Secure communication optimization method for double-IRS-assisted simultaneous wireless information and power transfer system based on graph neural network

Abstract

The invention relates to a secure communication optimization method for a double-IRS-assisted simultaneous wireless information and power transfer system based on a graph neural network, and belongs to the field of wireless communication security. According to the method, reflected beam forming of two IRSs and transmitted beam forming of a base station are jointly optimized to maximize the total information transmission rate of a system, the transmitted power constraint, the maximum eavesdropping rate constraint, the energy collection constraint and the modulo-one constraint of IRS reflection coefficients of the base station are met at the same time, and a machine learning method combining a Lagrangian multiplier method and GNN is adopted. Pilot frequency information used for estimating a channel is parameterized into a GNN representation vector, the representation vector is updated layer by layer according to a GNN connection graph, and beam forming at a BS position and a reflection coefficient at an IRS position are optimized according to a system target. Compared with a traditional method using channel estimation and non-convex optimization, the method provided by the invention obtains a higher information transmission rate on the premise of meeting the requirements of energy collection users and information collection users.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication security and relates to a method for optimizing secure communications of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network. Background Art

[0002] Intelligent reflecting surface (IRS) is a two-dimensional electromagnetic metasurface composed of a large number of low-cost passive components, which can change the wireless propagation environment and improve the performance of wireless communication systems. Physical layer security (PLS) is an information security technology that uses the randomness of wireless channels to achieve confidential communications. Since IRS can improve the quality of information collection user channels while reducing the quality of eavesdropping user channels, the two technologies have complementary advantages.

[0003] On the other hand, simultaneous wireless information and power transfer (SWIPT) technology has been widely studied since it was proposed. Because it can power a large number of low-power devices in wireless sensor networks and industrial Internet, it is considered to be one of the key technologies of the future Internet of Things.

[0004] Combining IRS with SWIPT can bring out the advantages of both. On the one hand, IRS can improve the information transmission capability. On the other hand, IRS can effectively compensate for the path loss of RF signals, thereby establishing an energy collection area near IRS and improving the energy transmission efficiency in this area.

[0005] The information transmission of the SWIPT system is vulnerable to eavesdropping, and information communication security has become one of the key issues that need to be solved urgently. For example, in order to meet the charging needs, energy collection users are usually closer to the base station BS than information collection users, so the signal quality is better, but this creates information security risks: since the data signal and the energy signal are transmitted at the same time, the energy collector can receive a stronger data signal than the information collection user, which makes it possible to eavesdrop on the privacy information of the information collection user. In recent years, in addition to the use of traditional encryption methods at the application layer, the concept of physical layer security (PLS) has been proposed, such as cooperative relaying, artificial noise, etc. Due to the outstanding performance of IRS in the field of wireless communication security, the IRS-assisted wireless communication PLS problem has become a hot topic of research.

[0006] Although many works have studied the secure communication of IRS-assisted wireless communication systems, there is no related technology for the secure communication of dual IRS-assisted SWIPT systems that considers channel estimation and reflection links between IRSs. The prior art with publication number CN113556164A discloses a beamforming optimization method based on energy efficiency priority in a single IRS-assisted SWIPT system, but does not consider communication security. The prior art with publication number CN116192220A discloses an IRS-assisted cognitive SWIPT system security rate optimization method, but there are a large number of variables and iterations, the problem convergence speed is slow, and the results obtained may not be close to the optimal solution. How to design the beamforming of the BS and the reflection vectors of the two IRSs to meet the needs of energy collection users and information collection users is a difficult point in the research. Compared with the optimization of a single IRS system, the dual IRS optimization problem not only needs to consider the reflection channels from the two IRSs to the users, but also needs to consider how to better utilize the channels between the two IRSs to achieve collaboration. Since the dual IRS-assisted wireless communication system has a large number of channel parameters and the system structure is more complex, it brings more difficult challenges to channel estimation. Channel estimation of dual IRS-assisted wireless communication systems is a time-consuming and arduous problem in itself. Considering the non-convexity of the optimization problem and the coupling of system variables, the traditional solution is to introduce slack variables to decouple the system variables, and then transform the non-convex optimization problem into an approximate convex optimization problem, and use alternating optimization to obtain an approximate solution to the problem. This method involves a large number of variables and iterations, has the disadvantage of slow convergence speed, and the result may not be close to the optimal solution. Summary of the invention

[0007] In view of this, the object of the present invention is to provide a method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network (GNN).

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] A method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network, the method comprising the following steps:

[0010] S1. Establish a dual IRS-assisted wireless energy-carrying communication system model, define the corresponding communication channel, the corresponding reflection phase shift matrix, the corresponding received signal representation and the corresponding receiving rate;

[0011] S2. Establish a secure communication optimization problem P0 under the dual IRS-assisted wireless energy-carrying communication system model with the goal of maximizing the average information collection user and information rate and satisfying the corresponding constraints;

[0012] S3, using the pilot estimation method to obtain all channel state information of the dual IRS assisted wireless energy communication system model and construct a corresponding pilot information vector;

[0013] S4, using graph neural network GNN to simulate the optimal mapping of pilot information vector to base station beamforming and IRS reflection phase shift, transforming the secure communication optimization problem P0 into P1;

[0014] S5, introduce Lagrange multiplier vector to further transform the secure communication optimization problem P1 into P2;

[0015] S6. Establishing a graph neural network GNN model under the dual IRS assisted wireless energy communication system model, which includes an input layer, a D update layer and an output layer;

[0016] S7. Train the graph neural network GNN model. After the training is completed, the base station beamforming and IRS reflection phase shift are solved with the secure communication optimization problem P2 as the objective function under the premise of given Lagrange multipliers.

[0017] Further, in step S1, the dual IRS-assisted wireless energy-carrying communication system model includes a base station BS, intelligent reflection surfaces IRS1 and IRS2, K information collection users U1, U2, ..., U K And N energy harvesting users E1, E2, …, E N , the base station BS has M transmitting antennas, all users have single antennas, IRS1 is deployed near the BS and the energy collection user, and it has L1 reflection units, IRS2 is deployed near the energy collection user, and it has L2 reflection units; the BS sends information to K information collection users and energy to N energy collection users at the same time, among which IRS1 reflects the signal transmitted by the base station to all users, and reflects the signal to IRS2 at the same time; IRS2 reflects the signal received from the base station and IRS1 to all users; among which,

[0018] The channels from BS to IRS1 and IRS2 are represented as and The channel from IRS1 to IRS2 is represented as The channels from BS, IRS1 and IRS2 to the information collection user are represented as and The channels from BS, IRS1 and IRS2 to the energy harvesting user are represented as and Among them, 1≤k≤K, 1≤n≤N, C x×y represents a set of complex-valued matrices of dimension x×y;

[0019] The reflection phase shift matrix of IRS1 is expressed as:

[0020]

[0021] represents the reflection coefficient of the l1th reflection unit, represents the reflection amplitude, Represents the reflection phase shift of the l1th reflection unit Where Q is the phase shift interval of a single element of the IRS;

[0022] The reflection phase shift matrix of IRS2 is expressed as:

[0023]

[0024] in, Represents the reflection coefficient and reflection amplitude of the l2th reflection unit Reflection Phase Shift

[0025] The base station's transmission signal is expressed as:

[0026]

[0027] in, For information collection user U k Confidential information of k ∈C M×1 is the corresponding beamforming matrix; For energy harvesting users E n Information, where z n ∈C M×1 is the beamforming matrix, and It follows a complex Gaussian distribution with a mean of 0 and a variance of 1;

[0028] Then the information collection user U k and energy harvesting users E n The received signals are expressed as:

[0029] y k =[h d,k +G1θ1h 1,k +G2θ2h 2,k +G1θ1D H θ2h 2,k ] T x+u k ,1≤k≤K

[0030] y n =[g d,n +G1θ1g 1,n +G2θ2g 2,n +G1θ1D H θ2g2,n ] T x+u n ,1≤n≤N

[0031] where u k and u n are the receiving noise at the information collection user and the energy collection user, respectively, and δ represents the variance of the complex Gaussian distribution;

[0032] Then U k The achievable rate at which the kth confidential information is received is:

[0033]

[0034] E n The achievable eavesdropping rate of the kth confidential information received at is:

[0035]

[0036] Energy harvesting users n The energy that can be collected is:

[0037]

[0038] Among them, η∈(0,1] is the energy collection efficiency.

