Multi-user semantic communication method and system

Abstract

The invention discloses a multi-user semantic communication method and system, and the method comprises the steps: constructing a multi-user semantic communication system assisted by an intelligent reflection surface based on a diversified topological structure; the method comprises the following steps: based on an intelligent reflecting surface assisted multi-user semantic communication system, carrying out signal transmission evaluation through a multi-user parallel transmission method combined with source channel coding, and constructing an optimized user scheduling strategy; and performing iterative optimization solution on the optimized user scheduling strategy based on a multi-user scheduling algorithm of deep reinforcement learning, constructing an optimization problem of user scheduling, and realizing multi-user semantic communication. According to the method, the decoding of the multi-user information can be assisted by utilizing the orthogonality of the semantic information in the semantic space and the orthogonality of the channel, so that the calculation cost of the semantic communication architecture and resources required by deployment are saved. The multi-user semantic communication method and system can be widely applied to the technical field of intelligent semantic communication.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent semantic communication, in particular to a multi-user semantic communication method and system. BACKGROUND

[0002] Most semantic communications are committed to researching more efficient semantic encoding and decoding algorithms and more intelligent removal of redundant information in the transmission process, while the nature of semantic information itself is not fully utilized, such as the sparsity of semantic space and the orthogonality of semantic features as high-dimensional vectors. In existing semantic communication research, only the improvement of semantic extraction process is often considered, and efforts are made to improve the compression ratio. In addition, the existing research on semantic communication and multiple access is usually based on a deep source-channel joint coding framework, in which a pair of neural network-based encoders and decoders are deployed at the receiving and transmitting ends. In a multi-user semantic communication system, multiple pairs of receiving and transmitting relationships are allowed, and only the nodes with matched encoders and decoders can complete communication. Such a method utilizes the orthogonality of semantic feature vectors, but each pair of receiving and transmitting relationship needs a pair of neural network-based encoders and decoders. This will result in a fixed number of pairs of transceivers in a multi-user system, lack of flexible network topology, and limited semantic orthogonality and channel orthogonality, which will lead to reduced transmission quality if too many users are simultaneously accessed.

[0003] In summary, with the continuous expansion of the Internet of Things and the increasing number of intelligent devices, the amount of data to be transmitted between Internet of Things nodes increases, and the types of data are diverse. Traditional communication cannot utilize limited wireless transmission resources to meet the growing transmission demand. SUMMARY

[0004] To solve the above technical problems, the purpose of the present application is to provide a multi-user semantic communication method and system which can utilize semantic information in semantic space orthogonality and channel orthogonality to assist in decoding multi-user information, saving the computational cost of the semantic communication architecture and the resources required for deployment.

[0005] The first technical solution adopted by the present application is a multi-user semantic communication method, comprising the following steps:

[0006] Based on the diversified topology structure, an intelligent reflecting surface assisted multi-user semantic communication system is constructed, which includes an intelligent reflecting surface and a plurality of half-duplex single-antenna users for semantic communication.

[0007] The intelligent reflecting surface assisted multi-user semantic communication system evaluates signal transmission by a multi-user parallel transmission method of joint source channel coding, and constructs an optimized user scheduling strategy.

[0008] The multi-user scheduling algorithm based on deep reinforcement learning iteratively optimizes and solves the optimized user scheduling strategy, constructs an optimization problem of user scheduling, and realizes multi-user semantic communication.

[0009] Further, the intelligent reflecting surface assisted multi-user semantic communication system evaluates signal transmission by a multi-user parallel transmission method of joint source channel coding, and constructs an optimized user scheduling strategy, which specifically includes:

[0010] The intelligent reflecting surface assisted multi-user semantic communication system constructs a channel model between users.

[0011] The semantic encoder and the channel encoder are combined to construct a joint source channel encoder.

[0012] Based on the joint source channel encoder, the semantic features of the data transmitted by the channel model between users are extracted, the channel state information is converted into semantic features, and multi-user signal superposition is performed to obtain the total signal sent by the user.

[0013] The total signal sent by the user is received, signal equalization processing is performed, and decoding is performed through a pre-trained decoder to obtain a reconstructed semantic feature vector.

[0014] The signal-to-interference-and-noise ratio of the reconstructed semantic feature vector is obtained, and semantic throughput calculation is performed to obtain the semantic throughput.

[0015] Based on the semantic throughput, the semantic similarity is evaluated by structural similarity, and an optimized user scheduling strategy is constructed.

