An intelligent metasurface phase shift optimization method and system based on unsupervised learning

Abstract

The application relates to an intelligent reflector surface phase shift optimization method and system based on unsupervised learning, belonging to the field of communication technology. A base station, a user and an intelligent reflector surface are included. The direct link between the base station and the user is blocked. The input data is indirect cascade channel state information. The data preprocessing step is to convert a one-dimensional channel state information vector into a two-dimensional complex matrix, split the real part and the imaginary part, and additionally add an absolute value part to form three-channel data as the input of a neural network. The whole network is composed of two convolution-attention blocks, a flattening layer and multiple fully connected layers. The attention mechanism selects a squeeze-and-excitation networks network. The model adopts an unsupervised learning method without labels. The loss function is defined as the negative value of the spectral efficiency. Considering that the output of the network needs to meet the phase shift constraint of the IRS, a penalty term is additionally added in the loss function to limit the output. The application has high calculation efficiency, and the phase shift optimization result has certain reference value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication, and is an optimization method for optimizing the reflection phase shift of an intelligent reflecting surface to maximize the spectral efficiency. BACKGROUND

[0002] Intelligent reflecting surface (IRS) is one of the most promising technologies in future wireless communication networks. Specifically, an intelligent reflecting surface is a planar metasurface equipped with a large number of passive reflecting elements connected to an intelligent controller, which can produce independent phase shifts or amplitude attenuations (collectively referred to as reflection coefficients) at each reflecting element in real time. By reasonably designing its reflection coefficients, the signals reflected by the IRS can be constructively added or destructively subtracted with other path signals to increase the signal strength at the receiver or reduce co-channel interference. Theoretical analysis shows that an IRS with N reflecting elements can achieve a total beamforming gain of N 2 .

[0003] In IRS-assisted communication, the design of the optimal IRS phase configuration is crucial to the enhancement of system performance. Due to the unit modulus constraint of the IRS element phase shift and the large number of elements to be optimized, the entire optimization problem is essentially non-convex, so it is difficult to achieve the optimal solution of the IRS phase configuration. The traditional solution is to use semi-definite programming, alternating optimization and other iterative methods to optimize the IRS phase shift. However, due to the iterative nature of these algorithms, the computational complexity of these algorithms is very high. In real-time communication scenarios, it will bring a lot of delay.

[0004] In the past few years, deep learning has shown its significant potential in dealing with non-convex optimization problems. Deep learning has strong expression ability and can learn complex function mapping, and its convergence speed is superior to traditional non-convex optimization methods. In addition, the calculation of the deep learning model can be parallelized, and the use of hardware acceleration technologies such as GPU can greatly speed up the training and inference of the model. In the context of IRS-assisted communication systems, researchers have used different deep learning methods to solve non-convex optimization problems, including supervised learning, unsupervised learning and deep reinforcement learning.

[0005] Therefore, based on the above problems, it is of great significance to provide an intelligent reflecting surface reflection phase shift optimization method based on deep learning. SUMMARY

[0006] The purpose of the present application is to overcome the shortcomings of the prior art and provide a convolutional neural network unsupervised learning scheme assisted by an attention mechanism to optimize the reflection phase shift of an intelligent reflecting surface to maximize the spectral efficiency.

[0007] A smart reflector phase shift optimization method based on unsupervised learning.

[0008] Complete the model feature design. The model includes a base station, a user, and a smart reflector. The direct link between the base station and the user is blocked, and only an indirect cascaded channel exists. The indirect channel is split into rows and columns and multiplied to form a new channel state information vector.

[0009] For channel state information data preprocessing, the one-dimensional channel state information vector is first transformed into a two-dimensional complex matrix, the real and imaginary parts are split, and the absolute value part is added to form three-channel data as input to the neural network.

[0010] The model network structure consists of two convolutional attention blocks, a flattening layer, and multiple fully connected layers. The attention mechanism uses Squeeze-and-excitation networks. The model employs an unsupervised learning method that does not require labels. The loss function is defined as the negative of the spectral efficiency. Considering that the network output needs to satisfy the phase shift constraint of the IRS, an additional penalty term is added to the loss function to limit the output and avoid overfitting.