[0039] Furthermore, the direct channel h d,k and d,n Subject to Rayleigh fading:

[0040]

[0041] in, They represent the large-scale fading factors of the channel, vec(·) represents matrix vectorization;

[0042] The reflection channels G1, G2, and D are subject to Rice fading, and the Rice channel is expressed as:

[0043]

[0044] Where LOS is the line-of-sight part of the channel, NLOS is the non-line-of-sight part of the channel, ε is the Rice factor, and β is the path loss of the channel; Obey standard complex Gaussian distribution The path loss coefficient β is related to the physical straight-line distance d between the two ends of the channel, is a function of azimuth and elevation; where Indicates that the user U is collected from information kThe azimuth and elevation angles to IRS1, then the steering vector of the l1th reflection unit of IRS1 is It is expressed as:

[0045]

[0046] Among them, d IRS1 is the distance between two adjacent reflection units in IRS1, λ c is the carrier wavelength, i1(l1)=mod(l1-1,L), It means rounding down, L and W are the length and width of IRS1, L×W=L1, then

[0047] Assume (x k ,y k ,z k ) is the information collection user U k The three-dimensional geographic coordinates, (x IRS1 ,y IRS2 ,z IRS3 ) is the three-dimensional geographic coordinate of IRS1, then the azimuth and elevation are expressed as:

[0048]

[0049] in, For U k Straight-line distance to IRS1;

[0050] Similarly, IRS1 to energy collection user E n Channel g 1,n , IRS2 to information collection user U k Channel h 2,k , IRS2 to energy collection user E n Channel g 2,n The generation method of is consistent with this;

[0051] For the channel G1 from BS to IRS1, let is the azimuth and elevation angle of the reflection unit on IRS1 reaching the BS, and the steering vector at the BS is expressed as:

[0052]

[0053] Among them, d BS is the distance between two adjacent BS antennas, let are the azimuth and elevation angles from IRS1 to BS, then It is expressed as:

[0054]

[0055] Assume (x BS ,y BS ,z BS ) is the three-dimensional geographic coordinates of the BS, then the azimuth and elevation are expressed as:

[0056]

[0057] Among them, d BS,IRS1 is the straight-line distance from BS to IRS1;

[0058] The generation method of the channel G2 from BS to IRS2 and the reflection channel D from IRS1 to IRS2 is consistent with this.

[0059] Further, in step S2, by jointly optimizing the reflection phase shift matrix θ1 of IRS1, the reflection phase shift matrix θ2 of IRS2 and the beamforming matrix {w k},{z n}, maximize the total information rate R of all information collection users under the constraints of the maximum eavesdropping rate and minimum energy of the energy collection users, the modulus-one constraint of the reflection coefficient of the IRS, and the maximum transmission power of the base station. k The security communication optimization problem P0 is expressed as:

[0060]

[0061] st:

[0062]

[0063] Among them, P BS is the transmit power constraint at the BS, is the maximum eavesdropping rate limit of the energy harvesting user, and ζ is the minimum energy harvesting limit of the energy harvesting user.

[0064] Further, in step S3, for the information collection user U k , let user U k The pilot signal sent is x k ∈C K×1 , where the pilot length is taken as the number of information collection users K, and the pilot signals sent by different users are orthogonal to each other, that is:

[0065]

[0066] Among them, P U is the power of the pilot transmitted by the information collection user;

[0067] Then the signal received by BS at time t is:

[0068]

[0069] Where t=1,2,...,τ u , τ u To estimate the number of pilots required, N t is the BS receiving noise, H k,t =h d,k +G1θ 1,t h 1,k +G2θ 2,t h 2,k +G1θ 1,t D H θ 2,t h 2,k ,1≤k≤K;

[0070] Using the orthogonality of the pilot signal, right-multiply x at the BS k ,get:

[0071]

[0072] in,

[0073] The number of uses is τ u The pilot information is used as the input of GNN:

[0074]

[0075] Similarly, energy harvesting user E n Using the same pilot estimation method, we can get E at the BS: n The number is τ u Pilot information:

[0076]

[0077] Then we get the information collection user pilot Y U and energy harvesting user pilot Y E , which are respectively expressed as:

[0078] Y U =[y1,y2,..,y k ],1≤k≤K

[0079] Y E =[y1,y2,..,y n ],1≤n≤N

[0080] Information collection user pilot Y U and energy harvesting user pilot Y E As the input feature vector of the graph neural network GNN.

[0081] Further, in step S4, the secure communication optimization problem P0 is transformed into P1:

[0082]

[0083] st:

[0084]

[0085] Among them, ({w k},{z n},θ1,θ2)=g(Y U ,Y E ) represents the use of graph neural network GNN to simulate receiving pilot Y from BS U and Y E Optimal mapping to BS beamforming and IRS reflection phase shifts.

[0086] Further, in step S5, the Lagrange multiplier vector is first introduced:

[0087]

[0088] By transferring the constraints C2 and C3 into the objective function, we obtain a partial primal-dual optimization problem, thereby transforming the secure communication optimization problem P1 into P2:

[0089]

[0090] st:

[0091]

[0092] It is solved by alternating optimization: first fix the dual variables to optimize the system variables to minimize the objective P2; then fix the system variables and optimize the dual variables to maximize P2;

[0093] For P2, given the system variables, the Lagrange multiplier and Update in the direction of gradient ascent based on the subgradient:

[0094]

[0095] Where l is the number of iterations and κ ≥ 0 is the update step size.

[0096] Furthermore, in step S6, a GNN model is designed to abstract the wireless communication environment into a graph G = (V, E), where V = [IRS1, IRS2, U1, ..., U k ,E1,...,E n ] is the set of nodes in the graph, and each node v∈V corresponds to a representation vector xv , v∈V, where the representation vector of the node is:

[0097]

[0098] The representation vector of the node is received by the BS with pilot information Y U and Y E get;

[0099] In GNN, the feature vector Y U and Y E After normalization and standardization, it is converted into a graph node representation vector. All representation vectors of the previous layer are used as input, and the representation vector is updated layer by layer. After multiple layers, the representation vector of each node is used to design the beamforming vector and reflection coefficient.

[0100] In GNN, the graph representation vector is learned through the input layer, where U k The input feature of the node is the received pilot y k , E n The input feature of the node is the received pilot y n , the input features of the IRS1 node By Y U and Y E The input features of IRS2 node are obtained by weighted averaging. By Y U and Y E Take the weighted average to get;

[0101] D updates the layer to update the representation vector, and gets

[0102] The output result is obtained through the output layer, first through a fully connected linear layer Then we will represent the vector The phase shift of the beamforming and IRS output to the BS is mapped by the normalization layer.