[0016] Further, the expression of the channel model between users is specifically as follows:

[0017]

[0018] In the above formula, h r,k represents the channel model, r represents user r, k represents user k, g r ,g k are the channels between user r and user k and the intelligent reflecting surface (IRS), respectively, and Φ represents the reflection coefficient matrix of the IRS.

[0019] Further, the expression of the total signal sent by the user is specifically as follows:

[0020]

[0021] In the above formula, sr represents the total signal sent by user r, B t represents the scheduling strategy, r represents user r, j represents user j, K represents the total number of users, represents the weighted signal.

[0022] Furthermore, the expression for calculating the semantic throughput is specifically as follows:

[0023]

[0024] In the above formula, Γ r,k represents semantic throughput, S represents the average semantic amount contained in the image (unit is the number of symbols), B represents the transmission bandwidth, I represents the number of bits of the original data, and C r Indicates the compression ratio of converting bits into the number of symbols, represents the trainable parameters of the decoder, ξ(·) represents the structural similarity, and γ r,k represents the signal-to-drying ratio, and θ represents the parameters of the semantic decoder.

[0025] Furthermore, the multi-user scheduling algorithm based on deep reinforcement learning iteratively optimizes and solves the optimized user scheduling strategy, constructs the optimization problem of user scheduling, and realizes the multi-user semantic communication step, which specifically includes:

[0026] Based on the user fairness criterion, the user scheduling strategy, IRS passive beamforming, and semantic codec parameters are jointly optimized to maximize the minimum value of user semantic throughput and construct a signal transmission optimization objective function.

[0027] The signal transmission optimization objective function is decomposed to obtain the problem of maximizing the minimum SINR of the user and the problem of maximizing the semantic similarity;

[0028] A dual-loop reinforcement learning algorithm is constructed, where the outer loop represents a reinforcement learning algorithm for determining user scheduling, and the inner loop represents a data-driven end-to-end training process for optimizing the semantic encoding strategy and passive beamforming of the IRS.

[0029] Based on the outer loop, the DDPG algorithm is used to generate a scheduling strategy to solve the problem of maximizing the minimum SINR of the user and obtain the user scheduling strategy;

[0030] Based on the inner loop, semantic encoding is performed on the combined source channel to solve the problem of maximizing semantic similarity and obtain the passive beamforming and semantic encoding strategies of the IRS.

[0031] Combining channel characteristics with data characteristics, a joint optimization problem of user scheduling, IRS passive beamforming, and semantic coding strategy is constructed to achieve multi-user semantic communication.

[0032] Further, the expression of the signal transmission optimization objective function is specifically as follows:

[0033]

[0034] In the above formula, B t [·] represents user scheduling, represents passive beamforming of the IRS, represents parameters of the semantic encoder-decoder, B represents B t of all time slots, that is, the scheduling strategy, represents parameters of the decoder, k represents the index of the receiving end user, represents a user set, T represents the number of time slots, t represents the index of the time slot, r represents the index of the transmitting user end, ξ represents semantic similarity, w r,k represents transmitted data, represents received data, represents the SINR at time t, represents the element on the diagonal, deg(·) represents taking the relative angle of a complex number, and N represents the number of reflecting surface transmitting units.

[0035] Further, the expression of the signal transmission optimization objective function is specifically as follows:

[0036]

[0037] In the above formula, represents the SINR at time t, represents a subset of, represents the user set selected as the transmitter at time t.

[0038] The second technical solution adopted by the application is: a multi-user semantic communication system, comprising:

[0039] A first module is configured to construct an intelligent reflecting surface assisted multi-user semantic communication system based on a diversified topology structure, wherein the multi-user semantic communication system comprises an intelligent reflecting surface and a plurality of half-duplex single-antenna users for semantic communication.

[0040] A second module is configured to perform signal transmission evaluation through a multi-user parallel transmission method of joint source channel coding based on the intelligent reflecting surface assisted multi-user semantic communication system, construct an optimized user scheduling strategy, and perform iterative optimization and solution on the optimized user scheduling strategy based on a deep reinforcement learning multi-user scheduling algorithm, construct an optimization problem of user scheduling, and realize multi-user semantic communication.

[0041] A third module is configured to perform iterative optimization and solution on the optimized user scheduling strategy based on a deep reinforcement learning multi-user scheduling algorithm, construct an optimization problem of user scheduling, and realize multi-user semantic communication.

