CN-121984606-A - Method and system for eliminating simultaneous same-frequency full duplex self-interference

CN121984606ACN 121984606 ACN121984606 ACN 121984606ACN-121984606-A

Abstract

The invention relates to the technical field of signal processing and discloses a method and a system for eliminating simultaneous same-frequency full duplex self-interference, wherein the method comprises the steps of carrying out data preprocessing on a generated self-interference signal to generate a self-interference signal containing nonlinear factors, and constructing a training data set according to the self-interference signal; the method comprises the steps of establishing a complex CNN-LSTM attention model for extracting nonlinear characteristics and time sequence characteristics of an interference signal, introducing an auxiliary sub-module in a training stage of the model to enable the model to have internal knowledge migration and realize knowledge distillation of a characteristic hierarchy, training the model by simultaneously taking a loss function of complex signal amplitude and phase into consideration, carrying out linear and nonlinear two-stage self-interference elimination on a received signal based on the trained model to realize self-interference suppression in a full duplex system, and effectively solving the problems of insufficient characteristic extraction capability, limited supervision information and low training speed of the existing deep learning-based full duplex nonlinear self-interference elimination method.

Inventors

LU MIN
Yang Aoshuang
DONG YUDAO
WU RUIWEN
SUN YUANMIN

Assignees

上海航天电子通讯设备研究所

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. The method for eliminating the simultaneous same-frequency full duplex self-interference is characterized by comprising the following steps of: S1, carrying out data preprocessing on the generated self-interference signals to generate self-interference signals containing nonlinear factors, and constructing a training data set according to the self-interference signals; S2, establishing a complex CNN-LSTM attention model, wherein the model comprises a complex convolution layer, a complex LSTM layer and an attention layer and is used for extracting nonlinear characteristics and time sequence characteristics of an interference signal; S3, introducing an auxiliary submodule in a training stage of the complex CNN-LSTM attention model to enable the model to have internal knowledge migration, and realizing knowledge distillation of a feature level through the submodule; S4, training the complex CNN-LSTM attention model based on a self-distillation mechanism and the training data set, and performing parameter optimization by adopting a loss function which simultaneously considers the amplitude and the phase of the complex signal; s5, based on the trained complex CNN-LSTM attention model, performing linear and nonlinear two-stage self-interference cancellation on the received signal to realize self-interference suppression in the full duplex system.
2. The simultaneous co-frequency full duplex self-interference cancellation method according to claim 1, wherein in step S1, the data preprocessing specifically comprises: s11 pair The lace transmitting signal at the moment Digital-to-analog conversion, low-pass filtering and IQ mixing, and converting into analog signal to obtain signal ; S12, signal matching Performing power amplification processing to obtain a radio frequency output signal Expressed as: , Wherein, the In order to be non-linear in order, Is an odd number of times, For the memory depth of the power amplifier, Representing the impulse response of the power amplifier; s13, the radio frequency output signal Obtaining self-interference signals through the influence of fading and noise of self-interference channels Expressed as: , Wherein, the The representation stage is Is a self-interference channel impulse response; S14, sorting the self-interference signal based on the processing result Expressed as: , Wherein, the Representing an equivalent channel response including a non-linearity, IQ imbalance, and a wireless channel impulse response of the power amplifier; S15, randomly generating The strip data is used as training samples of a deep learning model, and the acquired baseband signal samples are used as training samples of the deep learning model , Obtaining signals by executing the steps S11-S14 At the same time, the dimension is initialized as follows Is a matrix of (a) And a dimension of Is a matrix of (a) The initial values are all set to 0, and the initialization is carried out ; S16 for baseband signal samples Taking the above materials And calculate the statistical information of each of the repeated signals The signal to be extracted and its statistical information Copying to a matrix In, for signals Taking the above materials Copying to a matrix In order to make If (if) Step S16 is continuously executed, otherwise step S17 is executed; S17, obtaining the assigned matrix Corresponding to the feature vector of the training data set, obtaining the assigned matrix Representing tag values, will And Splicing according to the row direction to obtain a final preprocessed training data set 。
3. The simultaneous co-frequency full duplex self-interference cancellation method according to claim 2, wherein in step S2, said establishing a complex CNN-LSTM attention model comprises: S21, establishing inclusion A feature extraction module for layering the complex convolution layers, Training data Is of the characteristic value of (2) Dimension transformation into , Representing the complex domain, inputting the complex domain into the feature extraction module for layer-by-layer forward propagation, for the first Layer complex convolution layer, its output characteristic diagram Calculated by the following formula: , Wherein, the In order to be able to input the input, Is the first A complex convolution kernel weight matrix of a layer complex convolution layer, Respectively the first The number, size and step size of the convolution kernels of the layer complex convolution layers, Represent the first The processing functions of the layer complex convolution layer, Representing a complex activation function that is a function of the activation, Representing a complex convolution operation; s22, outputting characteristic diagram of the final complex convolution layer As input to the complex LSTM layers, the complex CNN-LSTM attention model includes a plurality of complex LSTM layers, each complex LSTM layer including a plurality of state cells, each state cell at a time instant The input of (2) is Based on the hidden state at the last moment Calculating the outputs of the input