[0011] Furthermore, the model's network structure employs a convolutional-attention module for feature extraction, including the following steps:

[0012] 1) Normalize the input data;

[0013] 2) Input the normalized data into two cascaded convolutional-attention modules for feature extraction;

[0014] 3) After passing through the convolutional attention module, the feature map is flattened by a flattening layer and then input into two fully connected layers for classification;

[0015] 4) Use linear activation in the last layer to obtain the output.

[0016] The convolution-attention module employs a channel attention mechanism, which learns the importance of each feature channel and assigns higher weights to more important feature channels. The overall network structure is activated by the Leaky-ReLU function, the output layer is activated by the linear function, and the optimizer uses the Adam algorithm.

[0017] Furthermore, the input data for the neural network is obtained in the following way:

[0018] Assuming the channel information state is known, the channel gain between the base station and the IRS is... The channel gain between the IRS and the user is , the IRS adjusts the phase angles of its reflecting elements to maximize the user-combined way of incident signals, defining the IRS phase shift vector as , whose diagonal reflection phase shift matrix is represented as where, , is the reflection phase shift of the intelligent reflecting surface, is the corresponding amplitude;

[0019] Assume that the transmitted signal vector is , and satisfies the transmitted power constraint where is the transmitted power of the base station, and the user receives the signal through the indirect channel assisted by the IRS, the received signal at the user is represented as: where, is defined as the equivalent cascaded channel, represents the complex additive white Gaussian noise (AWGN) at the user, is the noise power;

[0020] The spectral efficiency of the intelligent reflecting surface assisted MIMO communication system is represented as: , extracting the effect of each individual phase shift in , the relationship of the spectral efficiency function about the phase to be optimized is implicit, rewriting the spectral efficiency as an explicit function about , re-representing the indirect channel and as , where, represents the nth column of , represents the nth row of , since is a diagonal matrix, the indirect cascaded channel is rewritten as: where, , the optimization problem of the intelligent reflecting surface phase is rewritten as an explicit function about the IRS phase shift, for each , has the following representation:

[0021] ,

[0022] where,

[0023]

[0024]

[0025] and are all independent of , and is full rank, so is an invertible matrix, rewrite the objective function as:

[0026]

[0027] If all the variables are fixed, then and are fixed, and the objective function is equivalent to maximizing , and is determined by , where represents the non-zero unique eigenvalue of ;

[0028] The optimal required for the optimization problem is obtained by , i.e., by ;

[0029] Take as the input of the neural network.

[0030] The application also provides an intelligent reflector phase shift optimization system based on unsupervised learning for realizing the above method, the system is a MIMO downlink transmission system composed of a base station configured with M antennas and a user configured with K antennas, an intelligent reflector with N reflecting units is used to assist the communication between the base station and the user, wherein the direct link between the base station and the user is blocked, and only an indirect cascade channel exists, and the system further comprises:

[0031] The channel state information data preprocessing module is used for converting a one-dimensional channel state information vector into a two-dimensional complex matrix, splitting the real part and the imaginary part, and additionally adding an absolute value part to form three-channel data as the input of the neural network.

[0032] Further, in the model network structure of the system, the entire network is composed of two convolution-attention blocks, a flattening layer and a plurality of fully connected layers, wherein the attention mechanism selects a Squeeze-and-excitation networks network, the model adopts an unsupervised learning method without labels, the loss function is defined as the negative value of the spectral efficiency, and considering that the output of the network needs to meet the phase shift constraint of the IRS, a penalty term is additionally added in the loss function to limit the output.

[0033] Further, each convolution-attention module contains 40 2X2 size convolution kernels and a batch normalization layer.

[0034] The advantages and positive effects of the present application are: the present application adopts deep learning to optimize the IRS reflection phase shift, compared with the traditional convex optimization method, the deep learning can automatically learn the effective feature representation from the data, and the learned feature representation is more robust and has strong generalization performance, while the traditional optimization method is relatively sensitive; the calculation efficiency of the deep learning model is higher, and the GPU parallel computing method can be used to improve the calculation efficiency, and the use of deep learning is wider in real-time communication. In the present application, first, an effective input structure is designed for the model, which can greatly improve the learning ability of the model, and a special multi-channel input data is designed for the model, second, the attention mechanism in the visual field is introduced into the model, so that the network model can adaptively adjust the weight and pay more attention to the features with higher importance, and the unsupervised learning is used to avoid the high cost of labels, the present method is better than the traditional neural network method, and the phase shift optimization result has a certain reference value. BRIEF DESCRIPTION OF DRAWINGS

[0035] The technical solutions of the present application will be further described in detail below in combination with the drawings and embodiments, but it should be understood that these drawings are designed only for illustrative purposes, and therefore should not be regarded as limiting the scope of the present application. In addition, unless specifically indicated, these drawings are only intended to conceptually illustrate the structural configuration described herein.