[0042] The method and system have the advantages that the method and system construct an intelligent reflecting surface assisted multi-user semantic communication system based on diversified topological structures, further based on the intelligent reflecting surface assisted multi-user semantic communication system, perform signal transmission evaluation through a joint source channel coding multi-user parallel transmission method, use channel state information (CSI) of users as an identifier for distinguishing users, only need to train a pair of encoders and decoders regardless of how many users in the communication system, greatly reduce resources required for algorithm deployment, allow more flexible user scheduling, finally based on a multi-user scheduling algorithm of deep reinforcement learning, iteratively optimize and solve an optimized user scheduling strategy, construct an optimization problem of user scheduling, propose a hierarchical reinforcement learning framework, jointly optimize user scheduling strategy, intelligent reflecting surface and semantic encoding process, maximize throughput of the system, use sparsity of a semantic feature space, perform semantic encoding through joint source channel, use semantic information in semantic space orthogonality and channel orthogonality to assist decoding of multi-user information, and realize a multi-access based on semantic communication. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a step flowchart of a multi-user semantic communication method of the present application;

[0044] Figure 2 is a structural block diagram of a multi-user semantic communication system of the present application;

[0045] Figure 3 is a result schematic diagram of a semantic communication system provided by the embodiment of the present application;

[0046] Figure 4 is a semantic communication transmission step schematic diagram provided by the embodiment of the present application;

[0047] Figure 5 is a simulation result experiment schematic diagram provided by the embodiment of the present application;

[0048] Figure 6 is a total throughput change curve schematic diagram of the system in 5 time slots provided by the embodiment of the present application. DETAILED DESCRIPTION

[0049] The present application will be further described in detail below in combination with the drawings and specific embodiments. For the step numbers in the following embodiments, only the setting is for the convenience of description, and the order between the steps is not limited in any way, and the execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0050] With the development of communication technology, the performance of traditional communication architecture has approached the Shannon limit. However, in the face of the growing demand for transmission and limited transmission resources, the traditional communication architecture may be difficult to meet the service needs of intelligent systems, such as large-scale Internet of Things, Internet of Vehicles, industrial sensor network, edge computing system, etc. Semantic communication (SemCom) focuses on the semantic information of the transmitted data, and can further compress the transmitted data, which is a very promising method to solve the above problems. Unlike traditional data compression methods, semantic communication is based on the premise that the sender and the receiver share certain prior knowledge, extracts the semantic information of the transmitted data, and removes redundant information to achieve data compression. In addition, the semantic communication system transmits semantic information, not the bit stream based on the original data. Therefore, semantic communication has stronger robustness to errors, and can effectively complete information transmission even in a low signal-to-noise ratio environment. In summary, semantic communication has the advantages of high transmission efficiency and strong transmission robustness, and is one of the most promising technologies for designing the next generation of wireless communication systems, and is expected to be widely used in many application fields such as environmental data collection, multi-sensor information fusion, Internet of Vehicles data sharing, etc. to solve the technical problems of high data delay, low transmission rate and poor reliability in current applications.

[0051] In addition, semantic communication is considered as a very promising and application-oriented solution. Semantic communication focuses on the semantic information of the transmitted data rather than traditional data compression, and is considered to be able to break through the limitations of the Shannon paradigm and become the core technology of the next generation of communication. The present application believes that semantic communication allows us to design the coding algorithm more flexibly, so the present application utilizes the sparsity of the semantic feature space, performs semantic coding through joint source channel, utilizes the orthogonality of semantic information in the semantic space and the orthogonality of the channel to assist in decoding of multi-user information, and realizes a multi-access method based on semantic communication.

[0052] Based on this, in order to the frequency band utilization of the intelligent reflecting surface assisted multi-user semantic communication system, the embodiment of the present application combines channel orthogonality and semantic information orthogonality, and proposes a joint source channel coding multiple access method, which allows multiple semantic transmissions to be performed simultaneously in the system. Considering the finiteness of channel orthogonality and semantic information orthogonality, the present application designs a user scheduling method for the above-mentioned multiple access, which improves the overall throughput of the system under the premise of ensuring the fairness of users in the system. Since the user scheduling, channel regulation and semantic coding are coupled with each other, the parameter optimization of the multi-user semantic communication system is difficult to solve. The embodiment of the present application utilizes the monotonicity between the system performance of semantic communication and the signal-to-interference-and-noise ratio, separates the user scheduling from the semantic coding and channel coding, and proposes a two-stage iterative algorithm. The first stage of the algorithm only optimizes the scheduling strategy of the user, and the second stage jointly optimizes the channel regulation and semantic coding. Compared with the traditional baseline method and the existing semantic-based multi-user communication method, the simulation results of the present application prove that the proposed joint source channel coding multiple access method can significantly improve the throughput of the system. In addition, compared with the existing semantic method, the present application can save the computing cost of the semantic communication architecture and the resources required for deployment.