gate, the forget gate and the output gate, expressed as: , Wherein, the Respectively corresponding to the input doors Forgetful door And an output door , And Representing the input weight matrix and hidden state weight matrix corresponding to the gate, For the bias value of the weight parameter corresponding to the gate, Representing a complex Sigmoid activation function, Representing a complex convolution operation; Cell state And hidden state Updated by the following formula: , , Wherein, the Representing an element-level multiplication of the elements, And Respectively an input weight matrix for state update and a hidden state weight matrix, Updating the bias value of the weight parameter for the state; s23, integrating the outputs after passing through all the complex LSTM layers into Through the steps of The multiple neurons and the multiple full-connection layers with the activation function of the multiple sigmoids obtain the attention weight Will be And (3) with As input to the attention layer, and calculate the output of the attention layer by the following formula : , Wherein, the Representing element-level multiplication; s24, outputting the attention layer As input to a complex fully-connected layer comprising a neuron, an estimate of the self-interference signal is obtained 。
4. The simultaneous co-frequency full duplex self-interference cancellation method according to claim 3, wherein in step S3, said introducing an auxiliary sub-module further comprises: S31, introducing an auxiliary sub-module with the number of the complex convolution layers to the model structure, and outputting the characteristic diagram of each complex convolution layer Respectively input the first The number of auxiliary sub-modules is one, Calculating to obtain output : , Wherein, the Is the corresponding first A plurality of full connection layer weight matrices of the auxiliary sub-modules, Representing the corresponding first A plurality of full link layer processing functions of the auxiliary sub-modules, Is a complex scalar output.
5. The simultaneous co-frequency full duplex self-interference cancellation method according to claim 4, wherein said auxiliary sub-module structure comprises a plurality of full connection layers and an output layer, said auxiliary sub-module is used only in a training phase and is discarded in a deployment phase.
6. The simultaneous co-frequency full duplex self-interference cancellation method according to claim 4, wherein in step S4, said training said complex CNN-LSTM attention model comprises: S41 from the training data set Random selection of small batch data sets For each training instance therein Acquiring the tag value thereof Will be As the input of the complex CNN-LSTM attention model, the output of the complex CNN-LSTM attention model is obtained according to the processing flows of the steps S21-S24 and S31 And (d) Output of the auxiliary sub-modules ; S42, calculating a loss function by the following formula And inversely updating the parameters of the complex CNN-LSTM attention model and each auxiliary sub-module: , Wherein, the In order to assist in the total number of sub-modules, Weight is lost for the master model, and , Is the first The weight lost by each auxiliary sub-module meets the following conditions , 、 And Respectively outputs of 、 And The corresponding value of the amplitude angle is used, 、 And Respectively outputs of 、 And A corresponding modulus value; S43, repeating the steps S41-S42 until the model parameters are converged, and storing the parameters of the complex CNN-LSTM attention model.
7. The simultaneous co-frequency full duplex self-interference cancellation method according to claim 6, wherein in step S5, said performing linear and nonlinear two-stage self-interference cancellation on the received signal comprises: s51 for Time of day received signal Linear self-interference elimination is carried out by a least square method or an adaptive filtering method to obtain a signal of linear self-interference elimination ; S52, the received signal is processed Subtracting a signal from linear self-interference cancellation Obtaining a residual signal ; S53, the residual signal is processed As the input of the trained CNN-LSTM attention model, the model reasoning is performed to output the estimated value of the nonlinear self-interference signal ; S54 of residual signal Subtracting the estimated value of the nonlinear self-interference signal And obtaining the signal after nonlinear self-interference elimination at the receiver end.
8. A simultaneous co-frequency full duplex self-interference cancellation system, comprising: the self-interference signal generation and preprocessing module is used for carrying out data preprocessing on the generated self-interference signals, generating self-interference signals containing nonlinear factors and constructing a training data set according to the self-interference signals; The complex attention model construction module is used for establishing a complex CNN-LSTM attention model, and the model comprises a complex convolution layer, a complex LSTM layer and an attention layer and is used for extracting nonlinear characteristics and time sequence characteristics of the self-interference signals; The internal knowledge migration and self-distillation module is used for introducing an auxiliary sub-module in the training stage of the complex CNN-LSTM attention model, so that the model has internal knowledge migration, and knowledge distillation of a feature level is realized through the sub-module; The model training and optimizing module is used for training the complex CNN-LSTM attention model based on a self-distillation mechanism and the training data set, and carrying out parameter optimization by adopting a loss function which simultaneously considers the amplitude and the phase of the complex signal; And the two-stage self-interference elimination execution module is used for carrying out linear and nonlinear two-stage self-interference elimination on the received signal based on the trained complex CNN-LSTM attention model, and realizing self-interference suppression in the full duplex system.
9. A computer readable storage medium having stored thereon a computer program, wherein the program when executed by a processor implements the simultaneous co-frequency full duplex self interference cancellation method of any of claims 1-7.
10. An electronic device comprising one or more processors and storage means for storing one or more programs that, when executed by the one or more processors, cause the one or more processors to implement the simultaneous co-frequency full duplex self interference cancellation method of any of claims 1-7.

