Research On Signal Detection Method Based On Continuous Learning In Massive Mimo System Multi-Mode Channel

Massive Multiple-input Multiple-output(Massive MIMO)is one of the key technologies of Fifth Generation of Mobile Communication(5G),and the signal detection technology has important research significance in the 5G communication environment.The traditional massive MIMO signal detection method applied to the current communication system is difficult to balance the bit error rate performance and complexity,and the development of artificial intelligence technology has inspired current researchers.They combined signal detection technology with Deep Neural Network(DNN)combined to obtain a variety of signal detection networks,some detection networks can avoid matrix inversion through the design of iterative detection algorithms,and the automatic optimization of the weights of each layer through the neural network structure can be omitted Artificial calculation of some parameters.Although the current DNN-assisted signal detection technology can balance performance and complexity well,it still has the problems of learning overhead and training set overfitting.In a communication system with a changing wireless environment,the parameters of the channel will constantly change,and in the current signal detection network,the DNN-assisted signal detector trained for a specific wireless environment(ie,channel model)is usually not suitable for another environment(ie Channel model).In order to solve this problem and avoid the excessive learning overhead caused by repeated training.This paper will use the Continual Learning(CL)algorithm in Deep Learning(DL)to realize the real-time update of the signal detection network.The main contents are as follows:Aiming at the massive MIMO system,this paper proposes a signal detection method based on Elastic Weight Consolidation Continual Learning(EWCCL).In mobile communications,the wireless channel is affected by the wireless propagation environment,and the channel parameters often change,leading to changes in the channel model.The current signal detection network generally can only memorize a single channel model,which makes the use of signal detectors based on deep learning Limited robustness is poor.In this paper,the Rice channel model is adopted,and the multi channels is simulated by adjusting the parameters of the Rice channel.The basic idea of this paper is to calculate the importance of each parameter to the task through the EWCCL method after completing one training,and then in the next training,after adding a regular term to the loss function,avoid the excessive deviation of these parameters.The results of the last training are completely covered.This algorithm can enable the neural network to memorize the training results of multiple wireless channels,thereby having stronger robustness to changes in the wireless channel,and reducing the cost of retraining.Finally,the feasibility of the scheme is verified by simulation experiments,and it is proved that the signal detection network using EWCCL algorithm has sub-optimal performance in massive MIMO multi-channel systems.In this paper,under the millimeter wave massive MIMO system,the EWCCL algorithm is improved.Due to the poor penetration of millimeter waves,short wavelength,poor diffraction ability,and large path loss,obstacles,atmosphere,rain,communication distance extension,etc.Changes in environmental factors will change the scattering parameters of the channel.This paper uses a narrowband millimeter wave channel model,and simulates the nature of the millimeter wave channel that is susceptible to environmental influences by adjusting the scattering parameters of the millimeter wave channel model.Since the penalty term of the EWCCL algorithm increases linearly with the increase of tasks,the calculation overhead is too large,and the millimeter wave channel is more sensitive to changes in the environment.This article extends the usage scenario to the case of N millimeter wave channels.And through the approximate processing of the penalty item and the addition of the forgetting factor,not only the computational complexity is reduced,the computer storage allocation is optimized,but also the forgetting speed of the network can be controlled,and the performance between various tasks can be balanced.Compared with the previous EWCCL algorithm,it is also easier to implement in engineering.