D2D multicast communication method and system for semantic communication

Abstract

The present invention discloses a D2D multicast communication method for semantic communication, which realizes D2D multicast communication through semantic information and provides a new paradigm for D2D multicast communication. The method includes: a base station broadcasts an image semantic triple structure; calculates a user similarity index; realizes user clustering; trains image semantic codecs with different compression ratios; and designs semantic compression ratios and channel allocation strategies during semantic communication. In the method, semantic codecs are trained through convolutional neural networks, deep learning, etc., and game theory is introduced as a research framework. The utility function is calculated in a distributed manner through continuous iteration until convergence. The optimization method can be widely used in the fields of D2D communication and semantic communication.

Description

Technical Field

[0001] The present invention belongs to the technical field of D2D communication, and in particular to a D2D multicast communication method oriented to semantic communication. Background Art

[0002] In recent years, the rapid development of D2D communication technology has greatly facilitated people's communication experience. D2D multicast communication, as a common method of D2D communication technology, has attracted more and more attention. Specifically, the base station first transmits the pushed content, such as pictures, text, video and other data, to users with better channel conditions. These users are called cluster heads. Then the cluster heads transmit the pushed content to other users, which can greatly reduce the overhead of base station communication resources. In order to improve spectrum efficiency, the cluster head communicates with the target users in the cluster by multiplexing cellular links, and multiple cluster heads are allowed to access the same link. In this way, interference occurs between clusters. How to effectively eliminate interference and improve the service quality of target users has always been a hot issue in D2D communication.

[0003] The traditional way to achieve interference elimination is to use wireless resource management. This method improves user service quality through reasonable channel allocation strategy and power control, and this method has also been widely studied. However, with the rapid increase in the number of mobile users and massive data resources, coupled with the shortage of spectrum resources, the performance improvement of traditional wireless resource management solutions has reached a bottleneck, and it is urgent to develop new technical means to improve the communication ceiling.

[0004] Semantic communication is emerging as a new communication paradigm and has received widespread attention. Compared with traditional communication methods that transmit data in compressed bit streams, semantic communication focuses more on revealing the potential meaning in the original data to facilitate the communication process. By organically combining user needs with semantic features, semantic communication can significantly reduce the amount of information transmission, alleviate the communication burden, and reduce co-channel interference. On the other hand, the acquisition of semantic information is usually abstract and subjective, resulting in possible differences in the semantic information obtained for different tasks. Therefore, semantic communication places more emphasis on the actual user experience, and its ultimate goal is to facilitate the communication process in a task-oriented manner.

[0005] At present, there are some papers that study the core issue of using semantic information in wireless communications. The paper [Energy Efficient Semantic Communication Over Wireless Networks With Rate Splitting, IEEE J.Sel.Areas Commun., vol.41, no.5, pp.1484-1495, May 2023] considers the use of efficient semantic communication in wireless rate splitting networks and studies wireless resource allocation and semantic information extraction. The paper [Heterogeneous Semantic and Bit Communications: A Semi-NOMA Scheme, IEEE J.Sel.Areas Commun., vol.41, no.1, pp.155-169, Jan.2023] combines semantic communication with non-orthogonal multiple access technology and analyzes the performance of semantic communication and traditional bit communication under different signal-to-noise ratio conditions. The paper [Semantic Communication-Based Dynamic Resource Allocation in D2D Vehicular Networks, IEEE Trans. Veh. Technol., vol. 72, no. 8, pp. 10784-10796, Aug. 2023] considers the transmission of semantic information in D2D vehicular networks and solves the challenges of reliability and stability of semantic information transmission. The paper [Selection-Based Image Generation for Semantic Communication Systems, IEEE Commun. Lett., vol. 28, no. 1, pp. 34-38, Jan. 2024] proposes a semantic fidelity index to characterize the mutual information between the original image and the reconstructed image and the similarity of the neural network features, providing a practical and effective solution for image communication scenarios with limited bandwidth.

