Self-interference suppression method for low-complexity active reconfigurable intelligent surface based on meta-learning

Abstract

The invention relates to a self-interference suppression method for a low-complexity active reconfigurable intelligent surface based on meta-learning, and belongs to the technical field of wireless communication. The method comprises the following steps: firstly, initializing a base station precoding matrix, an RIS phase shift matrix, user weighting and rate, and establishing two small-scale neural networks to respectively carry out iterative optimization on the precoding matrix and the phase shift matrix by introducing a meta-learning method; in order to solve the problem that the dimension of the precoding moment is too high, a block gradient descent method is used for conducting dimension reduction processing on the precoding moment, and then complexity is reduced. A loss function is calculated and accumulated through multiple cycles, and original network parameters are updated through back propagation; and obtaining an approximate optimal solution of a final RIS phase shift matrix and a base station precoding matrix through multiple times of updating, thereby achieving the purpose of suppressing self-interference between active elements. Compared with a traditional alternating loop optimization (AO) algorithm, the method provided by the invention is better in performance, higher in convergence speed and higher in robustness.

Description

[0001] Technical Field

[0002] The present invention relates to a self-interference suppression method for a low-complexity active reconfigurable intelligent surface based on meta-learning, belonging to the technical field of wireless communication. Background Art

[0003] A reconfigurable intelligent surface (RIS) is a surface composed of a large number of low-cost and programmable reflecting elements, which can reconstruct the wireless environment by manipulating the phase of incident electromagnetic waves. It is considered a potential technology for enhancing the coverage and user weighted sum rate in the upcoming sixth-generation communication. Thanks to the passive phase-shift technology of RIS and the active beamforming technology of the base station (BS), the throughput and energy efficiency of wireless communication systems can be enhanced in various scenarios, especially in millimeter-wave (mmWave) communication scenarios.

[0004] In practical applications, common RISs include passive RISs, active RISs, etc. Among them, due to the gain of passive RIS being limited by the "multiplicative fading" effect, that is, the equivalent path loss of the transmitter-RIS-receiver link is the product of the path losses of the transmitter-RIS and RIS-receiver links, rather than the sum, it is usually thousands of times that of the direct signal. While active RIS can effectively overcome the "multiplicative fading" effect, in specific practices, there is a certain self-interference among the active elements in the active RIS. This is because RIS changes the propagation path of the signal through reflection or refraction. If the reflecting surface of RIS directly interferes with the user equipment or the base station, it may cause phase disorder between the reflected signal and the direct signal, thus generating self-interference and further affecting the reception quality. Therefore, there is an urgent need for a self-interference suppression method.

[0005] In recent years, the application of deep learning (DL) in wireless communication has attracted much attention because it can solve complex and challenging problems, as it is naturally good at extracting valuable features in high-dimensional spaces with low complexity. Among them, the meta-learning method is a mature deep learning model, which can quickly learn the features of a new task with only a small number of samples, and can transfer the knowledge learned from previous tasks to new tasks, thereby reducing the training time and data requirements. It can improve the sample efficiency and enhance the generalization ability. Therefore, the present invention intends to develop a self-interference suppression method for a low-complexity active reconfigurable intelligent surface based on meta-learning. Summary of the Invention

[0006] The purpose of the present invention is to provide a self-interference suppression method for a low-complexity active reconfigurable intelligent surface based on meta-learning, aiming to solve the technical problem of poor user communication effect caused by self-interference in active RIS-assisted communication.

[0007] To achieve the above object, the technical solution of the present invention is: a self-interference suppression method for a low-complexity active reconfigurable intelligent surface based on meta-learning, and the specific steps are as follows:

[0008] Step1: Establish an active RIS-assisted multi-user MISO system, which amplifies and forwards the data received by the active RIS from the base station to a single-antenna user within a specified area in each time slot;

[0009] Step2: Establish a problem of maximizing the user weighted sum rate of the system, and decompose the problem into a RIS phase shift matrix sub-problem and a base station precoding sub-problem; design a joint optimization scheme to jointly optimize the regular active RIS phase shift matrix and the digital precoding at the base station BS;

[0010] Step3: By introducing meta-learning, neural network, block gradient descent method, the gradient descent method is used to solve an approximate optimal solution for the joint variables, reduce the algorithm complexity, maximize the user weighted sum rate, and suppress the self-interference between active components.