[0053] Referring to Figure 1 and Figure 4 , the present application provides a multi-user semantic communication method, which comprises the following steps:

[0054] S100, based on a diversified topology structure, an intelligent reflecting surface assisted multi-user semantic communication system is constructed, and the multi-user semantic communication system comprises an intelligent reflecting surface and a plurality of half-duplex single-antenna users for semantic communication;

[0055] In the embodiment, as Figure 3 shown, the intelligent reflecting surface assisted multi-user semantic communication system is composed of an intelligent reflecting surface and K half-duplex single-antenna users capable of semantic communication. The users are represented as a set wherein user-k represents the kth user. At each time slot t∈{1,…,T}, the algorithm of the present application gives a scheduling strategy B t ∈{0,1} K×K , wherein B t [r,k]=1 represents that user-r is sending semantic information to user-k. The (joint source channel coding method) JSCE proposed by the present application allows more diverse network topologies, as Figure 3 shown, B t [r,i]=1,B t [r,k]=1,B t[j, k] = 1 means that the three transmission relationships occur simultaneously in the same frequency in the time slot, wherein the point-to-point communication further contains one-to-many and many-to-one transmission relationships.

[0056] S200, based on the intelligent reflecting surface assisted multi-user semantic communication system, the signal transmission evaluation is carried out through the multi-user parallel transmission method of joint source channel coding, and an optimized user scheduling strategy is constructed;

[0057] Specifically, based on the intelligent reflecting surface assisted multi-user semantic communication system, a channel model between users is constructed; a joint source channel encoder is constructed in combination with a semantic encoder and a channel encoder; based on the joint source channel encoder, semantic feature extraction is performed on the data transmitted by the channel model between the users and the channel state information is converted into semantic features, and multi-user signal superposition is performed to obtain the total signal sent by the user; the total signal sent by the user is received, signal equalization processing is performed, and decoding is performed through a pre-trained decoder to obtain a reconstructed semantic feature vector; the signal-to-interference-and-noise ratio of the reconstructed semantic feature vector is obtained, and semantic throughput calculation is performed to obtain the semantic throughput; based on the semantic throughput, the semantic similarity is evaluated through structural similarity, and an optimized user scheduling strategy is constructed.

[0058] In this embodiment, first, the channel h r,k can be expressed as follows:

[0059]

[0060] where Φ is the reflection coefficient matrix of IRS, g r ,g k are the channels between user-r and user-k and the intelligent reflecting surface (IRS), respectively, and the specific calculation method conforms to the characteristics of the Rice channel. For example, g k can be calculated by the following formula, and the expression is:

[0061]

[0062] where K is the Rice factor. w r,k represents the original data from user-r to user-k, and the present application designs two encoders for the original data and channel state information (CSI), respectively, which are and where θ s ,θ c are trainable parameters of the two encoders. For the data w r,k to be transmitted, semantic extraction is first performed on it:

[0063]

[0064] semantic feature sr,k Only the original data w r,k and the information of the encoder The feature of the present application is that all users use the same encoder-decoder, and when multiple pieces of semantic information from the same encoder are superimposed, the semantic information cannot be successfully demodulated. In order to enable users to successfully demodulate the information they need from the superimposed signal, the present application converts the CSI of each user into a part of the semantic feature:

[0065]

[0066] where ⊙ represents Hadamard product. The above formula is similar in form to the beamforming process, so the present application calls this method semantic beamforming. In order to simplify the notation, we will unify the two encoders into one JSCE encoder ε θ (h r,k ,w r,k ), where θ = {θ s , θ c} is the trainable parameter. If user-r needs to serve multiple users at the same time, the total signal s r emitted by user-r can be represented as follows:

[0067]

[0068] We assume that a complete semantic information transmission can be completed within one channel coherence time, then the signal received by user-k can be represented in the following form:

[0069]

[0070] where is a Gaussian noise, the index (i) indicates that the i-th symbol in the semantic feature is transmitted, is a vector of length L, that is,

[0071] From the above formula, it can be seen that the received signal has undergone two superposition processes: the first superposition occurs at the source encoding, and the second occurs in the signal propagation process. The present application proposes that JSCE allows one transmitter to serve multiple receivers, and also allows one receiver to receive multiple transmitters at the same time, and we unify the various network topologies through the above formula. After the received signal is equalized by the channel, it is directly decoded by a decoder:

[0072]

[0073] where denotes the received complete semantic feature vector. are trainable parameters of the decoder.