[0006] Existing literature has considered the application of semantic communication in wireless communication technology. However, the acquisition of semantic information is usually abstract and subjective, which may lead to differences in semantic information obtained for different tasks, which faces challenges in D2D multicast communication networks. On the other hand, the extraction process of semantic information is flexible and diverse. How to reasonably select semantic models in practical applications and how to give full play to the advantages of semantic information to improve the user's communication experience are still unresolved issues. These challenges have caused certain difficulties for the further development of semantic communication. Summary of the invention

[0007] The purpose of the present invention is to provide a semantic and wireless resource allocation scheme for D2D multicast semantic communication to address the problems existing in the prior art, and to provide an efficient resource management method with the average quality of experience of target users in a cell as the goal. The present invention innovatively utilizes semantic communication to enhance the delivery process of D2D multicast content.

[0008] The technical solution to achieve the purpose of the present invention is: a D2D multicast communication method for semantic communication, the method comprising the following steps:

[0009] Step 1, the base station broadcasts semantic information of the image to be transmitted, wherein the semantic information is a triple structure;

[0010] Step 2, defining a search list: the target user receives the semantic information of the image sent by the base station, selects an interest preference for the image according to the semantic information, and sets an interest list, i.e., a search list;

[0011] Step 3: cluster users based on the similarity between target users;

[0012] Step 4, constructing image semantic compression models with different compression ratios;

[0013] Step 5, perform semantic communication: construct a game model based on the service experience quality requirements of the target users in each cluster to solve the problem of maximizing the average experience quality of the target users within the area; the base station selects the image semantic compression ratio, inputs it into the image semantic compression model, and transmits the output corresponding semantic information to the cluster head. The cluster head reuses the transmission link of the cellular user and selects the transmission link to broadcast the information to the users in the cluster, completing the base station image transmission task.

[0014] Furthermore, step 1 specifically includes:

[0015] Semantically encode the image to be transmitted;

[0016] Based on the semantic segmentation technology of relational transformer and deep neural network, the triple structure of the image to be transmitted is obtained, which is expressed as follows:

[0017] τ n =D(x n )

[0018] Among them, τ n represents the set of triple structures, represents the semantic content of image n, m n represents the total number of triplets in image n, represents the first A triplet, x nDenotes the input image n, and D denotes a function of the semantic segmentation technique based on the relational transformer and deep neural network.

[0019] Furthermore, in step 2, the interest list of target user i for a set of images is expressed as:

[0020] q i =[q i,1 ,...,q i,2 ,q i,N ] T

[0021] Where N represents the total number of images pushed by the base station to the target user i, q i,n represents the interest level of target user i in image n, q i,n =1 means that target user i is interested in image n, q i,n =0 means that target user i has no interest in image n.

[0022] Furthermore, step 3 specifically includes:

[0023] Step 3-1, calculating the similarity between target users based on their interest lists and physical locations;

[0024] Step 3-2: Based on the similarity, target users are clustered using an improved k-center algorithm.

[0025] Furthermore, step 3-1 specifically includes:

[0026] Assume that the physical location of target user i is represented as (x i ,y i ), the normalized position coordinates are R is the radius of the target user’s area; similarly, the normalized interest vector is

[0027] Then the similarity between any two target users i and j is expressed as:

[0028]

[0029] Among them, F sim (u i ,u j ) represents the similarity between target users i and j, μ1 and μ2 represent the weights of interest similarity and physical location similarity, respectively. Respectively represent the normalized position coordinates of target users i and j, They represent the normalized interest vectors corresponding to target users i and j respectively; α is the large-scale fading coefficient.