[0103] Furthermore, in the input layer of GNN, the pilot information y k The real and imaginary components of are separated and then converted into column vectors to obtain:

[0104]

[0105] Among them, the information related to the node can be embedded in the column vector:

[0106]

[0107] Among them, loc k Collect the user's three-dimensional coordinate parameters for information;

[0108] Then Through a deep neural network layer As U k Initial representation vector of the node

[0109]

[0110] For energy harvesting users, the node representation vector is obtained in the same way:

[0111]

[0112] For an IRS node, its representation vector is formed by aggregating the representation vectors of the information collection user and the energy collection user:

[0113]

[0114] where ψ0(·) is an element-wise average function is a deep neural network layer;

[0115] The representation vector of IRS2 node is obtained in the same way as IRS1:

[0116]

[0117] The representation vector obtained by the input layer Contains the feature vectors of all nodes and edges and passes them to the D update layer;

[0118] The update of the representation vectors of all nodes in the update layer is based on the aggregation of the previous representation vector of itself and the representation vectors of its adjacent nodes. The specific mathematical expression is:

[0119]

[0120] Among them, V(v) represents the set of nodes adjacent to node v. is the aggregation function of the dth update layer, is a combination function;

[0121] The aggregation function of GNN is expressed as:

[0122]

[0123] Among them, ψ(·) is a function that is invariant to the input permutation, and Use deep neural networks;

[0124] For the IRS1 node, the representation vector comes from:

[0125]

[0126] in, It is a multi-layer perceptron, which is used to integrate the node representation vector and introduce the nonlinear activation function relu; ψ1(·) is implemented using the element pool averaging function:

[0127] ψ1(x1,x2,..,x k )=mean(x1,x2,..,x k )

[0128] The representation vector of the IRS2 node is updated from the neighboring nodes, including the representation vectors from the IRS1 node and from the information collection user and eavesdropping user nodes:

[0129]

[0130] in, is the MLP connection layer, and the element-wise average function ψ1(·) is selected as the aggregation function;

[0131] For the information collection user node, its representation vector is updated as follows:

[0132]

[0133] in, is an MLP connection layer, and selects the element pooling maximization function ψ2(·) as the aggregation function:

[0134] ψ2(x1,x2,..,x k )=max(x1,x2,..,x k )

[0135] For energy harvesting user nodes, the update of the representation vector is as follows:

[0136]

[0137] in, is the MLP connection layer, and selects the element pooling maximization function ψ2(·) as the aggregation function;

[0138] Updated node representation vector Passed to the output layer, the IRS1 reflection coefficient matrix is ​​finally obtained IRS2 reflection coefficient matrix and the beamforming matrix W at the BS = [w1,w2,..,w K ]∈C M×K and Z = [z1,z2,..,z N ]∈C M×N, while satisfying the unit mode constraints of v1 and v2 and the total transmission power constraint of the BS;

[0139] Will represent the vector and Pass through a linear layer with 2M units respectively get:

[0140]

[0141] The beamforming matrix is ​​then output through a normalization layer as follows:

[0142]

[0143]

[0144] Where W(i:j,u:v:) represents a submatrix of W, which is constructed by the indexes from the i-th row to the j-th row and the u-th column to the v-th column of W;

[0145] Pass a linear layer with 2L1 units Output As shown below:

[0146]

[0147] Then, the reflection phase shift of each reflection unit in IRS1 is obtained through the normalization layer, as follows:

[0148]

[0149] Where, [V1] ij represents the element in the i-th row and j-th column of the matrix, symbol represents the index subvector from i to j;

[0150] Following the same principle, Pass a linear layer with 2L2 units Output As shown below:

[0151]

[0152] Then, the reflection phase shift of each reflection unit in IRS2 is obtained through the normalization layer, as follows:

[0153]

[0154] Through the output layer, beamforming at the BS and reflection phase shift at the IRS are obtained.

[0155] Further, in step S6, the GNN model is trained with the goal of minimizing the Lagrangian relaxation function P2 under the premise of a given Lagrangian multiplier, and its loss function is:

[0156]

[0157] Among them, L is the loss function of the GNN model in the lth iteration process.

[0158] The beneficial effects of the present invention are:

[0159] The present invention proposes a secure communication optimization method for a dual IRS-assisted wireless energy-carrying communication system based on a GNN graph neural network. First, the Lagrange multiplier method is used to introduce the Lagrange multiplier vector to introduce the constraint condition into the objective function. A partial primal dual optimization problem is obtained. Then, the IRS and multiple users are modeled, a GNN connection graph is constructed, and the objective function with Lagrange multipliers is converted into a GNN loss function. Next, a pilot transmission is performed to obtain the training input of the GNN. The GNN completes the mapping from the pilot to the BS beamforming and the IRS reflection phase shift through training to maximize the system objectives, such as the total information transmission rate of the information collection user. The trained GNN has good generalization ability and can adapt to communication scenarios with different numbers of users. Compared with the traditional method of using channel estimation and non-convex optimization, a higher information transmission rate is obtained under the premise of meeting the needs of energy collection users and information collection users.

[0160] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0161] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0162] Figure 1 A schematic diagram of the structure of a dual IRS-assisted wireless energy-carrying communication system model provided by the present invention;

[0163] Figure 2 A GNN connection diagram of the dual IRS-assisted wireless energy-carrying communication system safety communication optimization method provided by the present invention;

[0164] Figure 3 A schematic diagram of the GNN structure of the dual IRS-assisted wireless energy-carrying communication system security communication optimization method provided by the present invention;

[0165] Figure 4 The present invention provides a schematic diagram of the overall process of the dual IRS-assisted wireless energy-carrying communication system security communication optimization method. DETAILED DESCRIPTION

[0166] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0167] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0168] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0169] See also Figure 1 to Figure 4 , which is a secure communication optimization method for dual IRS-assisted wireless energy-carrying communication system based on graph neural network.

[0170] Example

[0171] This embodiment provides a specific implementation method of a dual IRS-assisted wireless energy-carrying communication system security communication optimization method based on a graph neural network, which specifically includes the following contents:

[0172] Step 1: Establish a dual IRS-assisted wireless energy communication system model and set initial parameters, including the number of training samples, system parameters (number of BS antennas, number of IRS reflection units, number of information collection users and number of eavesdropping users), input flag input_flag, Rayleigh channel factor, Rice channel factor, noise power, channel transmission power, etc. Figure 1 As shown, the dual IRS-assisted wireless energy-carrying communication system model includes a base station BS, two IRSs, and K information collection users (denoted as U1, ..., U k ) and N energy harvesting users (denoted as E1, ..., E n ). The base station BS has M transmitting antennas, and all users have single antennas. IRS1 is deployed on a high-rise building close to the BS and energy collection users, and has L1 reflection units. IRS2 is deployed on a high-rise building close to the information collection users, and has L2 reflection units. The BS sends information to K information collection users and energy to N energy collection users at the same time. In order to meet the charging requirements, energy collection users are usually closer to the base station BS than information collection users, which makes energy collection users potential information eavesdroppers.

[0173] In order to meet the energy transmission requirements and improve the security of wireless information transmission, two IRSs can be used to assist communication in a coordinated manner. Specifically, IRS1 reflects the signal transmitted by the base station to all users, and at the same time reflects the signal to IRS2; IRS2 reflects the signal received from the base station and IRS1 to all users. Compared with the traditional single IRS auxiliary system communication, the dual IRS cooperative auxiliary system communication can not only create a new communication link for users, improve the information transmission rate and energy transmission power, but also the energy harvesting users receive the signals reflected by the dual IRS, and these signal energies act as artificial noise, thereby ensuring the security of information transmission.

[0174] More specifically, in the established dual IRS-assisted wireless energy-carrying communication system model, the reflection channel obeys Ricean fading, the direct channel obeys Rayleigh fading, and the IRS phase shift adopts a discrete form.