[0036] Figure 1 is a system model diagram.

[0037] Figure 2 is a multi-channel input data generation diagram.

[0038] Figure 3 is an attention mechanism assisted convolutional neural network unsupervised learning scheme diagram.

[0039] Figure 4 is a specific diagram of the attention mechanism network Squeeze-and-excitation networks. DETAILED DESCRIPTION

[0040] First of all, it should be noted that the specific structure, features and advantages of the present application will be described in detail below by way of example, however, all the descriptions are only used for illustration, and should not be understood as any limitation on the present application. In addition, any single technical feature described or implied in each embodiment mentioned in the present application, or any single technical feature shown or implied in each drawing, can still continue to be combined or deleted between these technical features (or their equivalents), so as to obtain more other embodiments of the present application which can not be directly mentioned in the present application.

[0041] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0042] In conjunction with specific embodiments of the present invention Figures 1-4 Please provide a detailed explanation:

[0043] This invention relates to an intelligent reflector phase shift optimization method based on unsupervised learning, which includes the following aspects:

[0044] Feature design: See appendix Figure 1 The system model diagram is shown below. The model training data used in this invention is generated by simulation. It is assumed that due to severe congestion, the direct link between the base station and the user is blocked, and only the indirect cascaded channel exists. The indirect channel is split into rows and columns and multiplied to form a new channel state information vector.

[0045] Data preprocessing: See appendix Figure 2 This invention first transforms the one-dimensional channel state information vector into a two-dimensional complex matrix, splitting its real and imaginary parts and adding an absolute value component to form three-channel data as input to the neural network. In this invention, for a two-dimensional convolutional neural network, the input data must be values ​​in the two-dimensional real number field; therefore, preprocessing of the input data is necessary to obtain the required two-dimensional real matrix form for the neural network. Currently, the common method for splitting complex values ​​is to separate them into real and imaginary parts, serving as two real-value channels for the neural network input data. This invention, however, adopts three-channel data, which offers better performance compared to two-channel data.

[0046] Network structure: see appendix Figure 3 and attached Figure 4The whole network of the application is composed of two convolution-attention modules, a flattening layer and a plurality of fully connected layers, wherein the attention mechanism selects a Squeeze-and-excitation networks network, the number of convolution-attention modules, the size of the convolution kernel and the size of the fully connected layer are obtained by experiment.The network structure of the application adopts a convolution-attention module for feature extraction.Specifically, the network structure first normalizes the input data.Then the data is input into two serial convolution-attention modules for feature extraction.Each convolution-attention module contains 40 2X2 size convolution kernels and a batch normalization layer.The channel attention mechanism is adopted in the convolution-attention module, and the more important feature channels are given higher weights by learning the importance of each feature channel, thereby improving the model performance.After the convolution-attention module, the feature map is flattened by a flattening layer, and then input into two fully connected layers for classification.The last layer uses linear activation to obtain the output.The overall network structure uses Leaky-ReLU function activation, the output layer uses linear function activation, and the optimizer uses Adam algorithm.The network structure can effectively utilize the distribution characteristics between feature channels, focus on important features through the channel attention mechanism, and ignore irrelevant features, thereby improving the generalization ability of the model.

[0047] Loss function: Unlike the supervised learning method requiring labels, the model designed in the application adopts an unsupervised learning method without labels, and the loss function is defined as the negative value of the spectral efficiency, and the increase of the spectral efficiency corresponds to the decrease of the loss function, considering that the output of the network needs to meet the phase shift constraint of the IRS, a penalty term is added in the loss function to limit the output to avoid overfitting.