[0030] Furthermore, step 3-2 specifically includes:

[0031] (1) Assume that the target users are divided into K clusters, and the set of cluster heads is represented as Ω = {h1,h2,...,h K}, h K represents the cluster head of the Kth cluster, and the cluster is represented as represents the Kth cluster, and the users in the cluster are represented as represents the first target users, represents the total number of target users in cluster k;

[0032] In order to ensure the similarity of users in the cluster, the following constraints are given:

[0033]

[0034] in, is the preset similarity threshold,

[0035] The silhouette coefficient is introduced to measure the clustering effect:

[0036]

[0037] Among them, the intra-cluster similarity is defined as:

[0038]

[0039] The inter-cluster similarity is defined as the maximum similarity between the target user and other clusters, expressed as:

[0040]

[0041] In the formula, SC k,i represents the silhouette coefficient corresponding to the target user i in cluster k, a k,i represents the similarity of target user i in cluster k, b k,i represents the maximum similarity between the target user i in cluster k and other clusters, represents the similarity between target users i and j in cluster k, represents the similarity between target user i in cluster k and target user j in cluster k', represents the k'th cluster, k'=1,2,...,K; represents the total number of target users in cluster k';

[0042] The purpose of clustering is to maximize the following objectives:

[0043]

[0044] Then the clustering problem in step 3-2 is expressed as:

[0045]

[0046] (3) Based on the definition in (1) above, the target users are clustered using the improved k-center algorithm. The specific process includes:

[0047] Step 3-2-1, enter the similarity values ​​of all target users and set the maximum number of cycles;

[0048] Step 3-2-2, set the number of clusters K, where N u Indicates that there are N u target users;

[0049] Step 3-2-3, for each given number of clusters K, loop, set the number of iterations t = 0, and iterate the following process;

[0050] Step 3-2-4, for each iteration t, initialize the cluster head to represents the cluster head of the Kth cluster in the tth iteration; for each cluster k, Cycle the following process;

[0051] Step 3-2-5, Settings

[0052] Step 3-2-6, for each cluster k∈[1,K], select the target user u ω , where u ω is the non-cluster head user in the cluster;

[0053] Step 3-2-7, if but Among them, F sim (u ω ,h k ) represents user u ω With cluster head h k similarity;

[0054] Step 3-2-8, if The target user u ω Enter cluster k * , Otherwise, user u ω are considered isolated points;

[0055] Step 3-2-9, traverse all users u ω, calculate the corresponding silhouette coefficient as SC0; randomly select a target user in cluster k as the cluster head, and loop through steps 3-2-4 to 3-2-8 to calculate a new silhouette coefficient SC1;

[0056] Then calculate:

[0057] ΔSC=SC1-SC0

[0058] If ΔSC is a positive number, the cluster head of cluster k is changed to the randomly selected user, otherwise, it remains unchanged;

[0059] The iteration ends until the maximum number of cycles is reached;

[0060] Step 3-2-10, compare the silhouette coefficients under all clustering numbers, where the maximum value is the optimal allocation number K;

[0061] Step 3-2-11, output the optimal number of clusters and the clustering and cluster head results.

[0062] Furthermore, step 4 specifically includes:

[0063] Step 4-1, construct the objective function of the image semantic codec with different compression ratios:

[0064] Construct a full-scale semantic extraction and reconstruction model, and the parameterized model weight coefficient is expressed as (ξ,ψ);

[0065] For the sub-semantic model derived from it, under a given ratio compression factor γ k After that, the parameterized model weight coefficient is expressed as

[0066] The objective function is defined as:

[0067]

[0068] in, represents the semantic reconstruction function under given parameter conditions, |Λ| represents the number of elements in the set Λ, represents the loss function; Indicates that the parameter set is The semantic encoding of Indicates that the parameter set is Semantic decoding of

[0069] Step 4-2, based on the objective function, using a small batch dynamic sub-model training method to train image semantic codecs with different compression ratios to generate image semantic compression models with different compression ratios;

[0070] The specific process includes:

[0071] Step 4-2-1, input training set X sample ,(ξ,ψ), set the maximum number of cycles;

[0072] Step 4-2-2, divide the proportional compression factor into N equal batches, and randomly select |Λ n | proportional compression factors and divided into n groups, where n = 1, 2, ..., N, and randomly select one of the batches Λ n ;

[0073] Step 4-2-3, for each batch of training data x batch ∈X sample ;

[0074] Loop, iterating the following process:

[0075] Step 4-2-4, given the proportional compression factor γ, calculate the decoding result of the size width under γ

[0076] Step 4-2-5, use the full-size model to calculate the decoding result under the full-size model

[0077] Step 4-2-6, calculate the loss function:

[0078]

[0079] Step 4-2-7, back propagation, update (ξ,ψ);

[0080] Step 4-2-8, output the weight coefficients (ξ, ψ) of the full-size semantic extraction and reconstruction model, and generate image semantic compression models with different compression ratios.