[0175] The baseband equivalent channel response is expressed as:

[0176] The channels from BS to IRS1 and IRS2 are represented as and The channel from IRS1 to IRS2 is represented as The channels from BS, IRS1 and IRS2 to the information collection user are represented as and The channels from BS, IRS1 and IRS2 to the energy harvesting user are denoted as g d,n ∈C M×1 , and Among them, 1≤k≤K, 1≤n≤N, C x×y Represents a collection of complex-valued matrices of dimension x-by-y.

[0177] The reflection phase shift matrix of IRS1 is expressed as represents the reflection coefficient of the l1th reflection unit, where is the reflection amplitude, It is its reflected phase shift, which has two main forms, one is continuous phase shift The other is discrete phase shift:

[0178]

[0179] Where Q is the phase shift interval of a single element on the IRS. In practical applications, it is very expensive to implement continuous phase shift in hardware, so discrete phase shift is more widely used. However, it is also important to study continuous phase shift because it is the upper limit of discrete phase shift.

[0180] The reflection phase shift matrix of IRS2 is expressed as represents the reflection coefficient of the l2th reflection unit, and the reflection amplitude is β 2l ∈[0,1], the reflection phase shift is:

[0181]

[0182] This embodiment only considers the reflection phase shift, so and

[0183] The base station's transmission signal is expressed as in For information collection user U k Confidential information of k ∈C M×1 is the corresponding beamforming matrix; For energy harvesting users E n Information, where z n ∈C M×1 is the beamforming matrix, and They all obey a complex Gaussian distribution with mean 0 and variance 1.

[0184] Then the information collection user U k and energy harvesting users E n The received signal can be expressed as:

[0185] y k =[h d,k +G1θ1h 1,k+G2θ2h 2,k +G1θ1D H θ2h 2,k ] T x+u k ,1≤k≤K

[0186] y n =[g d,n +G1θ1g 1,n +G2θ2g 2,n +G1θ1D H θ2g 2,n ] T x+u n ,1≤n≤N

[0187] where u k and u n are the receiving noises at the information collection user and the energy collection user, respectively, and they obey the complex Gaussian distribution with a mean of 0, that is,

[0188] It can be obtained that the achievable rate of receiving the kth confidential information at Uk is:

[0189]

[0190] The achievable eavesdropping rate of the kth confidential information received at En is:

[0191]

[0192] Energy collection The energy that can be collected by user En is:

[0193]

[0194] Among them, η∈(0,1] is the energy collection efficiency.

[0195] Assume h d,k , g d,n is an uncorrelated Rayleigh fading channel, that is:

[0196]

[0197] in, Respectively represent the large-scale fading factor of the channel, and vec(·) represents matrix vectorization. The remaining channels are all Rice fading channels. The Rice channel is obtained by the following formula:

[0198]

[0199] Among them, LOS is the line-of-sight part of the channel, NLOS is the non-line-of-sight part of the channel, ε is the Ricean factor, and β is the path loss of the channel. It obeys the standard complex Gaussian distribution, that is, The path loss coefficient β is set to 30+22log(d), where d is the physical straight-line distance between the two ends of the channel. is a function of azimuth and elevation. Specifically, let Indicates that the user U is collected from information k The azimuth and elevation angles to IRS1. Then the steering vector of the l1th reflection unit of IRS1 is It can be expressed as:

[0200]

[0201] Among them, d IRS1 is the distance between two adjacent reflection units in IRS1, λ c is the carrier wavelength, i1(l1)=mod(l1-1,L), ( To round down, L and W are the length and width of IRS1, L×W=L1). Therefore, Assume (x k ,y k ,z k ) is the information collection user U k The three-dimensional geographic coordinates, (x IRS1 ,y IRS2 ,z IRS3 ) is the three-dimensional geographic coordinate of IRS1. Then the azimuth and elevation can be expressed as:

[0202]

[0203] in, For U k Straight line distance to IRS1. IRS1 to energy harvesting user E n Channel g 1,n , IRS2 to information collection user U k Channel h 2,k , IRS2 to energy harvesting user E n Channel g 2,n can be generated in the same way.

[0204] For the channel G1 from BS to IRS1, let is the azimuth and elevation angle of the reflection unit on IRS1 to the BS, then the steering vector at the BS can be expressed as:

[0205]

[0206] Among them, d BS is the distance between two adjacent BS antennas, let are the azimuth and elevation angles from IRS1 to BS, then It can be expressed as:

[0207]

[0208] Assume (x BS ,y BS ,z BS ) is the three-dimensional geographic coordinates of the BS, the azimuth and elevation can be expressed as:

[0209]

[0210] Among them, d BS,IRS1 is the straight-line distance from BS to IRS1. The channel G2 from BS to IRS2 and the reflection channel D from IRS1 to IRS2 can be generated in the same way.

[0211] The second step is to establish a secure communication optimization problem; the present invention jointly optimizes the reflection phase shift matrix θ1 of IRS1, the reflection phase shift matrix θ2 of IRS2 and the beamforming matrix {w k},{z n}, ensuring that the eavesdropping rate of the energy collection user is less than or equal to the maximum eavesdropping rate, the received energy is not less than the minimum energy requirement, the reflection coefficient of the IRS satisfies the modulo-one constraint, and the base station does not exceed the maximum transmission power constraint to maximize the total information rate R of all information collection users k Therefore, the problem under consideration is described as follows:

[0212]

[0213] Among them, P BS is the transmit power constraint at the BS, is the maximum eavesdropping rate limit of the energy harvesting user, and ζ is the minimum energy harvesting limit of the energy harvesting user. Since the objective function and the system variables in constraints C2 and C3 are highly coupled, and constraints C4 and C5 are non-convex, it is difficult to find the global optimal solution to the problem.

[0214] The third step is to perform channel estimation and use the pilot method to obtain the channel state information (CSI) to extract the channel parameters of the information collection user and the energy collection user, thereby optimizing the objective function P0.

[0215] Due to the large number of reflection units in massive antennas and dual IRS, the complexity of channel estimation and the required pilot overhead increase exponentially. Many researchers are committed to designing efficient channel estimation schemes for different IRS architectures, aiming to achieve higher channel estimation accuracy with lower pilot overhead.

[0216] The present invention uses GNN graph neural network to simulate the optimal mapping from BS received pilot to BS beamforming and IRS reflection phase shift. The representation vector of GNN can be obtained from the result of channel estimation, that is, estimated CSI, or from the pilot information received by BS, because the purpose of channel estimation by pilot method is to design a reasonable estimation algorithm to recover more accurate channel CSI from pilot information, but GNN has the characteristics of information integration. Since CSI information is implicit in the pilot information, GNN can embed the pilot information as feature information into the graph representation vector. Doing so not only reduces the time overhead caused by the channel estimation algorithm, but also minimizes the calculation error of the objective function caused by the error generated by the channel estimation algorithm.

[0217] Many documents have proposed channel estimation schemes for large-scale MIMO communication systems assisted by dual IRS collaboration. Since the BS and IRS are generally fixed in position, the channel is generally high-dimensional and quasi-static, while the channels of information collection users and energy collection users need to move frequently, so they are low-dimensional and time-varying. For quasi-static channels, frequent estimation is not required, while for time-varying channels, frequent estimation is required. However, in combination with the objective function, the present invention does not need to obtain the estimated value of each channel, but only needs to obtain the total channel information from the BS to the energy collection user and the information collection user. For the information collection user U k In terms of BS and U k The total channel between k =h d,k +G1θ1h 1,k +G2θ2h 2,k +G1θ1D H θ2h 2,k ,1≤k≤K. Therefore, the present invention adopts the simplest pilot estimation method.