[0048] In the application scenario of the application, consider a MIMO downlink transmission system composed of a base station configured with M antennas and a user with K antennas, in order to further improve the performance of the communication system, the application uses an intelligent reflecting surface with N reflecting elements to assist the communication between the base station and the user, assuming that due to serious blockage, the direct link between the base station and the user is blocked, and only an indirect cascade channel exists. Due to high path loss, the signal reflected twice or more than twice by the intelligent reflecting surface can be ignored. In particular, it is assumed that the channel information state is known, wherein the channel gain between the base station and the IRS is , the channel gain between the IRS and the user is , and the IRS adjusts the phase angle of its reflection element to maximize the user combination incident signal. Define the IRS phase shift vector as , and its diagonal reflection phase shift matrix is represented as , wherein , is the reflection phase of the intelligent reflecting surface, For the corresponding amplitude, in practice, the design of each reflecting element of the IRS is such that the reflected signal is maximized, so typically taken as to obtain maximum reflection gain, i.e. each reflecting element mode is 1.

[0049] Assume the transmitted signal vector is and satisfies the transmit power constraint where is the transmit power of the base station. The user receives the signal through the indirect channel assisted by the IRS. Thus, the received signal at the user can be expressed as: where, is defined as the equivalent cascaded channel, denotes the complex additive white Gaussian noise (AWGN) at the user, is the noise power.

[0050] The spectral efficiency of the smart reflector assisted MIMO communication system can be expressed as: First, extract the effect of each individual phase shift in , since the relation of the spectral efficiency function with respect to the phase to be optimized is implicit. Therefore, rewrite the spectral efficiency as an explicit function with respect to . Represent the indirect channel and as , . Where denotes the nth column of , denotes the nth row of , and furthermore, since is a diagonal matrix, the indirect cascaded channel can be rewritten as: where, , in order to express the optimization problem with respect to the smart reflector phase explicitly, the explicit function with respect to the IRS phase shift for each has the following representation:

[0051] ,

[0052] where,

[0053]

[0054]

[0055] It can be observed that and are independent of , and is full rank, thus is a reversible matrix, the objective function can be rewritten as:

[0056]

[0057] This shows that if all the variables are fixed, then and are fixed, and the objective function is equivalent to maximizing , and can be determined by . Where represents nonzero unique eigenvalue. This shows that the optimal required by the optimization problem can be obtained by , that is, it can be calculated by , so that as the input of the neural network can greatly improve the performance and efficiency of the model, because this input can well reflect the role of each phase shift in the IRS reflection channel, so that the neural network can make better use of the information provided by the data.

[0058] Through the above implementation mode, the technical concept and technical solution of the present application are described in detail. It can be predicted that the present application designs an effective input structure for the input of the neural network, which can significantly improve the model efficiency, and designs an attention mechanism assisted unsupervised learning scheme to better mine channel state information. The present application solves the deficiencies of the previous deep learning algorithm in processing such problems, such as not fully mining channel state information and not effectively paying attention to key information and features of input data in network design. The simulation results prove that the proposed algorithm has the ability of fast convergence, and compared with the benchmark scheme, it can significantly improve the spectral efficiency. The present application provides a more efficient and effective solution for the IRS assisted MIMO downlink communication system, especially in the aspect of mining channel state information, which has important application value.

[0059] The present application proposes an effective model input structure, specifically, an effective one-dimensional input vector is first converted into a two-dimensional complex matrix, and then the real part, imaginary part and absolute value are extracted respectively to form a three-channel input matrix as the input of the neural network. Generating this input structure only requires low computational complexity, but has a significant impact on the learning ability of the neural network.

[0060] In addition, the application proposes an attention mechanism aided unsupervised neural network model (ACnet, Attention-aided Convolution net), the attention mechanism can help the neural network to adaptively focus on the key information and features in the input data, so as to better learn the input data, improve the model calculation efficiency, and additionally add a penalty term in the loss function to ensure that the output meets the constraint condition.

[0061] It should be further explained that the above embodiments have been described in detail, but the above content is only the preferred embodiment of the application and cannot be considered as limiting the scope of the application. Any equivalent changes and improvements made within the scope of the application shall still belong to the patent scope of the application.