[0011] The specific content of Step1 is as follows:

[0012] Step1.1: Construct an active RIS-assisted multi-user MISO system, which includes 1 BS with M antennas, 1 active RIS with N active reflection elements, and K single-antenna users. The BS and the users communicate through the direct link channel and the reflection channel. The direct link channel is from the base station to the user, and the reflection channel is from the base station to the active RIS to the user;

[0013] Step1.2: Let be the transmission link from the BS to the RIS, be the direct transmission link from the BS to the user, be the reflection link from the RIS to the user, be the precoding matrix at the BS side, and the data in the k-th column is defined as w k , and at this time, the received signal y k at user k is modeled as:

[0014]

[0015] where, respectively represent the channel vector between the BS and user k and the channel vector between the active RIS and user k, sj represents the transmission signal sequence of the base station represents the BS beamforming vector of the transmission signal s k , represents the phase shift matrix of the active RIS, θ i ∈(0, 2π), p n represents the amplification factor of the n-th active element; represents the noise introduced and amplified by the reflection amplifier. v is related to the input noise of the active RIS element and the inherent device noise. represents a complex multivariate Gaussian distribution with mean μ and variance ∑, I N is the N×N identity matrix, 0 N is the N×1 zero vector. is the noise power. represents the additive Gaussian white noise at user k, with mean 0 and variance σ 2 ;

[0016] Step1.3: The signal-to-interference-plus-noise ratio (SINR) is defined as the ratio of the signal to the sum of interference and noise in a system and is a major technical indicator for measuring the reliability of communication quality in a communication system. The larger the SINR, the better the communication quality. For a multi-user MISO system, the SINR at user k, SINR k is given by:

[0017]

[0018] where, is the cascaded channel from the BS to user k. is the direct transmission link. is the reflection link, w k represents the k-th column vector of.

[0019] Step1.4: Establish the power constraint. Let the total system power be P, the transmit power of the BS be P BS , and the reflection power of the RIS be P A . We get:

[0020] P = P BS + P A

[0021]

[0022] where, is the effective signal power of the BS. is the effective power signal of the RIS. is the amplified noise power of the RIS. E{·} represents the Euclidean expectation of the square of the radiated signal. Since the active RIS amplifies both the signal and the noise, the additional power consumption caused by noise amplification needs to be considered.

[0023] Step1.5: The reflected signal of the active RIS with self-interference is modeled as:

[0024] y = ΦX + ΦHy + Φv = (I N - ΦH) -1Φ(X + v)

[0025] Among them, is the self-interference matrix, ΦX is the effective signal, ΦHy is the self-interference signal, Φv is the dynamic noise. Due to the existence of active element self-interference, the phase shift matrix Φ is transformed into (I N - ΦH) -1 Φ; and it is stipulated that the phase shift matrix Ψ with self-interference = (I N - ΦH) -1 Φ;

[0026] Step1.6: Solve the equivalent cascaded channel in the case of self-interference The derived formula is:

[0027]

[0028] Through the first-order Taylor expansion, (I N - ΦH) -1 ≈ I N + ΦH. Let Ψ = diag(φ H ), and we get:

[0029]

[0030] Among them, α represents

[0031] The specific content of Step 2 is as follows:

[0032] Step2.1: Establish the problem of maximizing the weighted sum rate of users in the system under the premise of self-interference, and jointly optimize the RIS phase shift matrix and digital precoding at the BS. The optimization function is as follows:

[0033]

[0034] C4: |θ j | ∈ [0, 2π), j = 1, 2, 3, ……, N

[0035] Among them, ω k represents the weight of the user, W H represents the transpose matrix of the precoding matrix W, represents the second norm of Ψ, C1 and C2 are power constraints, represent the maximum power of the base station and RIS respectively; C3 is the discrete phase constraint, C4 is the phase angle change constraint, and θ j represents the phase shift angle of the RIS element;