[0218] Since information is collected from user U k and energy harvesting users E n The pilot transmission method is the same as that of k Assume that user U k The pilot signal sent is x k ∈C K×1 , where the pilot length is taken as the number of information collection users K, and the pilot signals sent by different users are orthogonal to each other. That is:

[0219]

[0220] Among them, P U is the power of the pilot signal transmitted by the information collection user. The signal received by the BS at time t is:

[0221]

[0222] Where t=1,2,...,τ u , τ u To estimate the number of pilots required, N t is the BS receiving noise, H k,t =h d,k +G1θ 1,t h 1,k +G2θ 2,t h 2,k +G1θ 1,t D H θ 2,t h 2,k ,1≤k≤K. Using the orthogonality of the pilot signal, multiply x on the right at the BS. k ,get:

[0223]

[0224] in, Note that due to H k,t It is a complex combined channel. It is a difficult problem to obtain the CSI of each channel. The traditional method requires step-by-step quasi-static channel estimation from BS to IRS, and then use the estimated results to perform cascade channel estimation between IRSs, and finally perform user channel estimation. The present invention uses GNN to realize the mapping from channel CSI to BS beamforming and IRS reflection phase shift, and converts the target optimization function P0 to P1. The present invention uses GNN to realize the mapping from channel CSI to BS beamforming and IRS reflection phase shift, and does not need to obtain accurate channel estimation results. Since the pilot information already contains the complete channel state information CSI, the pilot is embedded into the GNN as a feature vector. Specifically, the number of τ is used u The pilot information is used as the input of GNN.

[0225]

[0226] Energy harvesting users n Using the same pilot estimation method, we can get E at the BS: n The channel CSI is expressed as follows:

[0227]

[0228] After the above steps, the input feature vector of GNN is obtained, that is, the information collection user pilot Y U and energy harvesting user pilot Y E , the two are expressed as follows:

[0229] Y U =[y1,y2,..,yk ],1≤k≤K

[0230] Y E =[y1,y2,..,y n ],1≤n≤N

[0231] The fourth step is to use the graph neural network GNN to simulate receiving the pilot signal Y from the BS U and Y E The optimal mapping to BS beamforming and IRS reflection phase shift transforms the objective optimization problem P0 into P1:

[0232]

[0233] st:

[0234]

[0235] In the fifth step, since the system variables in the constraints C2 and C3 in this problem are highly coupled and the constraints are non-convex, the constraints are difficult to handle directly. Therefore, the Lagrange multiplier method is used to introduce the constraints into the target P1.

[0236] First, introduce the Lagrange multiplier vector:

[0237]

[0238] By transferring the constraints C2 and C3 into the objective function, we obtain a partial primal-dual optimization problem as follows:

[0239]

[0240] st:

[0241]

[0242] In general, alternating optimization is the preferred method for solving the maximum-minimum problem. First, the dual variable is fixed to optimize the system variable to minimize the target P2. Then the system variable is fixed and the dual variable is optimized to maximize P2. For P2, given the system variable, the Lagrange multiplier and The update can be performed based on the subgradient in the direction of gradient ascent.

[0243]

[0244] Where; is the number of iterations, κ≥0 is the update step size.

[0245] The sixth step is to design the GNN model, such as Figure 2The neural network architecture diagram of GNN shown in the figure abstracts the wireless communication environment into a graph G = (V, E), V = [IRS1, IRS2, U1, ..., U k ,E1,...,E n ] is a set of nodes in the graph, for example, IRS and information collection users and energy collection users are regarded as a node in the graph. Each node v∈V has its corresponding representation vector x v , v∈V. For example, the representation vector of a node is as follows:

[0246]

[0247] The representation vector of the node is received by the BS with pilot information Y U and Y E In GNN, the feature vector Y U and Y E After normalization and standardization, it is converted into a graph node representation vector, and all the representation vectors of the previous layer are used as input to update the representation vector layer by layer. After multiple layers, the representation vector of each node will contain enough information to design the beamforming vector and reflection coefficient. Specifically, GNN is trained in such a way that the phase shift of IRS1 can be obtained by node IRS1, the phase shift of IRS2 can be obtained by node IRS2, and the BS beamforming matrix can be obtained by node U k and E n If there is a relationship between nodes representing vector updates, a directed edge is used to connect them, where the edge from IRS to information collection users and energy collection users represents the existence of data transmission and energy transmission relationships, the edge between information collection users represents the existence of link interference, and the edge from energy collection users to information collection users represents the existence of interference, the edge between energy collection users represents the existence of energy interference, and the edge from information collection users to energy collection users represents the existence of interference.

[0248] Compared with fully connected neural networks, GNNs more naturally capture the interactions between information collection users, energy collection users, and IRSs. In particular, the update of each information collection user node is a function of all its neighboring nodes, energy collection user nodes, and IRS nodes, which enables GNNs to learn to avoid interference that increases the information rate of information collection users. The update of each energy collection user node is a function of all information collection user nodes, all its neighboring energy collection user nodes, and IRS nodes, which enables GNNs to learn to avoid energy collection users eavesdropping on information collection users while meeting energy collection requirements. The representation vector update of the IRS1 node comes from all energy collection users and information collection users, thereby better meeting the needs of all users. The representation vector update of the IRS2 node comes from all energy collection users and information collection users as well as the IRS1 node, which means that IRS2 needs to receive reflection information from the IRS1 node to collaboratively complete the service. The layer structure of the graph neural network is represented as follows Figure 3 shown.

[0249] The entire GNN consists of three layers, and the graph representation vector is learned through the input layer, where U k The input feature of the node is the received pilot y k , where E n The input feature of the node is the received pilot y n , the input features of IRS1 node are Y U and Y E The weighted average is obtained, denoted as Similarly, the input features of IRS2 node are composed of Y U and Y E The weighted average is obtained, denoted as D updates the layer to update the representation vector, and gets The output result is obtained through the output layer. Specifically, it is first obtained through a fully connected linear layer Then we will represent the vector The phase shift of the beamforming and IRS output to the BS is mapped by the normalization layer.

[0250] 1) Input layer: Since TensorFlow does not support complex-valued operations, the pilot information y is first k The real and imaginary components of are separated and then converted into column vectors to obtain:

[0251]

[0252] The representation vector of each node in GNN can be obtained by embedding the useful information of the node. Therefore, any useful information related to the node can be embedded into the representation vector to obtain sufficiently detailed and complete effective information. For example, adding information to collect the user's three-dimensional geographic coordinates:

[0253]

[0254] where loc k Collect the user's three-dimensional coordinate parameters for information.

[0255] Then Through a deep neural network layer As U k Initial representation vector of the node

[0256]

[0257] For energy harvesting users, the node representation vector is obtained in the same way:

[0258]

[0259] For an IRS node, its representation vector is formed by the aggregation of the representation vectors of the information collection user and the energy collection user. That is:

[0260]

[0261] where ψ0(·) is an element-wise average function:

[0262]

[0263] The reason for this design is that the pilot signals sent by the information collection user and the energy collection user already contain the information of the IRS reflection vector, and the IRS information contained in the pilot signals sent by each user is equivalent. is a deep neural network layer.

[0264] The representation vector of IRS2 node is obtained in the same way as IRS1:

[0265]

[0266] Representation vector Contains the feature vectors of all nodes and edges, and is then passed to the D update layer.

[0267] D Update layer:

[0268] The update of the representation vectors of all nodes in the update layer is based on the aggregation of the previous representation vector of itself and the representation vectors of its adjacent nodes. The specific mathematical expression is:

[0269]

[0270] Here V(v) represents the set of nodes adjacent to node v. is the aggregation function of the dth update layer, is a combination function.