Claims

1. An intelligent reflecting surface phase shift optimization method based on unsupervised learning, characterized by: Completing the model feature design, the model includes a base station, a user and an intelligent reflecting surface, the direct link between the base station and the user is blocked, only an indirect cascade channel exists, and the indirect channel is respectively split by rows and columns and multiplied to form a new channel state information vector; Preprocessing the channel state information data, first converting the one-dimensional channel state information vector into a two-dimensional complex matrix, splitting the real part and the imaginary part, and additionally adding an absolute value part to form three-channel data as the input of the neural network; Among them, in the model network structure, the entire network is composed of two convolution-attention blocks, a flattening layer and multiple fully connected layers, the attention mechanism selects the Squeeze-and-excitation networks network, the model adopts an unsupervised learning method without labels, the loss function is defined as the negative value of the spectral efficiency, and considering that the output of the network needs to meet the phase shift constraint of the IRS, a penalty term is additionally added in the loss function to limit the output to avoid overfitting.

2. The method of claim 1, wherein: The network structure adopts a convolution-attention module for feature extraction, including the following steps: 1) Normalizing the input data; 2) Inputting the normalized data into two serial convolution-attention modules for feature extraction; 3) After the convolution-attention module, the feature map is flattened by a flattening layer and then inputted into two fully connected layers for classification; 4) Using linear activation in the last layer to obtain the output.

3. The method of claim 2, wherein: The channel attention mechanism is adopted in the convolution-attention module, which learns the importance of each feature channel and gives higher weights to more important feature channels; the overall network structure adopts Leaky-ReLU function activation, the output layer uses linear function activation, and the optimizer adopts Adam algorithm.

4. The intelligent reflecting surface phase shift optimization method based on unsupervised learning according to claim 1, characterized by: The input data of the neural network is obtained by the following method: Assume the channel information state is known, where the channel gain between the base station and the IRS is , the channel gain between the IRS and the user is , the IRS adjusts its reflection element phase angles to maximize the user combined way incident signal, define the IRS phase shift vector as , its diagonal reflection phase shift matrix is represented as , where , is the reflection phase shift of the intelligent reflecting surface, is the corresponding amplitude; Assume the transmit signal vector is and satisfies the transmit power constraint where is the transmit power of the base station, the user receives the signal through the indirect channel assisted by the IRS, and the received signal at the user is represented as: where, is defined as the equivalent cascaded channel, represents the complex additive white Gaussian noise (AWGN) at the user, is the noise power; The spectral efficiency of the smart reflector assisted MIMO communication system is expressed as: , extracting the role of each individual phase shift , the relationship of the spectral efficiency function with respect to the phase to be optimized is implicit, rewriting the spectral efficiency as an explicit function of , the indirect channel and is re-expressed as , where denotes the nth column of , denotes the nth row of , since is a diagonal matrix, the indirect cascaded channel is rewritten as: where, , the optimization problem of the smart reflector phase is rewritten as an explicit function of the IRS phase shift, for each , has the following expression: ， wherein, and are both independent of , and full rank, so is invertible, rewriting the objective function as: If all the variables are fixed, then and are fixed, the objective function is equivalent to maximizing , and is determined by , where denotes the non-zero unique eigenvalue. The optimal solution required for the optimization problem Depend on Obtain, that is, through To calculate; Will As input to a neural network.

5. An unsupervised learning based smart reflector phase shift optimization system implementing the method of any one of claims 1-4, characterized by: The system is a MIMO downlink transmission system composed of a base station with M antennas and a user with K antennas, an intelligent reflecting surface with N reflecting units is used to assist communication between the base station and the user, wherein the direct link between the base station and the user is blocked, only an indirect cascade channel exists, and the system further includes: A channel state information data preprocessing module for converting a one-dimensional channel state information vector into a two-dimensional complex matrix, splitting the real part and the imaginary part, and additionally adding an absolute value part to form three-channel data as the input of the neural network.

6. The intelligent metasurface phase shift optimization system based on unsupervised learning of claim 5, wherein: In the model network structure of the system, the entire network is composed of two convolution-attention blocks, a flattening layer and multiple fully connected layers, the attention mechanism selects the Squeeze-and-excitation networks network, the model adopts an unsupervised learning method without labels, the loss function is defined as the negative value of the spectral efficiency, and considering that the output of the network needs to meet the phase shift constraint of the IRS, a penalty term is additionally added in the loss function to limit the output.

7. The intelligent metasurface phase shift optimization system based on unsupervised learning of claim 6, wherein: Each convolution-attention module contains 40 2x2 size convolution kernels and one batch normalization layer.

Images