[0036] Step2.2: Train the gradient-based meta-learning model, update the model parameters, and solve the approximate optimal solutions of W and Ψ. Use the known G and H d, f, ω k The channel matrix is input into the neural network NN, and the initialization gradients of W and Ψ are input simultaneously. and Iterative updates are performed using meta - learning. The iteration includes inner iteration, outer iteration, and epoch iteration;

[0037] The inner iteration is used to cyclically optimize the target variables: Define two sub - networks called the precoding network CN and the phase network PN, which are responsible for optimizing W and Ψ respectively. The j - th update process is expressed as:

[0038] W * = CN(W (0,j) , Ψ * ),

[0039] Ψ * = PN(W * , Ψ (0,j) )

[0040] where W (i,j) and Ψ (i,j) represent W and Ψ in the i - th inner iteration of the j - th outer iteration. W * represents the precoding matrix obtained after N i inner loops, and Ψ * represents the phase - shift matrix obtained after N i inner loops;

[0041] The outer iteration is used to accumulate the loss: Each outer iteration has N i inner iterations and works in a self - supervised learning manner. The loss function of the j - th outer iteration is expressed as the negative value of the user - weighted sum rate WSR:

[0042]

[0043] The epoch iteration is used to update the parameters of NN: After completing N o outer iterations, the losses are added up and the average loss is:

[0044]

[0045] Backpropagation is performed, and the NN parameters N W , N Ψ in the precoding network and the phase - shift network are updated using the Adam optimizer. The formula is:

[0046]

[0047] where α W , α Φ are the learning rates of the two networks respectively. Denote the updated precoding network parameters, Denote the updated phase shift network parameters;

[0048] Step2.3: Detail the design of the precoding network CN and the phase network PN;

[0049] In the presence of self-interference, let the WSR in the i-th inner iteration and the j-th outer iteration be Denoted as:

[0050]

[0051] where, is the initialized or updated phase shift matrix. In the working process of CN, first calculate the WSR, and use the block gradient descent method to calculate the gradient Δw of w (i,j) with respect to the WSR (i,j) feed it into CN, and add it to the original w (i,j) , and the updated column matrix at this time is w * ; splice and restore the updated column vector into the precoding matrix W * and perform power constraint on it to obtain the updated W * , which is W (i+1,j) ;

[0052] w * = w (i,j) + Δw (i,j)

[0053]

[0054] Update the CN network parameters through epoch iteration to obtain the final precoding matrix W opt ;

[0055] Optimize the PN layer, and express the WSR in the i-th inner iteration and the j-th outer iteration as The formula is:

[0056]

[0057] where, is the initialized or updated precoding vector, and Ψ (i,j) is the corresponding phase shift matrix at this time. Use the gradient descent method to feed the gradient ΔΨ of Ψ (i,j) with respect to the WSR into PN. Since the trigonometric function has periodicity and non-monotonic change, a modulation function δ(·) with λ as the amplification factor is designed to limit ΔΨ within (0, 2π) to obtain the updated gradient and add it to Ψ (i,j) to obtain the result Ψ (i+1,j), which is as follows:

[0058]

[0059] Update the PN network parameters through epoch iteration to obtain the final phase shift matrix Ψ opt .

[0060] The specific content of Step3 is as follows:

[0061] Step3.1: For the given channel matrix, randomly initialize W, ψ, and Set the convergence accuracy ε and the maximum number of iterations;

[0062] Step3.2: Solve ΔW and ΔΨ as the inputs of the NN, and update W iteratively through the CN network * to obtain the optimal W opt , and calculate the average loss at this time

[0063] Step3.3: Input W opt into the PN network, and update Ψ iteratively * to obtain the optimal Ψ opt , and calculate the updated average loss

[0064] Step3.4: Substitute the obtained Ψ opt , Ψ opt into the optimization function to calculate the user weighted sum rate R(W, Ψ), and iterate until the algorithm converges to obtain the optimal result, achieving the purpose of suppressing self-interference, then end the iteration, otherwise continue to execute Step3.2.

