Multipath Channel Estimation And Signal Demodulation Based On MAML

In the process of data detection under multipath channels,multipath channel estimation and signal demodulation are key issues that the receiver needs to solve.First,the receiver uses the pilot sequence to estimate the multipath channel,and then uses the channel state information(CSI)to demodulate the transmission signal with a suitable detection algorithm.The channel estimation and demodulation algorithm in the traditional communication field is based on a solid mathematical model and supported by mature communication theory.However,in some extreme scenarios,the performance of existing algorithms may drop sharply or even fail to work: for example,Internet of Things scenarios with limited pilot resources,new communication scenarios lacking mature mathematical models such as molecular channels,scenarios with highly complex channels such as satellite communications,and scenarios where CSI changes strongly over time such as shortwave communication.In this regard,this thesis uses neural network to replace the traditional receiver,and completes multipath channel estimation and signal demodulation in a datadriven manner.In order to further improve the robustness of the designed neural network,a model-agnostic meta-learning(MAML)algorithm with stronger adaptability,fewer training samples,and model-independent in the field of machine learning is studied to achieve the rapid convergence of the training process,which has practical application value.The main research work is as follows:First,the problem of signal demodulation under multipath channels is abstracted into a fewshot learning problem.In order to fit the classification problem data set,it is necessary to build a multipath channel model,and extract the sequence of the sender and receiver to form training samples and labels,so as to migrate the MAML algorithm in the machine learning field to the multipath channel estimation and demodulation problem.In the environment of dynamic changes of CSI such as block fading channel,a MAML-based group channel estimation and demodulation scheme is proposed.Specifically,the MAML training process is performed on the pilot block.Network parameters with high universality are found through the inner and outer loops,which are used as the initialization parameters for executing the MAML test process on each data block in the group.The experimental results further illustrate that under the same degree of mastery of CSI,the demodulation accuracy of this scheme is better than that of the traditional communication method,and the training efficiency is better than that of the other two machine learning algorithms.Based on the group channel estimation and demodulation scheme,optimization schemes are proposed for the two stages of MAML execution.First,adjust the position of the decoder for the meta-training network,embed the decoder inside the neural network.In this way,the error correction ability of channel coding and decoding can be made of full use to reduce the error samples in the network,and assist the efficient training of network parameters.Experiments show that this scheme reduces the number of iterations required for network parameter convergence by half.As to the test network,this thesis proposes a dynamic channel estimation and demodulation scheme based on the idea of meta-learning.It reencode and re-modulate the demodulation result of the previous fading block for repeated training process,providing initialization for the test process of the next fading block.Experiments show that,compared with the group channel estimation scheme,the optimization scheme has the ability to quickly adapt to the dynamic environment,and is less dependent on pilot resources,which can significantly improve the data transmission efficiency of the device.