Description

Method and system for eliminating simultaneous same-frequency full duplex self-interference Technical Field The invention relates to the technical field of signal processing, in particular to a method and a system for eliminating simultaneous same-frequency full duplex self-interference. Background With the rapid development of mobile communication technology, the types and the number of wireless services continue to increase, and the demand for scarce spectrum resources is also increasing. The same-frequency full duplex technology allows the communication equipment to simultaneously transmit and receive signals in the same frequency band, so that the spectrum utilization rate can be theoretically doubled, and the technology is considered as one of key technologies in 5G and future 6G communication. However, this technique faces a fundamental challenge in that the high power signal generated by the transmitter is directly coupled into the receiver of the same device, creating severe self-interference. If the interference signal cannot be effectively eliminated, the self-interference signal can submerge the weak expected receiving signal, so that the system cannot work normally. Conventional self-interference cancellation techniques mainly include spatial domain, analog domain, and digital domain approaches. Spatial domain methods rely on antenna isolation, but are limited by device size and isolation is limited. Analog domain methods cancel in the radio frequency link, but the non-idealities of the analog devices introduce new distortions. The digital domain method is processed in the baseband and becomes the current mainstream. The polynomial-based digital elimination method is widely applied, but the number of model parameters thereof increases exponentially along with the nonlinear order, so that the calculation complexity is too high, and the real-time requirement is difficult to meet. In recent years, deep neural networks have been introduced into the field of self-interference cancellation due to their strong nonlinear fitting capability. Initial studies have employed fully connected networks for modeling, and the subsequent use of convolutional neural networks has further improved the ability to extract nonlinear features. Considering that the communication signal is complex in nature, subsequent researches further provide a model based on a complex convolution network, and performance superior to that of a real network is achieved. However, the existing deep learning-based cancellation method still has the defects that firstly, most models (such as CNN) have limited modeling capability on time memory characteristics (such as memory effect of a power amplifier) of signals, secondly, the existing method usually only considers the real part and the imaginary part of the signals during training or splits complex numbers into the real part and the imaginary part for separate processing, so that important phase information is lost, model precision is limited, and finally, network depth is often simply increased to improve performance, so that model complexity and training time are obviously increased, and practical deployment is not facilitated. Therefore, how to design a self-interference elimination method capable of fully extracting space-time characteristics of signals and effectively utilizing complex phase information while maintaining high-efficiency training and reasoning speed is a technical problem to be solved in the art. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a simultaneous same-frequency full-duplex self-interference elimination method based on a complex attention mechanism and self-distillation, so as to solve the problems of insufficient feature extraction capability, limited supervision information and low training speed of the full-duplex self-interference elimination method based on deep learning. In one aspect, a method for simultaneous same-frequency full duplex self-interference cancellation is provided, including the following steps: S1, carrying out data preprocessing on the generated self-interference signals to generate self-interference signals containing nonlinear factors, and constructing a training data set according to the self-interference signals; S2, establishing a complex CNN-LSTM attention model, wherein the model comprises a complex convolution layer, a complex LSTM layer and an attention layer and is used for extracting nonlinear characteristics and time sequence characteristics of an interference signal; S3, introducing an auxiliary submodule in a training stage of the complex CNN-LSTM attention model to enable the model to have internal knowledge migration, and realizing knowledge distillation of a feature level through the submodule; S4, training the complex CNN-LSTM attention model based on a self-distillation mechanism and the training data set, and performing parameter optimization by adopting a loss functio