[0065] The beneficial effects of the present invention are as follows: In traditional algorithms, the method for suppressing self-interference mainly decomposes the problem into several sub-problems through alternating optimization, and uses auxiliary variables, Lagrange multipliers, etc. to convert non-convex problems into convex problems. The search speed is slow and it is easy to fall into local optimal solutions. However, the present invention uses an active RIS-assisted communication. According to the distribution of users, it uses the meta-learning method in the deep learning model, comprehensively considers the overall situation of variables through three-layer iteration, jointly optimizes the phase shift angle matrix of the active reconfigurable intelligent surface and digital precoding at the base station, maximizes the user weighted sum rate, realizes fast search efficiency and stronger robustness, suppresses the self-interference between active components, and improves the communication quality of the system. Description of the Drawings

[0066] Figure 1 is the system model diagram in the embodiment of the present invention;

[0067] Figure 2 is the schematic diagram of the algorithm framework;

[0068] Figure 3 It is a relationship diagram of the user weighted sum rate and the self-interference factor in the example of the present invention;

[0069] Figure 4 It is a comparison diagram of the algorithm iteration convergence in the example of the present invention. Detailed implementation manners

[0070] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation manners.

[0071] Example 1: As Figure 1 shown, it includes a base station equipped with M antennas, which provides different communication services for K single-antenna users respectively, and the K users are respectively distributed in a circular area with a radius of 100 m; an active RIS equipped with N active elements, operating in full-duplex (FD) mode, and can simultaneously complete data transmission and reception. The direct link between the base station and each user is a strong direct link. At this time, the known channel state information (CSI) is: is the transmission link from the BS to the RIS, is the direct complex fading channel from the BS to the user, is the complex fading reflection channel from the RIS to the k-th user. For the proposed optimization problem, the base station precoding matrix and the RIS phase shift matrix are jointly optimized by using the manifold element learning method to solve the approximate optimal solution of the original problem, maximize the user weighted sum rate, and thus suppress self-interference.

[0072] The specific steps are as follows:

[0073] Step1: Establish an active RIS-assisted multi-user MISO system, which amplifies and forwards the data received by the active RIS from the base station to the single-antenna users within the specified area in each time slot;

[0074] Step1.1: Construct an active RIS-assisted multi-user MISO system, which includes 1 BS with M antennas, 1 active RIS with N active reflection elements, and K single-antenna users. The BS and the users communicate through the direct channel and the reflection channel. The direct channel is base station → user, and the reflection channel is base station → active RIS → user;

[0075] Step1.2: Let be the transmission link from the BS to the RIS, be the direct transmission link from the BS to the user, be the reflection link from the RIS to the user, be the precoding matrix at the BS end, and the k-th column data is defined as wk At this time, the received signal y at user k k is modeled as:

[0076]

[0077] where respectively represent the channel vector between the BS and user k and the channel vector between the active RIS and user k, s j represents the transmission signal sequence of the base station represents the transmission signal s k is the BS beamforming vector of the transmission signal s, represents the phase shift matrix of the active RIS, θ i ∈(0, 2π), p n represents the amplification factor of the nth active element; represents the noise introduced and amplified by the reflection amplifier, v is related to the input noise of the active RIS element and the inherent device noise, represents a complex multivariate Gaussian distribution with mean μ and variance ∑, I N is the N×N identity matrix, 0 N is the N×1 zero vector, is the noise power, represents the additive Gaussian white noise at user k, with mean 0 and variance σ 2 ;

[0078] Step1.3: According to the multi-user MISO system, the signal-to-interference-plus-noise ratio SINR at user k k is:

[0079]

[0080] where is the cascaded channel from the BS to user k, is the direct transmission link, is the reflection link, w k represents the kth column vector of;

[0081] Step1.4: Establish the power constraint condition. Let the total system power be P, the BS transmission power be P BS , and the RIS reflection power P A , and we get:

[0082] P = P BS + P A

[0083]

[0084] where is the effective signal power of the BS, is the effective power signal of the RIS, is the amplified noise power of the RIS, and E{·} represents the Euclidean expectation of the square of the radiated signal;

[0085] Step1.5: The reflected signal of the active RIS with self-interference is modeled as:

[0086] y = ΦX + ΦHy + Φv = (I N - ΦH) -1 Φ(X + v)

[0087] where, is the self-interference matrix, ΦX is the effective signal, ΦHy is the self-interference signal, Φv is the dynamic noise. Due to the self-interference of active components, the phase shift matrix Φ is transformed into (I N - ΦH) -1 Φ; and it is stipulated that the phase shift matrix Ψ with self-interference is Ψ = (I N - ΦH) -1 Φ;

