Deep Learning Based Transceiver Design And Training Method Research

The communication physical layer based on deep learning can learn and fit the real communication channel in real time according to the actual transmitted data,avoiding the off-line modeling of the channel based on statistical model,which has attracted extensive attention in academic circles in recent years.The transmitter and receiver based on deep learning do not need to be strictly divided into different modules,so the system structure is simple and has the potential to achieve better performance.In addition,due to the unified physical layer design,the physical layer based on deep learning has the ability of automatic air interface adaptation,and the implementation of the algorithm avoids the dependence on the hardware achieving or environment.However,there are still many problems to be solved in the research of communication physical layer based on deep learning.Firstly,the current physical layer implementation methods based on deep learning are still much larger than the traditional methods in storage and computational complexity.In particular,Beyond 5G and 6G mobile communications put forward higher requirements for transmission rate,which undoubtedly further increases the complexity of physical layer implementation.Secondly,facing the unknown channel model,we will not be able to train the transmitter through gradient back propagation.In order to further solve these two problems,we propose a communication physical layer method based on Binary Neural Network and Generative Adversarial Network.Experiments show that using the Binary Neural Network,we can reduce the number of parameters and computational complexity to more than 10% of the convolutional neural network with the same structure,so as to solve the problems of complexity and storage space.With the Generative Adversarial Network,we can simulate the unknown channel and update the parameter of transmitter by using the back propagation.In this thesis,we implement a modulation and demodulation model with Binary Neural Network and Generative Adversarial Network under pre training based on deep learning,and use this structure to complete the whole process of communication transmission,which lays the foundation for the subsequent implementation of the physical layer.In this thesis,we have conducted many experiments and built multiple neural networks based on bit and symbol structures,and compared the bit error rate and constellation map,which proved the effectiveness of the pre training based Binary Generative Adversarial Network we used in the thesis.In addition,in order to further verify the generalization of the modulation and demodulation model based on deep learning,we test the ideal Gaussian white noise channel and the phase noise channel in the actual millimeter wave and terahertz system.The experimental results show that no matter which modulation mode or channel we adopt,the methods based on deep learning we used can achieve the expected performance and reduce the computational and storage complexity at the same time,which lays a corresponding foundation for our subsequent research.