology development,signal modulation recognition is one of the hot research topics.With the development of 5G,the communication environment has become more complex,and there are more and more types of signal modulation,and it is difficult for traditional signal modulation identification methods to meet the needs of communication networks.As a current hot artificial intelligence technology,convolutional neural network as the core has been able to solve complex problems such as feature fitting,multi-dimensional classification and pattern recognition,however the method based on neural network requires a large number of training samples,it is very difficult to obtain a large number of reliable and high-quality data,the use of data augmentation combined with convolutional neural network for automatic recognition of modulated signals has become an important research direction in the field of communication,the main research content of this thesis includes:1.Summarize the research background of signal modulation recognition methods,the research status at home and abroad,and analyze the problems existing in the current research.The basic theories of neurons,activation functions and other basic theories based on neural networks and the types of modulated signals are summarized,and the forward propagation,backpropagation and optimization algorithms of the neural network training process are derived in detail,focusing on the analysis of the basic principles of convolution,residual and long short-term memory networks,so as to provide theoretical support and experimental comparison conditions for the construction of neural network structures in subsequent chapters.2.Two optimized convolutional neural network models are designed by introducing the improved residual module and the long short-term memory network respectively,which transform the signal modulation recognition problem into the signal classification problem in deep learning,and the signal recognition accuracy is improved by 30% compared with the original model.At the same time,the data format transformation of the input signal is carried out,the recognition results of I/Q and A/P data transformation formats under different convolutional neural networks are studied,and the identity mapping layer and long short-term memory neural network are used to extract high-dimensional information and time series information to improve the signal recognition accuracy,and the overall recognition accuracy of the two models reaches 91% and 92%,respectively.In order to verify the separability of the neural network model,the t-SNE algorithm is used to visualize the overall distribution of the output data at the network layer,and the signal modulation recognition process of the neural network is studied and analyzed.3.The data augmentation method of time series transformation and constellation map transformation is proposed to combine with the spatiotemporal multi-channel neural network to extract the temporal and spatial characteristics of the signal and effectively improve the accuracy of signal modulation and recognition under small sample conditions.Combining the advantages of Convolutional Neural Network(CNN),Long Short-Term Memory(LSTM),and Deep Neural Networks(DNN),the classification result can reach 93.5% when the training sample is only 360,000.In particular,the average recognition accuracy of QAM16 and QAM64 signals is increased by about 25%,which solves the problem of easy confusion between the two signals and provides an effective method for signal modulation and recognition under small sample conditions. |