[0088] Step1.6: Solve the equivalent cascaded channel in the case of self-interference The derived formula is:

[0089]

[0090] Through the first-order Taylor expansion, (I N - ΦH) -1 ≈ I N + ΦH. Let Ψ = diag(φ H ), and we get:

[0091]

[0092] where, α represents

[0093] Step2: Establish the problem of maximizing the user weighted sum rate of the system, and decompose the problem into the RIS phase shift matrix sub-problem and the base station precoding sub-problem; design a joint optimization scheme to jointly optimize the regular active RIS phase shift matrix and the digital precoding at the base station BS;

[0094] Specifically, the known objective function is:

[0095]

[0096] C4: |θ j | ∈ [0, 2π), j = 1, 2, 3, ……, N

[0097] where, ω kDenotes the weight of the user, W H Denotes the transpose matrix of the precoding matrix W Denotes the second norm of Ψ, C1 and C2 are power constraints Denote the maximum powers of the base station and RIS respectively; C3 is the discrete phase constraint, C4 is the phase angle change constraint, θ j Denotes the phase shift angle of the RIS element

[0098] At this time, an attribute is introduced, indicating that maximizing the user weighted sum rate requires the full power of the BS. The cascaded channel of the k-th user is expressed as

[0099] After determining the target variables, meta-learning is used to solve the approximate optimal solutions for W and Ψ. The specific network design is as Figure 2 shown

[0100] The inner iteration is used to cyclically optimize the target variables: Define two sub-networks called the precoding network CN and the phase network PN, which are responsible for optimizing W and Ψ respectively. The j-th update process is expressed as

[0101] W * = CN(W (0,j) , Ψ * ),

[0102] Ψ * = PN(W * , Ψ (0,j) )

[0103] where W (i,j) and Ψ (i,j) represent W and Ψ in the i-th inner iteration of the j-th outer iteration, W * represents the precoding matrix obtained after N i inner loops, and Ψ * represents the phase shift matrix obtained after N i inner loops

[0104] The outer iteration is used to accumulate the loss: Each outer iteration has N i inner iterations and works in a self-supervised learning manner. The loss function of the j-th outer iteration is expressed as the negative value of the user weighted sum rate WSR

[0105]

[0106] The epoch iteration is used to update the parameters of the NN: After completing N o outer iterations, the losses are added up and the average loss is

[0107]

[0108] Perform backpropagation and update the NN parameters N in the precoding network and the phase shift network using the Adam optimizer W ,N Ψ , the formula is:

[0109]

[0110] where, α W ,α Φ are the learning rates of the two networks respectively represents the updated precoding network parameters represents the updated phase shift network parameters, and the learning rate α W = 1e -3 ,α Φ = 1.5e -3 , and e is the base of the exponent

[0111] Design the precoding network CN and the phase network PN in detail:

[0112] In the presence of self-interference, let the WSR in the i-th inner iteration and the j-th outer iteration be expressed as:

[0113]

[0114] where, is the initialized or updated phase shift matrix. In the working process of CN, first calculate the WSR, and use the block gradient descent method to calculate the gradient Δw of w (i,j) with respect to the WSR (i,j) feed it into CN and add it to the original w (i,j) , and the updated column matrix at this time is w * ; splice and restore the updated column vector into the precoding matrix W * and perform power constraint on it to obtain the updated W * , which is W (i+1,j) ;

[0115] w * = w (i,j) + Δw (i,j)

[0116]

[0117] Update the CN network parameters through epoch iteration to obtain the final precoding matrix W opt .

[0118] The PN network specifically includes an input fully connected layer (containing N neurons), a hidden layer (containing 200 neurons), a rectified linear layer, and a fully connected layer (containing N neurons).