[0074] According to the expression of the received signal, we can give the signal-to-interference-plus-noise ratio (SINR) of user-k when demodulating the information from user-r, which is expressed as:

[0075]

[0076] where σ 2 is the noise power, P t is the normalized signal power, and respectively represent the interference introduced by the encoding process and the propagation process, which can be calculated by the following formula:

[0077]

[0078] In order to evaluate the performance of semantic communication, the present application introduces the concept of semantic throughput, which is expressed as:

[0079]

[0080] where S is the average semantic quantity contained in the image, B represents the bandwidth of transmission, I represents the number of bits of the original data, and C r is the compression ratio of converting bits to symbols. For actual pictures, it is difficult to calculate the true value of S, so the present application uses a traditional encoding method to approximate the semantic quantity of the picture, that is, At the same time, for the JSCE proposed by the present application, there is the following relationship: C r I = L. The semantic similarity indicates that it is a function that can reflect SINR and semantic encoding parameters. For the image transmission task in this paper, the present application selects structural similarity (SSIM) as ξ, which is calculated as follows:

[0081]

[0082] where μ represents the mean value at the pixel level, σ 2 represents the variance, is the covariance of the two pictures, and c1 and c2 are two constants used to prevent the denominator from being too small.

[0083] S300, a multi-user scheduling algorithm based on deep reinforcement learning, iteratively optimizes and solves the user scheduling strategy, constructs an optimization problem of user scheduling, and realizes multi-user semantic communication.

[0084] Specifically, based on the fairness criterion of users, jointly optimize the user scheduling strategy, the passive beamforming of IRS and the parameters of the semantic encoder-decoder, maximize the minimum of the user semantic throughput, construct a signal transmission optimization objective function; decompose the signal transmission optimization objective function to obtain the minimum SINR problem of maximizing users and the semantic similarity problem of maximizing; construct a double-loop reinforcement learning algorithm, wherein the outer loop represents a reinforcement learning algorithm for determining the scheduling of users, and the inner loop represents a data-driven end-to-end training process for optimizing the semantic encoding strategy and the passive beamforming of IRS; based on the outer loop, use the DDPG algorithm to generate a scheduling strategy to solve the minimum SINR problem of maximizing users, and obtain the user scheduling strategy; based on the inner loop, jointly encode the semantic source channel to solve the semantic similarity problem of maximizing, and obtain the passive beamforming of IRS and the semantic encoding strategy; combine the channel characteristics and data characteristics to construct a joint optimization problem of user scheduling, IRS passive beamforming and semantic encoding strategy, and realize multi-user semantic communication.

[0085] In the embodiment, JSCE can realize parallel transmission of multiple users through joint source channel encoding, but too many users accessing at the same time will still cause interference, so the application considers a multi-time slot scheduling problem. Since we assume that all users in the system are homogeneous, the application starts from the user fairness, jointly optimizes the user scheduling B t , the passive beamforming Φ of IRS and the parameters θ of the semantic encoder-decoder, and maximizes the minimum of the user semantic throughput:

[0086]

[0087] The constraint one is the IRS reflection coefficient limit, and the constraint two is to limit all users in the half-duplex mode. The above optimization problem cannot be solved by traditional methods because there is no explicit expression between ξ and . And in order to ensure that users can successfully demodulate information, it is necessary to ensure the orthogonality of semantics and the spatial orthogonality of the channel at the same time.

[0088] In order to solve this complex problem, the application decomposes the above problem into two sub-problems, namely maximizing the minimum SINR of users and maximizing the semantic similarity, and then alternately optimizes the two sub-problems. The application proposes an interpretable reinforcement learning algorithm. The algorithm has two loops, the outer loop is a reinforcement learning algorithm for determining the scheduling of users B t , and the inner loop is a data-driven end-to-end training process for optimizing the semantic encoding strategy and the passive beamforming Φ of IRS.

[0089] Given the scheduling strategy B tFinally, since all users are homogeneous and all users have the same encoder, we only need to focus on maximizing the semantic similarity. Subproblem 1 can be expressed as follows:

[0090]

[0091] In order to solve sub-problem 1, the present invention uses a deep neural network (DNN) to fit the encoding and decoding process of semantic communication. The backbone of the DNN is mainly composed of a basic residual module (BRB), an inverse residual module (IBRB) and a channel attention module (CAB). BRB is a module in Resnet, which is equivalent to a convolution operation with a convolution kernel size of 3 and a convolution step size of 2. IBRB is implemented by replacing the convolution in BRB with deconvolution. During the encoding process, in order to promote DNN to better fit the distribution characteristics of user CSI, the present invention first maps the scalar or low-dimensional CSI to a high-dimensional vector space, and calls this process channel information vectorization (C2V). The C2V mapping requires that similar CSI also show certain similarities in the vector space, and uses position encoding as a specific implementation method of C2V. Without loss of generality, the present invention considers a 2-D scenario and assumes that all users are distributed in the same plane to simplify the mathematical model. Given a user's CSI g k , we can get the vector feature corresponding to the CSI as e k =C2V(g k ). Then the CAB module calculates the feature map s r,k The channel mean vector, the mean vector and e k After fusion, a multi-layer perceptron (MLP) and a softmax function are used to obtain the final CSI-based attention weight:

[0092] a r,k =softmax(MLP(s r,k +e k )).