[0271] The key to designing GNN is to choose appropriate aggregation functions and combination functions so that it has good scalability. The selection of aggregation functions is as follows:

[0272]

[0273] where ψ(·) is a function that is invariant to the input permutation, such as element pool maximization or element pool averaging function. and Both can be implemented using deep neural networks.

[0274] For the IRS1 node, the representation vector comes from:

[0275]

[0276] It is a layer of MLP. MLP (Mutil-Layer-Perceptron, MLP) multi-layer perceptron is used to integrate node representation vectors and introduce nonlinear activation function relu, so that GNN can learn more complex nonlinear feature mapping relationships. ψ1(·) is implemented using element pooling averaging function, as shown below:

[0277] ψ1(x1,x2,..,x k )=mean(x1,x2,..,x k )

[0278] The representation vector of the IRS2 node is updated from the neighboring nodes, including the representation vectors from the IRS1 node and from the information collection user and eavesdropping user nodes:

[0279]

[0280] in is an MLP connection layer, and the element-wise average function ψ1(·) is selected as the aggregation function, which is consistent with the fact that IRS needs to adjust its reflection vector to serve the information collection users and energy collection users as much as possible and meet the information eavesdropping rate and energy collection constraints of the energy collection users.

[0281] For the information collection user node, its representation vector is updated as follows:

[0282]

[0283] in is an MLP connection layer, and the element pooling maximization function ψ2( ) is selected as the aggregation function, which is consistent with the fact that multi-user interference is usually dominated by the user with the strongest signal.

[0284] ψ2(x1,x2,..,x k )=max(x1,x2,..,x k )

[0285] For energy harvesting user nodes, the update of the representation vector is as follows:

[0286]

[0287] in is an MLP connection layer, and the element-wise pooling maximization function ψ2(·) is selected as the aggregation function, which is consistent with the fact that multi-signal interference is usually dominated by the strongest signal.

[0288] Output layer:

[0289] After the update layer, the node representation vector of GNN update Passed to the output layer. Finally, the IRS1 reflection coefficient matrix is ​​obtained IRS2 reflection coefficient matrix and the beamforming matrix W at the BS = [w1,w2,..,w K ]∈C M×K and Z = [z1,z2,..,z N ]∈C M×N , while satisfying the unit mode constraints of v1 and v2 and the total transmission power constraint of the BS.

[0290] Will represent the vector and Pass through a linear layer with 2M units respectively get:

[0291]

[0292] The beamforming matrix is ​​then output through a normalization layer as follows:

[0293]

[0294] Where W(i:j,u:v:) represents a submatrix of W, which is constructed by the indices from the i-th row to the j-th row and the u-th column to the v-th column of W.

[0295] Pass a linear layer with 2L1 units Output As shown below:

[0296]

[0297] Then, the reflection phase shift of each reflection unit in IRS1 is obtained through the normalization layer, as follows:

[0298]

[0299] Where [V1] ij represents the element in the i-th row and j-th column of the matrix, symbol Represents the subvector of indices from i to j.

[0300] Following the same principle, Pass a linear layer with 2L2 units Output As shown below:

[0301]

[0302] Then, the reflection phase shift of each reflection unit in IRS2 is obtained through the normalization layer, as follows:

[0303]

[0304] Through the output layer, beamforming at the BS and reflection phase shift at the IRS are obtained.

[0305] The GNN-based machine learning algorithm in the above process is shown in Table 1 below:

[0306] Table 1

[0307]

[0308]

[0309] Step 7: Train the GNN model, output the training results, and calculate the objective function using the obtained base station beamforming vector, IRS phase shift, and related channel parameters. The training process is:

[0310] Input: Received pilot information at the BS as training data set and the learning step size of the Lagrange multiplier

[0311] Output: trained GNN model;

[0312] Step 1: Set the number of samples to 1024, system parameters (number of BS antennas, number of IRS reflection units, number of information collection users and number of energy collection users), input flag input_flag, Rayleigh, Rice channel factors, noise power, etc.

[0313] Step 2: Estimate the channel according to the proposed channel estimation scheme and obtain the pilot information:

[0314]

[0315] Step 3: Initialize the Lagrange multipliers:

[0316] Step 4: Set the learning rate, maximum number of iterations I, and maximum number of training times per iteration J. Set the training loss unchanged flag No_change to 0. Perform a GNN training once as training data, obtain the training loss train_loss[0], and set the best training loss best_loss = train_loss[0]; train a certain number of rounds with a small batch of data of size b for the training data set, and perform the following steps in the jth training round of the i-th iteration round:

[0317] Perform multiple pilot transmissions to obtain pilot data as a training data set

[0318] Randomly select a mini-batch β from the training dataset b The BS beamforming and IRS reflection phase shift are used as input for training, and the Lagrangian relaxation loss function at the mini-batch level is obtained. The sub-gradient of the Lagrangian multiplier is calculated. After completing the mini-batch data set training, the Lagrangian multiplier is updated and the next round of training is entered.

[0319] After J rounds of training, use the pilot Perform a training to get the training loss, and compare the current training loss with the optimal loss. If it is less than the optimal loss, it proves that the model performance has been improved, and the optimal loss needs to be updated and the next round of iterations needs to be entered. If it is greater than the optimal loss, it proves that the model performance has not been improved, and it is necessary to determine whether there has been no improvement after 5 consecutive rounds of iterations. If it proves that the model performance has reached the optimal level, the training needs to be terminated. If not, continue to the next round of iterations. Finally, if the maximum number of iterations is reached, the training ends.

[0320] The specific training process is shown in Table 2:

[0321] Table 2

[0322]

[0323]

[0324] The input of the algorithm is the received pilot information at the BS as the training data set And the learning step size of the Lagrange multiplier. GNN trains a certain number of rounds of small batches of data of size b for the training data set. The BS beamforming and IRS reflection phase shift strategies generated by the GNN model are obtained, and then the value of the Lagrange relaxation function is calculated using the current Lagrange multiplier. Then the GNN model updates the model learning parameters according to the Lagrange relaxation function, which is the loss function. During each small batch training round, the sub-gradient of the Lagrange multiplier is calculated. Finally, after each training round, the Lagrange multiplier is updated based on the sub-gradient. Among them, σ c (·) is a mapping function that maps the data in the constraint set to the Lagrange multiplier space.

[0325] After the GNN model is completed, its learning parameters and Lagrange multipliers will be updated to the problem P2 to be stable and satisfy the constraints. Once the GNN model is trained, it can be directly applied to a dual IRS-assisted wireless energy communication example in a similar scenario without considering constraints and Lagrange multipliers.

[0326] Use the trained GNN model to optimize the given dual variable λ (1) and λ (2) System variables {w k},{z n}, θ1, θ2, GNN is used to extract the feature vector of each node from the BS received pilot and learn the beamforming at the BS and the IRS reflected beam vector through training to maximize the average information collection user and information rate under the constraints. The GNN model will be trained with the goal of minimizing the Lagrangian relaxation function P2 under the premise of a given Lagrangian multiplier, that is, using P2 as the objective function,

[0327]

[0328] L is the loss function of the GNN model during the lth iteration.

[0329] In the eighth step, this embodiment tests the trained GNN model and tries the output effect of GNN when the number of users and the number of energy collection users are different. Compared with the traditional non-convex optimization method, the GNN-based machine learning method has better generalization ability and stronger scalability, can achieve better confidentiality and security effects, and takes less time.