[0119] Optimize the PN layer and represent the WSR in the i-th inner iteration and the j-th outer iteration as The formula is:

[0120]

[0121] Among them, is the precoding vector after initialization or update, and Ψ (i,j) is the corresponding phase shift matrix at this time. Use the gradient descent method to feed the gradient ΔΨ of Ψ (i,j) with respect to WSR into PN. Since the trigonometric function has periodicity and non-monotonic change, a modulation function δ(·) with λ as the amplification factor is designed to limit ΔΨ within (0, 2π) to obtain the updated gradient and add it to Ψ (i,j) to obtain the result Ψ (i+1,j) after one iteration, which is:

[0122]

[0123] Update the PN network parameters through epoch iteration to obtain the final phase shift matrix Ψ opt ;

[0124] The specific block gradient descent method is as follows:

[0125] Divide the precoding matrix into m blocks (usually of equal size), and each block contains b samples (i.e., M×K = m×b). At this time, each sub-block

[0126] After that, in each iteration, only update some variables (one or more blocks), and keep other variables unchanged.

[0127] For block X i , define the block gradient which represents the partial derivative of the loss function with respect to the i-th variable block:

[0128]

[0129] In the k-th iteration, select a certain to update, where α is the learning rate of the precoding network, and obtain the result after one iteration

[0130]

[0131] The other blocks remain unchanged:

[0132]

[0133] For the gradient update criterion, select the block with the largest gradient norm as the next-generation update target to accelerate convergence.

[0134] Finally, each optimized block X i is restored to the precoding matrix W. Update the CN parameters using the epoch loop, and the final W opt .

[0135] The PN network specifically includes an input fully connected layer (containing 2K neurons), a hidden layer (containing 200 neurons), a rectified linear layer, and a fully connected layer (containing 2K neurons).

[0136] Conduct a complexity analysis of the system: If Φ is a diagonal matrix, then at this time the complexity of At this time, the complexity of calculating the weighted sum rate of k users is NN Φ the computational complexity of the network is That is, the complexity of PN is where N i , N o , N e are the number of times of the inner loop, outer loop, and epoch loop respectively. The complexity of CN is mainly introduced during the block gradient calculation, and the complexity is NN W the network complexity is In summary, the final complexity of this algorithm is while the complexity of the traditional AO algorithm is N1 and N2 are the number of times of the inner loop of the AO algorithm, and N3 is the number of times of the outer loop of the AO algorithm.

[0137] Compared with the traditional algorithm, the present invention realizes a reduction in the complexity of M from cubic to linear and a reduction in the complexity of N from quadratic to linear. It can effectively reduce the complexity compared with the traditional AO algorithm.

[0138] Step3: By introducing meta-learning, neural networks, block gradient descent method, and gradient descent method, solve the approximate optimal solution for the joint variables, reduce the algorithm complexity, maximize the user weighted sum rate, and suppress the self-interference between active components.

[0139] Step3.1: For the given channel matrix, randomly initialize W, Ψ, and set the convergence accuracy ε and set the maximum number of iterations;

[0140] Step 3.2: Solve for ΔW and ΔΨ as the inputs to the NN, and iteratively update W through the CN network * to obtain the optimal W opt and calculate the average loss at this time

[0141] Step 3.3: Input W opt into the PN network and iteratively update Ψ * to obtain the optimal Ψ opt and calculate the updated average loss

[0142] Step 3.4: Substitute the obtained W opt , Ψ opt into the optimization function to calculate the user weighted sum rate R(W, Ψ), and iterate until the algorithm converges to obtain the optimal result, achieving the purpose of suppressing self-interference, then end the iteration; otherwise, continue to execute Step 3.2

[0143] In this embodiment, the BS and the active RIS are located at (0m, 0m) and (300m, 0m) respectively, and 4 users are randomly distributed within a circular area with the center at (300m, 10m) and a radius of 100m. The number of BS antennas M = 64, the number of RIS elements N = 100, and the number of users K = 4. The noise power is set to The total system power consumption where Due to the existence of self-interference in the RIS, the self-interference factor δ = [-60, -55, -50, -45, -40, -35, -30, -25, -20]. Theoretically, the larger the self-interference factor, the more obvious the impact of self-interference on the system. According to the settings of these parameters, the system is simulated using python

[0144] As Figure 3 shown, the user weighted sum rate decreases as the self-interference factor δ increases, but the degree of decrease in the result obtained by the algorithm proposed in the present invention is much smaller than that in the traditional algorithm, indicating that the proposed scheme can effectively suppress the self-interference between active components, thereby improving the system communication quality

[0145] As Figure 4 shown, the comparison of the number of iterations of this algorithm with that of the traditional algorithm shows that the convergence speed is significantly faster than that of the traditional algorithm. Therefore, it shows that the present invention uses an effective optimization strategy or method, can approach the optimal solution faster, reduces computational redundancy, and has a significant advantage over the traditional algorithm

[0146] The specific implementation methods of the present invention have been described in detail in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned implementation manners. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention.