[0093] IRS can not only improve the channel spatial orthogonality during the coding process, but also indirectly participate in a r,k In order to optimize the passive beamforming of IRS, the present invention regards IRS as a part of DNN structure. During the training process, the IRS reflection coefficient Φ will also be updated through back propagation. After the training is completed, the constraints of the IRS reflection coefficient may not be fully satisfied, so the present invention quantizes the neural network layer and quantizes the phase angle of the reflection coefficient to {0,π}. For the optimization goal of maximizing semantic similarity, the present invention uses the mean square error (MSE) loss function for joint optimization. and Φ. The training process is similar to autoencoder, the input image is used as the ground truth to calculate the MSE loss between the output of DNN and the input.

[0094] After fixing the passive beamforming and semantic encoding strategy of IRS, the optimization problem of user scheduling can be rewritten as follows:

[0095]

[0096] The solution of the scheduling strategy is defined on a discrete set, which makes the problem difficult to solve.

[0097] In summary, the embodiments of the present application have the following technical differences compared with the prior art:

[0098] 1) The present application uses the sparsity of high-dimensional feature space to improve the orthogonality of semantic features by designing a certain encoding method, which is used to assist multiple access and improve the frequency band utilization.

[0099] 2) The present application proposes a joint source-channel encoding method (JSCE) that uses the channel state information (CSI) of users as the identifier to distinguish users. Only one pair of encoder and decoder is needed regardless of the number of users in the communication system, which greatly reduces the resources required for algorithm deployment and allows more flexible user scheduling. In addition to traditional point-to-point transmission, the present application also provides a one-to-many and many-to-one network topology, further improving the system throughput.

[0100] 3) The present application uses the proposed JSCE to design a multi-slot user scheduling strategy. We divide the time axis into multiple time slots, and only allow a part of users to use semantic multiple access for parallel transmission in each time slot. The scheduled users and the reflecting surface will affect the orthogonality of the channel, and the semantic encoder based on deep learning will affect the orthogonality of the semantics. Therefore, the present application proposes a hierarchical reinforcement learning framework to jointly optimize the user scheduling strategy, the intelligent reflecting surface and the semantic encoding process, maximizing the system throughput.

[0101] Therefore, the embodiments of the present application have the following advantages compared with the prior art:

[0102] 1) The proposed JSCE is based on semantic multiple access using joint source-channel encoding, which uses channel orthogonality and semantic orthogonality to realize multi-user decoding and achieve concurrent simultaneous transmission.

[0103] 2) The proposed JSCE is designed to optimize the user scheduling, intelligent reflecting surface and semantic encoding process. The algorithm adjusts the subsequent user scheduling strategy based on historical scheduling records, and then optimizes the reflecting surface and semantic encoding based on user scheduling, thereby maximizing the system throughput while ensuring user fairness.

[0104] 3) The MDP process modeling and deep reinforcement learning control method are designed based on the multi-time slot multi-user scenario. The semantic encoding based on deep learning is used, and the user scheduling, intelligent reflecting surface and neural network parameters are coupled together, which is a complex high-dimensional control problem. The MDP method is used to model the task, and then the DDPG algorithm is used to train the model.

[0105] Therefore, the JSCE proposed in the embodiment of the present application uses the channel state information (CSI) of the user as the identifier for distinguishing the user, and only one pair of encoder and decoder needs to be trained regardless of the number of users in the communication system, which greatly reduces the resources required for algorithm deployment and allows more flexible user scheduling. Compared with the point-to-point transmission provided by the existing semantic multiple access, the JSCE also provides two possible network topologies of one-to-many and many-to-one, which further improves the flexibility of user scheduling and improves the throughput of the system.

