Joint Signal Detection For MCM Based On Neural Networks

Multi-carrier modulation technology is an indispensable technical means for combating frequency-selective fading in wireless broadband communication systems.Artificial neural net-work is a rapidly developing "artificial intelligence" technology in the past decade,and is widely used in image processing,speech recognition and upper layers of wireless communication sys-tems,and has great application potential in the physical layer.The combination of neural network technology and multi-carrier modulation technology can improve system performance and break through the limitations of traditional block-structure communication systems,and cope with more complex channel environments.This thesis mainly studies the classical multi-carrier modulation technology,multi-carrier joint signal detection system based on neural network and end-to-end multi-carrier system based on auto-encoder.This thesis first introduces the principles and characteristics of multi-carrier modulation tech-nology and artificial neural network technology.In this thesis,the structure and principle of the transceiver of OFDM and FBMC systems are elaborated.Taking FBMC as an example,the received signal under actual channel conditions is analyzed and deduced.Next,the channel estimation algo-rithm and channel equalization algorithm commonly used in the FBMC system will be described.Then,starting from the basic neuron unit structure,this thesis introduces 3 typical artificial neu-ral network models:the deep neural network(DNN),convolutional neural network(CNN)and auto-encoder(AE).And their advantages,disadvantages and areas of expertise are compared.Then this thesis studies the multi-carrier joint signal detection system based on neural network,and proposes the joint signal detection system of DNN-FBMC and CNN-FBMC.This thesis first introduces the basic model of FBMC system based on neural network,and analyzes the difference with traditional FBMC system.Then,the DNN-FBMC system is proposed,which can adaptively learn channel characteristics.Aiming at the time-frequency dimension characteristics of multi-carrier modulation system,a CNN feature extraction model is introduced and CNN-FBMC system is proposed.The CNN-FBMC system is better at learning FBMC time-frequency two-dimensional data features and enhancing system robustness.Based on the original loss function,this thesis pro-poses an optimized loss function from the perspective of analyzing the characteristics of FBMC output signals.The simulation results show that the optimized loss function can greatly improve the convergence speed of NN-FBMC in the early stage of training.In addition,the performance of DNN-FBMC and CNN-FBMC systems are better than that of traditional FBMC systems with least squares channel estimation.Besides,CNN-FBMC has better BER performance,stronger robust-ness and less dependence on pilots than DNN-FBMC.Finally,the multi-carrier end-to-end system based on auto-encoder is studied,taking FBMC as an example.Inspired by the single-carrier end-to-end system based on auto-encoder,this the-sis establishes a parallel coding/decoding network model by using the feature that auto-encoder can reconstruct the input data,and proposes an AE-FBMC end-to-end system.The system breaks through the limitations of traditional block-structure communication systems,giving the entire sys-tem more freedom and self-driving capabilities.Then,this thesis introduces the Spatial Transform Network(STN)algorithm in the image domain,and transforms it into a Communication Transform Network(CTN)model according to the characteristics of the multi-carrier system,and applies it to the AE-FBMC system,so that the CTN-AE-FBMC end-to-end system was proposed.The system not only has the advantages of AE-FBMC,but also can learn more channel features in combina-tion with the transform network.The simulation results show that the AE-FBMC system has better BER performance than the traditional FBMC system with least squares channel estimation,and the CTN-AE-FBMC system performance is much better than the AE-FBMC system.Its BER per-formance is very close to the traditional FBMC system with minimum mean square error channel estimation.In addition,the CTN-AE-FBMC system has low dependence on pilots and can achieve better robustness without pilots.Further simulations also show that the CTN-AE-FBMC system has a performance improvement compared to the CNN-FBMC system.