[0330] In this embodiment, Figure 1The present invention provides a dual IRS assisted SWIPT multi-user communication system model, the scene is set in a city with tall buildings, in the figure, there are direct channels and reflection channels between the base station and the user, the BS directly transmits information and energy to the information collection user and the energy collection user through wireless transmission, or reflects the signal to the user through the IRS. At the same time, the N single-antenna energy collection users in the system will eavesdrop on the information collection user information. Figure 2 This is the GNN connection diagram of the dual IRS-assisted downlink SWIPT multi-user communication system, and the connections between the various nodes of the model are given. The GNN connection diagram is based on the signal transmission and interference in the actual communication environment, abstracting IRS1 and IRS2 as well as information collection users and energy collection users as nodes, and abstracting communication links and interference links as edges associated between nodes. The diagram is a directed graph, in which the IRS node sends data to all user nodes and receives feedback from users. The information collection user receives information from the IRS node and is interfered by the signals of other information collection users and energy collection users; the energy collection user receives energy from the IRS node and needs to eavesdrop on the information of the information collection user, so it is interfered by the signals of all energy collection users and other information collection users other than the target information user. Figure 3 The structure diagram of the GNN model includes an input layer, an update layer, and an output layer. The input layer converts the pilot information into a node representation vector and then passes it to the update layer; the update layer updates the representation vectors of all nodes from the aggregation of itself and adjacent nodes. The update layer can have multiple layers. Generally speaking, the GNN model will adopt a two-layer structure, especially when the number of nodes is small, the effect of multiple layers is not improved much; after the update is completed, in order to obtain the base station beamforming vector and IRS phase shift, the representation vector of the relevant node needs to be passed to the output layer, and the BS beamforming vector and the reflected phase shift of the two IRSs are output under the conditions of satisfying the phase shift modulus constraint and the total power constraint at the BS. Figure 4 This is the training process of the GNN model.

[0331] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on graph neural network, characterized in that: The method comprises the following steps: S1. Establish a dual IRS-assisted wireless energy-carrying communication system model, define the corresponding communication channel, the corresponding reflection phase shift matrix, the corresponding received signal representation and the corresponding receiving rate; S2. Establish a secure communication optimization problem P0 under the dual IRS-assisted wireless energy-carrying communication system model with the goal of maximizing the average information collection user and information rate and satisfying the corresponding constraints; S3, using the pilot estimation method to obtain all channel state information of the dual IRS assisted wireless energy communication system model and construct a corresponding pilot information vector; S4, using graph neural network GNN to simulate the optimal mapping of pilot information vector to base station beamforming and IRS reflection phase shift, transforming the secure communication optimization problem P0 into P1; S5, introduce Lagrange multiplier vector to further transform the secure communication optimization problem P1 into P2; S6. Establishing a graph neural network GNN model under the dual IRS assisted wireless energy communication system model, which includes an input layer, a D update layer and an output layer; S7. Train the graph neural network GNN model. After the training is completed, the base station beamforming and IRS reflection phase shift are solved with the secure communication optimization problem P2 as the objective function under the premise of given Lagrange multipliers.

2. According to claim 1, a method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network is characterized in that: In step S1, the dual IRS-assisted wireless energy-carrying communication system model includes a base station BS, intelligent reflection surfaces IRS1 and IRS2, K information collection users U1, U2, ..., U K And N energy harvesting users E1, E2, …, E N , the base station BS has M transmitting antennas, all users have single antennas, IRS1 is deployed near the BS and the energy collection user, and it has L1 reflection units, IRS2 is deployed near the energy collection user, and it has L2 reflection units; the BS sends information to K information collection users and energy to N energy collection users at the same time, among which IRS1 reflects the signal transmitted by the base station to all users, and reflects the signal to IRS2 at the same time; IRS2 reflects the signal received from the base station and IRS1 to all users; among which, The channels from BS to IRS1 and IRS2 are represented as and The channel from IRS1 to IRS2 is represented as The channels from BS, IRS1 and IRS2 to the information collection user are denoted as h d,k ∈C M×1 , and The channels from BS, IRS1 and IRS2 to the energy harvesting user are denoted as g d,n ∈C M×1 , and Among them, 1≤k≤K, 1≤n≤N, C x×y represents a set of complex-valued matrices of dimension x×y; The reflection phase shift matrix of IRS1 is expressed as: represents the reflection coefficient of the l1th reflection unit, represents the reflection amplitude, Represents the reflection phase shift of the l1th reflection unit l1={1,..,L1}, ​​where Q is the phase shift interval of a single element of the IRS; The reflection phase shift matrix of IRS2 is expressed as: in, Represents the reflection coefficient and reflection amplitude of the l2th reflection unit Reflection Phase Shift The base station's transmission signal is expressed as: in, For information collection user U k Confidential information of k ∈C M×1 is the corresponding beamforming matrix; For energy harvesting users E n Information, where z n ∈C M×1 is the beamforming matrix, and It follows a complex Gaussian distribution with a mean of 0 and a variance of 1; Then the information collection user U k and energy harvesting users E n The received signals are expressed as: y k =[h d,k +G1θ1h 1,k +G2θ2h 2,k +G1θ1D H θ2h 2,k ] T x+u k ,1≤k≤K y n =[g d,n +G1θ1g 1,n +G2θ2g 2,n +G1θ1D H θ2g 2,n ] T x+u n ,1≤n≤N where u k and u n are the receiving noise at the information collection user and the energy collection user, respectively, and δ represents the variance of the complex Gaussian distribution; Then U k The achievable rate at which the kth confidential information is received is: E n The achievable eavesdropping rate of the kth confidential information received at is: Energy harvesting users n The energy that can be collected is: Among them, η∈(0,1] is the energy collection efficiency.

3. According to claim 2, a method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network is characterized in that: Direct Channel d,k and d,n Subject to Rayleigh fading: in, They represent the large-scale fading factors of the channel, vec(·) represents matrix vectorization; The reflection channels G1, G2, and D are subject to Rice fading, and the Rice channel is expressed as: Where LOS is the line-of-sight part of the channel, NLOS is the non-line-of-sight part of the channel, ε is the Rice factor, and β is the path loss of the channel; Obey standard complex Gaussian distribution The path loss coefficient β is related to the physical straight-line distance d between the two ends of the channel, is a function of azimuth and elevation; where Indicates that the user U is collected from information k The azimuth and elevation angles to IRS1, then the steering vector of the l1th reflection unit of IRS1 is It is expressed as: Among them, d IRS1 is the distance between two adjacent reflection units in IRS1, λ c is the carrier wavelength, i1(l1)=mod(l1-1,L), It means rounding down, L and W are the length and width of IRS1, L×W=L1, then Assume (x k ,y k ,z k ) is the information collection user U k The three-dimensional geographic coordinates, (x IRS1 ,y IRS2 ,z IRS3 ) is the three-dimensional geographic coordinate of IRS1, then the azimuth and elevation are expressed as: Among them, d Uk,IRS1 For U k Straight-line distance to IRS1; Similarly, IRS1 to energy collection user E n Channel g 1,n , IRS2 to information collection user U k Channel h 2,k , IRS2 to energy collection user E n Channel g 2,n The generation method of is consistent with this; For the channel G1 from BS to IRS1, let is the azimuth and elevation angle of the reflection unit on IRS1 reaching the BS, then the steering vector at the BS is expressed as: Among them, d BS is the distance between two adjacent BS antennas, let are the azimuth and elevation angles from IRS1 to BS, then It is expressed as: Assume (x BS ,y BS ,z BS ) is the three-dimensional geographic coordinates of the BS, then the azimuth and elevation are expressed as: Among them, d BS,IRS1 is the straight-line distance from BS to IRS1; The generation method of the channel G2 from BS to IRS2 and the reflection channel D from IRS1 to IRS2 is consistent with this.