[0106] Finally, the JSCE and user scheduling algorithm designed in the present application are verified in simulation simulation experiments, and the results are as follows. The simulation environment is a multi-user semantic communication scenario, and the model is as shown in Figure 3 There are a total of 5 users in the scene, and the total duration contains 5 time slots, i.e. K=5, T=5. The intelligent reflecting surface is placed at the origin (0, 0, 0), and the coordinates of the users are randomly generated, as follows: (1.13, 0.50), (-0.01, -0.21), (-1.10, -0.28), (0.19, 1.01), (0.20, 0.01). The noise power is 2 =0.1, and the Rician factor K=10.

[0107] The feasibility analysis is shown in the simulation results as Figure 5 The user located at (1.13, 0.50) is scheduled as the sender, and simultaneously sends three different 512x512 RGB pictures to the users located at (-0.01, -0.21), (-1.10, -0.28), and (0.19, 1.01), Figure 3are the results obtained by three receiving end users respectively demodulating. As shown in the figure, the JSCE proposed by the application can achieve results comparable to the existing semantic multiple access DeepMA with the help of the reflecting surface, and the JSCE occupies only 27.29MB of storage space, while the DeepMA needs to occupy 147.7MB of space under the same environment. Figure 5 It is also shown that the intelligent reflecting surface has a PSNR improvement of about 4dB for image transmission of JSCE, proving that JSCE combines the use of channel orthogonality and semantic orthogonality, rather than simply relying on neural networks and semantic orthogonality.

[0108] effectiveness analysis, Figure 6 As shown, in the above 5-user scenario, the total throughput curve of the system in 5 time slots during the training process of the deep reinforcement learning user scheduling algorithm proposed by the application is shown. Among them, time division multiple access (TDMA) is combined with bit communication and semantic communication as a baseline algorithm, and then DeepMA and semantic NOMA are used as comparative algorithms. In order to control the variable, the semantic encoding process used by semantic NOMA is consistent with JSCE. Semantic NOMA mainly relies on channel differences when dealing with multi-user problems, without utilizing the orthogonality of semantic information; DeepMA only utilizes the orthogonality of semantics, without considering channel differences, and the semantic encoding model required by DeepMA occupies a larger space. In summary, both of the above comparative methods have certain limitations, and JSCE can allow more users to access by jointly utilizing the orthogonality of the source and channel. As shown in the figure, the JSCE method proposed by the application can significantly improve the throughput of the system.

[0109] Reference Figure 2 A multi-user semantic communication system, comprising:

[0110] The first module 201 is configured to construct an intelligent reflecting surface assisted multi-user semantic communication system based on a diversified topology structure, wherein the multi-user semantic communication system comprises an intelligent reflecting surface and a plurality of half-duplex single-antenna users for semantic communication.

[0111] The second module 202 is configured to perform signal transmission evaluation based on the intelligent reflecting surface assisted multi-user semantic communication system by using a multi-user parallel transmission method of joint source channel coding, and construct an optimized user scheduling strategy.

[0112] The third module 203 is configured to iteratively optimize and solve the optimized user scheduling strategy based on a deep reinforcement learning multi-user scheduling algorithm, construct an optimization problem of user scheduling, and realize multi-user semantic communication.

[0113] The contents in the method embodiments are applicable to the system embodiments, the system embodiments specifically implement the same functions as the method embodiments, and achieve the same beneficial effects as the method embodiments.

[0114] The above is a specific description of the preferred embodiments of the application, but the application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the application. These equivalent modifications or replacements are all included in the scope defined by the claims of the application.

Claims

1. A multi-user semantic communication method, characterized in that: The following steps are involved: Based on a diversified topology, a multi-user semantic communication system assisted by an intelligent reflector is constructed. The multi-user semantic communication system includes an intelligent reflector and several half-duplex single-antenna users performing semantic communication. A multi-user semantic communication system assisted by intelligent reflectors uses a multi-user parallel transmission method with joint source-channel coding to evaluate signal transmission and build an optimized user scheduling strategy. The multi-user scheduling algorithm based on deep reinforcement learning iteratively optimizes and solves the optimized user scheduling strategy, constructs the optimization problem of user scheduling, and realizes multi-user semantic communication.

2. A multi-user semantic communication method according to claim 1, characterized in that: The multi-user semantic communication system based on intelligent reflective surface assistance performs signal transmission evaluation through a multi-user parallel transmission method with joint source-channel coding, and constructs an optimized user scheduling strategy, which specifically includes: A multi-user semantic communication system assisted by intelligent reflective surfaces is used to build a channel model between users; Combine the semantic encoder and the channel encoder to construct a joint source-channel encoder; Based on the joint source-channel encoder, semantic features are extracted from the data transmitted by the channel model between users, the channel state information is converted into semantic features, and multi-user signals are superimposed to obtain the total signal sent by the user; Receive the total signal sent by the user, perform signal equalization processing and decode it through a pre-trained decoder to obtain a reconstructed semantic feature vector; Obtain the signal-to-interference-and-noise ratio of the reconstructed semantic feature vector and perform semantic throughput calculation to obtain the semantic throughput; Based on semantic throughput, the semantic similarity is evaluated through structural similarity to build an optimized user scheduling strategy.

