Method and device for eliminating self-interference

Abstract

The invention discloses a self-interference elimination method and device, comprising: obtaining a baseband signal x(t) at time t, and importing the baseband signal x(t) into a training model of an input layer and a convolution layer to obtain an output signal; the training model includes: Introduce a three-dimensional tensor in the input layer, set a complex convolutional layer structure in the convolutional layer; input the output signal into the LSTM layer, the LSTM layer is used to process sequences with timing, and output the results to the fully connected layer, which is used for The output result performs dimension transformation of the data to obtain the dimension transformation result; according to the dimension transformation result, it is input to the output layer, and the output layer outputs two neurons. The present invention introduces three-dimensional tensors in the input layer, sets a complex convolution layer structure in the convolution layer, designs two network structures for reconstructing self-interference signals, and makes full use of the advantages of convolutional neural network local perception and weight sharing , so as to learn more abstract low-dimensional features in high-dimensional features, so as to improve the effect of self-interference cancellation.

Description

technical field [0001] The present invention relates to the technical field of full-duplex communication, in particular to a self-interference elimination method and device. Background technique [0002] In the face of nonlinear effects caused by radio frequency devices contained in self-interference signals, traditional methods need to combine relevant prior knowledge to establish mathematical models to describe nonlinear effects, and then obtain model parameters through channel estimation methods to reconstruct self-interference signals. [0003] Due to the extreme reliance on relevant prior knowledge, if the model mismatch occurs, the elimination effect will be seriously deteriorated, and the method of manually designing the model to estimate the relevant parameters is relatively inefficient. And combined with the traditional deep neural network, there are the following disadvantages. The fully connected layer neural network only uses the multi-layer perceptron and nonlin...

AI Technical Summary

Problems solved by technology

[0003] Due to the extreme dependence on relevant prior knowledge, if the model mismatch occurs, the elimination effect will be seriously deteriorated, and the method of manually designing the model to estimate the relevant parameters is relatively inefficient

And combined with the traditional deep neural network, there are the following disadvantages. The fully connected layer neural network only uses the multi-layer perceptron and nonlinear activation function to reduce the feature extraction process and the ability to treat the nonlinear approximation of the fitting function to a certain extent, but cannot target Unique high-dimensional data has characteristics such as space-frequency correlation and time correlation for processing

Images