| With the increasing demand for communication services in the whole society,wireless communication technology is constantly evolving.At present,the fifth generation(5g)mobile communication is becoming mature and becoming the mainstream technology for the new generation of interconnected society.As one of the key technologies of physical layer,5G puts forward higher requirements for the reliability,transmission rate and spectral efficiency of Orthogonal Frequency Division Multiplexing(OFDM)system.The performance of OFDM is closely related to channel estimation and modulation classification in the receiver.The fading and time-varying characteristics of multipath channels are aggravated by complex and changeable communication scenarios,while the traditional channel estimation and modulation classification algorithms have limited performance or lack of flexibility under such conditions,which cannot meet the requirements.In recent years,the rapid development of artificial intelligence and deep learning(DL)has shown excellent performance and extensibility in various fields,providing new ideas for wireless communication technology.In order to improve the performance of OFDM receiver in time-frequency doubly selective channel,the channel estimation and modulation classification algorithms based on DL are studied in this paper,and their application prospects are explored.Firstly,driven by the image super-resolution,a doubly-selective channel estimation scheme based on U-shaped neural network(UCENet)is proposed.The neural network is used to model the direct mapping between the low-resolution channel response and the accurate high-resolution channel response at the pilot locations,so as to improve the accuracy of least square(LS)estimation.The network realizes feature extraction and abstraction in the down-sampling stage,and restores the location information and the complete channel according to the abstract features in the up-sampling stage.Due to the different changing rates of the channel response in frequency and time direction,the features of different scales are fused through skip connection.In the up-sampling stage,the attention mechanism is also used to readjust the feature intensity in the dimensions of channel and space.Furthermore,in order to improve the robustness of the algorithm against noise,a two-stage neural network channel estimation scheme based on signal-to-noise ratio(SNR)classification is designed,which is called SC-CE.In the first stage,the training classification network obtains the SNR information of the input signal without any prior knowledge,and classifies it into high-quality and low-quality signals with 1 d B as the boundary.In the second stage,the UCENets with/without attention modules are used for channel estimation of the two kinds signals.Using the accurate channel information output by the neural network,the signals are equalized by zero forcing(ZF)algorithm to compensate the amplitude and phase distortion,and then sent to the subsequent modulation classification module.On the basis of fully considering the signal characteristics,an adaptive fusion neural network(AFNet)is proposed for OFDM signal modulation classification.The network can intelligently extract and aggregate the multi-scale spatial features of in-phase/quadrature(I/Q)signals,so as to improve the ability of feature representation.In addition,a new confidence weighted loss function is proposed to solve the imbalance of sample quality,which is realized by two-stage learning strategy.Through two-stage learning,neural network can focus on high confidence samples with more real information and extract effective features,so as to improve the overall classification accuracy.In order to evaluate the performance of the proposed algorithms,an OFDM communication system is constructed using MATLAB to generate the dataset for off-line training.The neural network models are designed based on Python and Tensorflow,and then they are deployed to the OFDM system for joint simulation after achieving the best parameters.The simulation results of channel estimation show that the performance of UCENet is improved by more than 10 d B compared with LS algorithm,and 2～8 d B compared with the other neural network.For SC-CE,the classification accuracy of SNRs in the first stage is higher than 99%,while the estimation performance of the complete scheme is further improved on the basis of UCENet,and better robustness is obtained under the SNRs of-10～20 d B.And it still maintains a great advantage when the number of pilots is halved and the moving speed becomes higher.The joint simulation results of channel estimation,equalization and AMC show that the combined scheme of SC-CE and CW-AFNET has an average classification accuracy of 54.39 % and a maximum accuracy of 90.20 % at-10～20 d B,which is the best performance in the absence of any prior information.In addition,the confidence loss function also improves the highest accuracy of other modulation classification models by 0.98%～2.31%,which verifies its effectiveness and universality. |