[0081] Furthermore, step 5 specifically includes:

[0082] Step 5-1, construct the following optimization problem P0 according to different user needs:

[0083] P0:

[0084] stC1:

[0085] C2:

[0086] C3:

[0087] C4:

[0088] in,

[0089]

[0090] In the formula, is the experience quality indicator of target user n in cluster k. C1, C2, C3, and C4 are all constraints. is the delay of target user n in cluster k, yes The threshold of k,n ∈[0,1] represents the weight of the fidelity of target user n in cluster k, 1-μ k,n is the weight of the target user n’s delay; λ1,λ2,λ3 are the parameters that determine the range of the logic function change, ensuring that the scores of fidelity and delay are between [0,1]; μ k,n =0 means that the target user n in cluster k is more concerned about receiving fast recommended content and only needs to meet the content requirements; μ k,n =1 means that the user is more concerned about the image quality of the recommended content, rather than the timeliness of the recommended content; N u Indicates that there are N u target users; min Represents γ k The lower limit of the value range; ρ(s k ,l) represents the indicator function, l is the designated channel, s k represents the cluster head h k The link selection strategy set, represents the cluster head h k The transmission power, is the cluster head h k The channel gain between the base station and the m For cellular users m The maximum tolerable interference value; K' represents the number of bits of unit semantic information; S k is the amount of semantic data of the transmitted image in cluster k under the full-size model; Cluster The achievable rate of the receivers in the cluster; is the compression factor γ of cluster k at a given ratio k The semantic fidelity of the corresponding sub-model; (σ1, σ2, σ3, σ4) are hyperparameters; For target users The semantic fidelity threshold of

[0091] Step 5-2, using the cluster head link selection algorithm based on parameter correction to solve the optimization problem P0, the specific process includes:

[0092] Step 5-2-1, generating a cluster transmission link selection strategy and an image semantic compression ratio that meet the above C1, C2, C3, and C4 constraints as initial variables;

[0093] The following process is executed in a loop:

[0094] Step 5-2-2, randomly select cluster k, and use one-dimensional enumeration method to update the optimal image semantic compression ratio that meets the constraints;

[0095] Step 5-2-3, update the transmission link selection strategy of cluster k, calculate the utility function of all optional transmission link selection strategies of the cluster head corresponding to cluster k, and then follow the hybrid strategy π k (i+1) selection strategy, updated using the Boltzmann exploration strategy criterion:

[0096]

[0097] Among them, U k (l,s -k ) represents the utility of the cluster head corresponding to cluster k when selecting the transmission link l', σ(i+1)≥0 is the learning parameter, which is positively correlated with the number of iterations. is the probability that the cluster head corresponding to cluster k selects the transmission link l', s -k represents the transmission link selection strategy of clusters other than cluster k, s k represents the transmission link selection strategy corresponding to cluster k, U k (s k ,s -k ) represents the utility function of cluster k, represents the set of transmission link selection strategies corresponding to cluster k;

[0098] Step 5-2-4, update the transmission link selection strategy s k ;

[0099] The loop ends until all target users no longer change their transmission link selection strategies and semantic compression ratios.

[0100] Furthermore, the utility function calculation formula in step 5-2-3 is:

[0101]

[0102] in,

[0103]

[0104] Where V k (s k ,s -k ) is the benefit of cluster k, V k′ (s k′ ,s -k′ ) is the benefit of cluster k' that does not take the transmission link selection strategy action, represents the clusters in the cluster set C except cluster k, s k' represents the transmission link selection strategy of cluster k', s -k' represents the transmission link selection strategy of other clusters except cluster k', V k (s k′ ,s -k′\k ) is the benefit of cluster k' after cluster k adopts the transmission link selection strategy.