Claims

1. A self-interference suppression method for low-complexity active reconfigurable smart surfaces based on meta-learning, characterized by: Step 1: Establish an active RIS-assisted multi-user MISO system, which amplifies the data received by the active RIS from the base station in each time slot and forwards it to the single-antenna user in the specified area; Step 2: Establish the user weighted sum rate maximization problem of the system, and decompose the problem into RIS phase shift matrix sub-problem and base station precoding sub-problem; design a joint optimization scheme to jointly optimize the regular active RIS phase shift matrix and the digital precoding at the base station BS; Step 3: By introducing meta-learning, neural networks, and block gradient descent, the gradient descent method approximates the optimal solution for the joint variables, reduces the algorithm complexity, maximizes the user weighted sum rate, and suppresses self-interference between active components.

2. According to claim 1, a self-interference suppression method for low-complexity active reconfigurable smart surface based on meta-learning is characterized in that: The Step 1 is specifically as follows: Step 1.1: construct an active RIS-assisted multi-user MISO system, which includes a BS with M antennas, an active RIS with N active reflective elements, and K single-antenna users. The BS and the user communicate through a direct channel and a reflection channel. The direct channel is base station→user, and the reflection channel is base station→active RIS→user. Step 1.2: Set is the transmission link between BS and RIS, It is the direct transmission link between BS and user. It is the reflection link from RIS to the user. is the precoding matrix at the BS end, where the kth column data is defined as w k , at this time the received signal y at user k k Modeled as: in, denote the channel vector between BS and user k and the channel vector between active RIS and user k, s j The base station sends a signal sequence s = [s1, s2, ..., s k ], Indicates sending signal s k The BS beamforming vector, represents the phase shift matrix of active RIS, θ i ∈(0,2π),p n represents the amplification factor of the nth active element; represents the noise introduced and amplified by the reflection amplifier, v is related to the input noise and inherent device noise of the active RIS elements, represents a complex multivariate Gaussian distribution with mean μ and variance ∑, I N is the N×N identity matrix, 0 N is an N×1 zero vector, is the noise power, represents the additive Gaussian white noise at user k, with a mean of 0 and a variance of σ 2 ; Step 1.3: According to the multi-user MISO system, the SINR of user k k for: in, is the cascade channel from BS to user k, It is a direct transmission link. is the reflection link, w k express The k-th column vector of ; Step 1.4: Establish power constraints, assuming that the total system power is P and the BS transmission power is P BS , RIS reflected power P A ,get: P=P BS +P A in, is the effective signal power of BS, is the effective power signal of RIS, is the amplified noise power of RIS, E{·} represents the Euclidean expectation of the square of the radiated signal; Step 1.5: The reflected signal of active RIS with self-interference is modeled as: y=ΦX+ΦHy+Φv=(I N -ΦH) -1 Φ(X+v) in, is the self-interference matrix, ΦX is the effective signal, ΦHy is the self-interference signal, and Φv is the dynamic noise. Due to the existence of active element self-interference, the phase shift matrix Φ is transformed into (I N -ΦH) -1 Φ; and stipulate that there is a phase shift matrix Ψ=(I N -ΦH) -1 Φ; Step 1.6: Solve the equivalent cascade channel in the presence of self-interference The export formulas are: Through the first-order Taylor expansion, we have (I N -ΦH) -1 ≈I N +ΦH, let Φ=diag(φ H ),get: Among them, α represents 3. According to claim 1, a self-interference suppression method for low-complexity active reconfigurable smart surface based on meta-learning is characterized in that: The Step 2 is specifically as follows: Step 2.1: Establish the user weighted sum rate maximization problem of the system under the premise of self-interference, jointly optimize the RIS phase shift matrix and the digital precoding at the BS, and the optimization function is as follows: C4:|θ j |∈[0,2π),j=1,2,3,……,N Among them, ω k represents the user's weight, W H represents the transposed