3. A multi-user semantic communication method according to claim 2, characterized in that: The expression of the channel model between the users is specifically as follows: In the above formula, h r,k represents the channel model, r represents user r, k represents user k, g r ,g k are the channels between user r and user k and the intelligent reflecting surface (IRS), and Φ represents the reflection coefficient matrix of IRS.

4. A multi-user semantic communication method according to claim 3, characterized in that: The expression of the total signal sent by the user is specifically as follows: In the above formula, s r represents the total signal sent by user r, B t represents the scheduling strategy, r represents user r, j represents user j, K represents the total number of users, represents the weighted signal.

5. A multi-user semantic communication method according to claim 4, characterized in that: The expression for calculating the semantic throughput is specifically as follows: In the above formula, Γ r,k represents semantic throughput, S represents the average semantic amount contained in the image (unit is the number of symbols), B represents the transmission bandwidth, I represents the number of bits of the original data, and C r Indicates the compression ratio of converting bits into the number of symbols, represents the trainable parameters of the decoder, ξ(·) represents the structural similarity, and γ r,k represents the signal-to-drying ratio, and θ represents the parameters of the semantic decoder.

6. A multi-user semantic communication method according to claim 5, characterized in that: The multi-user scheduling algorithm based on deep reinforcement learning iteratively optimizes and solves the optimized user scheduling strategy, constructs the optimization problem of user scheduling, and realizes the multi-user semantic communication step, which specifically includes: Based on the user fairness criterion, the user scheduling strategy, IRS passive beamforming, and semantic codec parameters are jointly optimized to maximize the minimum value of user semantic throughput and construct a signal transmission optimization objective function. The signal transmission optimization objective function is decomposed to obtain the problem of maximizing the minimum SINR of the user and the problem of maximizing the semantic similarity; A dual-loop reinforcement learning algorithm is constructed, where the outer loop represents a reinforcement learning algorithm for determining user scheduling, and the inner loop represents a data-driven end-to-end training process for optimizing the semantic encoding strategy and passive beamforming of the IRS. Based on the outer loop, the DDPG algorithm is used to generate a scheduling strategy to solve the problem of maximizing the minimum SINR of the user and obtain the user scheduling strategy; Based on the inner loop, semantic encoding is performed on the combined source channel to solve the problem of maximizing semantic similarity and obtain the passive beamforming and semantic encoding strategies of the IRS. Combining channel characteristics with data characteristics, a joint optimization problem of user scheduling, IRS passive beamforming, and semantic coding strategy is constructed to achieve multi-user semantic communication.

7. A multi-user semantic communication method according to claim 6, characterized in that: The expression of the signal transmission optimization objective function is specifically as follows: In the above formula, B t [·] represents user scheduling, Φ represents the passive beamforming of IRS, θ represents the parameters of semantic codec, and B represents B of all time slots. t The set of scheduling strategies, represents the decoder parameters, k represents the receiving end user index, represents the user set, T represents the number of time slots, t represents the time slot index, r represents the index of the sending user end, ξ represents the semantic similarity, w r,k Indicates sending data. Indicates receiving data. represents the SINR at time t, express For the elements on the diagonal, deg(·) represents the relative angle of the complex number, and N represents the number of emitting units on the reflective surface.

8. A multi-user semantic communication method according to claim 7, characterized in that: The expression of the signal transmission optimization objective function is specifically as follows: In the above formula, represents the SINR at time t, express A subset of , representing the set of users selected as transmitters at time t.

9. A multi-user semantic communication system, characterized in that: Includes the following modules: The first module is used to build a multi-user semantic communication system assisted by an intelligent reflector based on a diversified topology structure. The multi-user semantic communication system includes an intelligent reflector and a plurality of half-duplex single-antenna users performing semantic communication; The second module is used for a multi-user semantic communication system assisted by an intelligent reflector. It uses a multi-user parallel transmission method with joint source-channel coding to evaluate signal transmission and build an optimized user scheduling strategy. The third module is used to iteratively optimize and solve the optimized user scheduling strategy based on the multi-user scheduling algorithm based on deep reinforcement learning, construct the optimization problem of user scheduling, and realize multi-user semantic communication.

Images