[0105] The present invention takes advantage of semantic communication. On the one hand, it uses semantic feature analysis to perform semantic inference on image transmission content, thereby obtaining the real-time preferences of target users. On the other hand, it uses semantic coding technology to communicate with a small amount of transmission data, reconstructs the image at the user, and optimizes the management of semantic resources and wireless resources with the goal of maximizing the user's quality of experience.

[0106] Compared with the prior art, the present invention has the following significant advantages:

[0107] (1) This method considers the image transmission process. First, the base station broadcasts a small amount of image semantic information to the target users in the cell. The target users are divided into different clusters according to their physical location and similarity of interest in the image, and a cluster head is selected. The image is further semantically compressed at different ratios at the base station. According to the different requirements of different clusters for image fidelity and latency, different amounts of semantic information are transmitted to different cluster heads, thereby maximizing the average experience quality of all target users in the cell.

[0108] (2) The present invention introduces game theory as an analysis and solution tool. The original problem is transformed into an exact potential game, and an improved spatial adaptive resource allocation algorithm is proposed to find the Nash equilibrium point, which has certain advantages in terms of convergence and optimality.

[0109] (3) The present invention improves the traditional k-center algorithm and can achieve efficient user clustering.

[0110] (4) The present invention proposes a small-batch dynamic sub-model training method, which can quickly train semantic coding models with multiple-size compression ratios.

[0111] (5) The present invention proposes a cluster head link selection algorithm based on parameter modification. The algorithm can adaptively change its learning parameters according to the number of iterations. When the number of iterations is large, the learning step size increases, which can converge more quickly; when the number of iterations is small, the step size decreases, which can fully explore unknown areas; the decision variables are expanded from the traditional one-dimensional to two-dimensional, and the semantic compression ratio within the cluster is integrated with the cluster head link selection, and the two independent variables are treated as combined variables.

[0112] (6) This optimization strategy can be widely applied to semantic communication and D2D communication scenarios, but is not limited to the scope listed above.

[0113] The present invention is further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 It is a flow chart of a D2D multicast communication method for semantic communication.

[0115] Figure 2 This is a schematic diagram of a D2D multicast communication model based on semantic information.

[0116] Figure 3 The diagram is a schematic diagram of how clustering silhouette coefficients vary with similarity thresholds under different algorithms in one embodiment.

[0117] Figure 4 The figure is a schematic diagram showing how the loss function of the algorithm training process in an embodiment changes with the number of training cycles.

[0118] Figure 5 The figure is a schematic diagram showing the changes of user fidelity and delay scores with weight coefficients in an embodiment.

[0119] Figure 6 The diagram is a schematic diagram of how the average user experience quality changes with the number of clusters under different algorithms in an embodiment. DETAILED DESCRIPTION

[0120] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0121] It should be noted that if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0122] The present invention provides a D2D multicast communication method for semantic communication, comprising:

[0123] Step 1: The base station broadcasts the semantic information of the transmitted image: semantic encoding of the image. Since the encoding of image data is often complex, the semantic segmentation technology of deep neural network can be used to quickly obtain the triple structure of the target image. This structure has a small amount of data and is suitable for the base station to quickly and efficiently share the image information content.

[0124] Step 2: Define the search list: The target user receives the semantic information of the image sent by the base station, selects his interest preference for the image according to the semantic information, and sets his interest list;

[0125] Step 3: User clustering process: Based on the interest list and physical location of the target users, define the similarity between the target users. The target users with high similarity are divided into a cluster. Define the silhouette coefficient to measure the performance index of the clustering result. Improve the K-center algorithm, use this algorithm to implement user clustering, and select the appropriate cluster head.

[0126] Step 4: Image semantic compression process: Next, the base station performs semantic encoding on the image by using a flexible semantic model to characterize the image fidelity, considering semantic encoders of different sizes, outputting semantic information of different data volumes, and using different semantic encoders and decoders according to the different requirements of each cluster for image fidelity and latency.