matrix of the precoding matrix W, represents the second norm of Ψ, C1 and C2 are power constraints, Represent the maximum power of the base station and RIS respectively; C3 is the discrete phase constraint, C4 is the phase angle change constraint, θ j represents the phase shift angle of the RIS element; Step 2.2: Train the gradient-based meta-learning model, update the model parameters, solve the approximate optimal solution of W,Ψ, and transform the known G,H d ,f,ω k The channel matrix is ​​input into the neural network NN, and the initialization gradients of W and Ψ are input at the same time. and Iterative updating is performed using meta-learning, where the iteration includes inner iteration, outer iteration and epoch iteration; The inner iteration is used to cyclically optimize the target variable: two sub-networks are defined, namely the precoding network CN and the phase network PN, which are responsible for optimizing W and Ψ respectively. The j-th update process is expressed as: W * =CN(W (0,j) ,Ψ * ), P * =PN(W * ,P (0,j) ) Among them, W (i,j) and (i,j) represents W and Ψ in the i-th inner iteration of the j-th outer iteration, W * Indicates that after N i The precoding matrix obtained by the inner loop is Ψ * Indicates that after N i The phase shift matrix obtained by the inner loop; The outer iteration is used to accumulate losses: each outer iteration has N i The jth inner iteration works in a self-supervised learning manner, and the loss function of the jth outer iteration is It is expressed as the negative value of the user weighted sum rate WSR: epoch iteration is used to update the parameters of NN: N o After the outer iterations, the losses are summed and the average loss is obtained for: Perform back propagation and use the Adam optimizer to update the NN parameters N in the precoding network and phase shift network W ,N Ψ , the formula is: Among them, α W ,α Φ are the learning rates of the two networks respectively, represents the updated precoding network parameters, represents the updated phase shift network parameters; Step 2.3: Detailed design of precoding network CN and phase network PN; In the presence of self-interference, let the WSR in the i-th inner iteration and the j-th outer iteration be It is expressed as: in, is the initialized or updated phase shift matrix. In the CN workflow, WSR is first calculated, and w is converted to (i,j) Gradient Δw relative to WSR (i,j) Feed into CN and add to the original w (i,j) , the updated column matrix is ​​w * ; Concatenate the updated column vectors to restore the precoding matrix W * And perform power constraints on it to obtain the updated W * , is W (i+1,j) ; In * =in (i,j) +Δw (i,j) Update the CN network parameters through epoch iteration to obtain the final precoding matrix W opt ; Optimize the PN layer and express the WSR in the i-th inner iteration and the j-th outer iteration as The formula is: in, is the initialized or updated precoding vector, Ψ (i,j) For the corresponding phase shift matrix at this time, use the gradient descent method to convert Ψ (j,j) The gradient ΔΨ relative to WSR is fed into PN. Since the trigonometric function is periodic and non-monotonic, a regulation function δ(·) with λ as the amplification factor is designed to limit ΔΨ to (0,2π) and obtain the updated gradient And add to Ψ (i,j) , and get the result after one iteration Ψ (i+1,j) ,for: Update the PN network parameters through epoch iteration to obtain the final phase shift matrix Ψ opt .

4. According to claim 1, a self-interference suppression method for low-complexity active reconfigurable smart surface based on meta-learning is characterized in that: The Step 3 is specifically as follows: Step 3.1: For a given channel matrix, randomly initialize W, ψ, and Set the convergence accuracy ε and the maximum number of iterations; Step 3.2: Solve ΔW and ΔΨ as the input of NN, and iteratively update W through CN network * , get the optimal W opt , calculate the average loss at this time Step 3.3: W opt Input PN network, update Ψ through iteration * , and get the optimal Ψ opt , calculate the average loss after update Step 3.4: The obtained Ψ opt ,Ψ opt Substitute into the optimization function Calculate the user weighted sum rate R(W,Ψ) in the iterative process until the algorithm converges to the optimal result and achieves the purpose of suppressing self-interference. Then the iteration ends, otherwise continue to execute Step 3.2.

Images