Research On Wireless Communication Demodulation Based On Machine Learning

With the popularity of intelligent mobile devices and the increasing demand for wireless services,how to achieve high spectrum efficiency and low bit error rate has been received considerable attention under the 5G scene with lower latency,large con-nections.Wireless signals in the transmission process may be affected by Doppler fre-quency shift,multipath effect and strong interference,which leads to the formation of non-Gaussian channels.The conventional optimal demodulator is designed for additive white Gaussian noise channel,which can not satisfy the low bit error rate requirements in non-Gaussian channel scenarios.Machine learning(ML),with good robustness and fault tolerance,can adequately approximate any nonlinear system by training to extract data features in the complex environment.Therefore,by using the the advantage of ML to deal with complex non-linear problems,the demodulators based on ML are designed to improve the system demodulation performance.The main work and contributions of this dissertation can be concluded as follows:(1)We first establish a flexible end-to-end wireless communication system under the complex wireless communication environment,which contains vector signal gener-ator,transmitting antenna,receiving antenna,vector signal analyzer and ML based de-modulator.Based on the system model,we then measure and collect the database in the actual indoor scenes,which include binary phase shift keying(BPSK)and quadrature amplitude modulation(QAM)of single-carrier modulation,and orthogonal frequency division multiplexing(OFDM)of multi-carrier modulation.In order to accelerate the rate of convergence,the collected database are normalized.It is worth noting that the single-carrier modulation database is openly accessed online.We upload the database to Baidu Cloud and Google Cloud(2)Two demodulators based on ML are designed for single-carrier modulation,namely demodulator based on deep belief network(DBN)and support vector machine(SVM),and Adaptive Boosting(AdaBoost)demodulator based on ?-Nearest Neighbor(?NN).In the DBN-SVM based demodulator,we first design the DBN demodulator and one-versus-one(OVO)-SVM demodulator,and then the demodulator based on DBN-SVM is derived by adopting DBN for feature extraction and OVO-SVM for feature classification.In the AdaBoost demodulator based on ?NN,the stronger classifier is formed by cascading multiple ?NN weak classifiers.In each iteration,the signal error rate(SER)is reduced by increasing the weight of the error demodulated signal and reducing the weight of the correct demodulated signal.The test results indicate that the BER of the proposed demodulators based on ML is significantly lower than that of the maximum likelihood estimation(MLE)demodulator.(3)Two demodulators base on ML are modified for multi-carrier modulation.In the improved DBN-SVM demodulator,on the one hand,directed acyclic graphs(DAG)-SVM with lower test complexity than OVO-SVM classifiers is used to classify the features outputted by DBN.On the other hand,the structural parameters of DB-N network,such as the length of training batch and the number of neurons,are also adjusted accordingly.In modified AdaBoost demodulator based on ?NN,in order to overcome the redundancy calculation of the original demodulator,a regularization term is added.Meanwhile,the iteration error threshold is setted to prevent over-fitting.The results show that the BER of two modified demodulators is lower than that of coherent demodulation.