4. According to claim 2, a method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network is characterized in that: In step S2, by jointly optimizing the reflection phase shift matrix θ1 of IRS1, the reflection phase shift matrix θ2 of IRS2 and the beamforming matrix {w k },{z n }, maximize the total information rate R of all information collection users under the constraints of the maximum eavesdropping rate and minimum energy of the energy collection users, the modulus-one constraint of the reflection coefficient of the IRS, and the maximum transmission power of the base station. k The security communication optimization problem P0 is expressed as: st: Among them, P BS is the transmit power constraint at the BS, is the maximum eavesdropping rate limit of the energy harvesting user, and ζ is the minimum energy harvesting limit of the energy harvesting user.

5. According to claim 4, a method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network is characterized in that: In step S3, for information collection user U k , let user U k The pilot signal sent is x k ∈C K×1 , where the pilot length is taken as the number of information collection users K, and the pilot signals sent by different users are orthogonal to each other, that is: Among them, P U is the power of the pilot transmitted by the information collection user; Then the signal received by BS at time t is: Where t=1,2,...,τ u , τ u To estimate the number of pilots required, N t is the BS receiving noise, H k,t =h d,k +G1θ 1,t h 1,k +G2θ 2,t h 2,k +G1θ 1,t D H θ 2,t h 2,k ,1≤k≤K; Using the orthogonality of the pilot signal, right-multiply x at the BS k ,get: in, The number of uses is τ u The pilot information is used as the input of GNN: Similarly, energy harvesting user E n Using the same pilot estimation method, we can get E at the BS: n The number is τ u Pilot information: Then we get the information collection user pilot Y U and energy harvesting user pilot Y E , which are respectively expressed as: Y U =[y1,y2,..,y k ],1≤k≤K Y E =[y1,y2,..,y n ],1≤n≤N Information collection user pilot Y U and energy harvesting user pilot Y E As the input feature vector of the graph neural network GNN.

6. According to claim 5, a method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network is characterized in that: In step S4, the secure communication optimization problem P0 is transformed into P1: st: Among them, ({w k },{z n },θ1,θ2)=g(Y U ,Y E ) represents the use of graph neural network GNN to simulate receiving pilot Y from BS U and Y E Optimal mapping to BS beamforming and IRS reflection phase shifts.

7. The method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network according to claim 6 is characterized in that: In step S5, the Lagrange multiplier vector is first introduced: By transferring the constraints C2 and C3 into the objective function, we obtain a partial primal-dual optimization problem, thereby transforming the secure communication optimization problem P1 into P2: st: It is solved by alternating optimization: first fix the dual variables to optimize the system variables to minimize the objective P2; then fix the system variables and optimize the dual variables to maximize P2; For P2, given the system variables, the Lagrange multiplier and Update in the direction of gradient ascent based on the subgradient: in is the number of iterations, and κ ≥ 0 is the update step size.

8. The method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network according to claim 7, characterized in that: In step S6, a GNN model is designed to abstract the wireless communication environment into a graph G = (V, E), where V = [IRS1, IRS2, U1, ..., U k ,E1,...,E n ] is the set of nodes in the graph, and each node v∈V corresponds to a representation vector x v , v∈V, where the node representation vector is: The representation vector of the node is received by the BS with pilot information Y U and Y E get; In GNN, the feature vector Y U and Y E After normalization and standardization, it is converted into a graph node representation vector. All representation vectors of the previous layer are used as input, and the representation vector is updated layer by layer. After multiple layers, the representation vector of each node is used to design the beamforming vector and reflection coefficient. In GNN, the graph representation vector is learned through the input layer, where U k The input feature of the node is the received pilot y k , E n The input feature of the node is the received pilot y n , the input features of the IRS1 node By Y U and Y E The input features of IRS2 node are obtained by weighted averaging. By Y U and Y E Take the weighted average to get; D updates the layer to update the representation vector, and gets The output result is obtained through the output layer, first through a fully connected linear layer Then we will represent the vector The phase shift of the beamforming and IRS output to the BS is mapped by the normalization layer.

9. The method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on a graph neural network according to claim 8, characterized in that: In the input layer of GNN, the pilot information y is first k The real and imaginary components of are separated and then converted into column vectors to obtain: Among them, the information related to the node can be embedded in the column vector: Among them, loc k Collect the user's three-dimensional coordinate parameters for information; Then Through a deep neural network layer As U k Initial representation vector of the node For energy harvesting users, the node representation vector is obtained in the same way: For an IRS node, its representation vector is formed by aggregating the representation vectors of the information collection user and the energy collection user: where ψ0(·) is an element-wise average function is a deep neural network layer; The representation vector of IRS2 node is obtained in the same way as IRS1: The representation vector obtained by the input layer Contains the feature vectors of all nodes and edges and passes them to the D update layer; The update of the representation vectors of all nodes in the update layer is based on the aggregation of the previous representation vector of itself and the representation vectors of its adjacent nodes. The specific mathematical expression is: Among them, V(v) represents the set of nodes adjacent to node v. is the aggregation function of the dth update layer, is a combination function; The aggregation function of GNN is expressed as: Among them, ψ(·) is a function that is invariant to the input permutation, and Use deep neural networks; For the IRS1 node, the representation vector comes from: in, It is a multi-layer perceptron, which is used to integrate the node representation vector and introduce the nonlinear activation function relu; ψ1(·) is implemented using the element pool averaging function: ψ1(x1,x2,..,x k )=mean(x1,x2,..,x k ) The representation vector of the IRS2 node is updated from the neighboring nodes, including the representation vectors from the IRS1 node and from the information collection user and eavesdropping user nodes: in, is the MLP connection layer, and the element-wise average function ψ1(·) is selected as the aggregation function; For the information collection user node, its representation vector is updated as follows: in, is an MLP connection layer, and selects the element pooling maximization function ψ2(·) as the aggregation function: ψ2(x1,x2,..,x k )=max(x1,x2,..,x k ) For energy harvesting user nodes, the update of the representation vector is as follows: in, is the MLP connection layer, and selects the element pooling maximization function ψ2(·) as the aggregation function; Updated node representation vector Passed to the output layer, the IRS1 reflection coefficient matrix is ​​finally obtained IRS2 reflection coefficient matrix and the beamforming matrix W at the BS = [w1,w2,..,w K ]∈C M×K and Z = [z1,z2,..,z N ]∈C M×N , while satisfying the unit mode constraints of v1 and v2 and the total transmission power constraint of the BS; Will represent the vector and Pass through a linear layer with 2M units respectively get: The beamforming matrix is ​​then output through a normalization layer as follows: Where W(i:j,u:v:) represents a submatrix of W, which is constructed by the indexes from the i-th row to the j-th row and the u-th column to the v-th column of W; Pass a linear layer with 2L1 units Output As shown below: Then, the reflection phase shift of each reflection unit in IRS1 is obtained through the normalization layer, as follows: Where, [V1] ij represents the element in the i-th row and j-th column of the matrix, symbol represents the index subvector from i to j; Following the same principle, Pass a linear layer with 2L2 units Output As shown below: Then, the reflection phase shift of each reflection unit in IRS2 is obtained through the normalization layer, as follows: Through the output layer, beamforming at the BS and reflection phase shift at the IRS are obtained.

10. The method for optimizing secure communication of a dual IRS-assisted wireless energy-carrying communication system based on graph neural network according to claim 8, characterized in that: In step S6, the GNN model is trained with the goal of minimizing the Lagrangian relaxation function P2 given the Lagrangian multiplier, and its loss function is: Among them, L is the The loss function of the GNN model during the iteration.

Images