[0127] Step 5: Semantic communication process: According to the service experience quality requirements of the target users in each cluster, a game model is constructed to solve the problem of maximizing the average experience quality of the target users in the cell. The base station selects an appropriate image semantic compression ratio and transmits the corresponding semantic information to the cluster head. The cluster head reuses the links of cellular users and selects an appropriate link to broadcast the information to the users in the cluster, completing the base station image transmission task.

[0128] In one embodiment, according to Figure 1 Given the implementation scenario diagram, the implementation steps of the D2D multicast semantic communication solution proposed in the present invention mainly include:

[0129] Step 1: We adopt a novel end-to-end scene graph generation framework, called Relation Transformer, which simplifies the acquisition of image triplets. The framework can detect triplet structures with only visual appearance and simultaneously predict a fixed set of structures. This process follows an encoder-decoder architecture, where the encoder infers the visual feature context and the decoder exploits this context for triplet inference. Note that this process can be implemented with other pre-trained image generation techniques based on different tasks. In any case, a set of image triplets can be inferred from image n, represented as:

[0130] τ n =D(x n )

[0131] where τ n represents the set of triple structures, represents the semantic content of image n, m n represents the total number of triplets in image n, x n Represents the input image n. The base station transmits this triple graph structure to the target users in the cell in the form of broadcast. The target users match this triple graph structure with their own preferences and interests. The image with the highest matching score is the image they want to retrieve most.

[0132] Step 2: Assume that the total number of target users in the cell is N u , the interest list of target user i for a set of images is q i =[q i,1 ,...,q i,2 ,q i,N ] T , where N represents the total number of images that the base station wants to push to the target user, In this way, each target user has an interest vector related to himself. This encoding method is inspired by the one-hot encoding commonly used in machine learning. Indicates that user i is interested in image n, In particular, the interest vector corresponding to the base station is a vector in which all elements are 1.

[0133] Step 3: User clustering process

[0134] First, the concept of similarity is defined. In the present invention, similarity is defined to consider the similarity of interests between users, while also taking into account the proximity of physical locations, so that the communication process is more efficient. Assume that the physical location of the target user is represented by (x i ,y i ), so the normalized position coordinates can be obtained as Similarly, the normalized interest vector is Therefore, the similarity between any two users i and j can be expressed as:

[0135]

[0136] Where μ1 and μ2 represent the weights of interest similarity and physical location similarity. Usually, they take the same value. α is regarded as a large-scale fading coefficient, which is more realistic in definition. On the other hand, assuming that the target users are divided into K clusters, the set of cluster heads is represented by Ω = {h1,h2,...,h K}, the clustering is represented as The users in the cluster are represented as In order to ensure the similarity of users in the cluster, the following constraints are given:

[0137]

[0138] in, is the given similarity threshold, Indicates the similarity between user i and the cluster head in cluster k. On the other hand, in order to characterize the quality of clustering results, the silhouette coefficient is introduced to measure the clustering effect, and the intra-cluster similarity index is defined as:

[0139]

[0140] Similarly, the inter-cluster similarity is defined as the maximum similarity between the target user and other clusters, expressed as:

[0141]

[0142] From the above definition, we can see that the purpose of the clustering process is to increase the value of a and reduce the value of b. On this basis, the silhouette coefficient of the cluster is obtained by the following formula:

[0143]

[0144] When the silhouette coefficient is close to 1, it means that the user similarity in a cluster is higher, while the similarity between clusters is very low. On the contrary, when the silhouette coefficient is close to -1, it means that the similarity between clusters is much greater than the similarity within a cluster. The goal of clustering is to make the silhouette coefficient as close to 1 as possible. Specifically, the goal of clustering is to maximize the following objectives:

[0145]

[0146] Therefore, the clustering problem can be expressed as:

[0147]

[0148] In order to achieve efficient user clustering, the present invention improves the traditional k-center algorithm and proposes a fast and efficient clustering method. The specific implementation process is shown in Table 1.

[0149] Table 1 User clustering algorithm

[0150]

[0151] Here, it should be noted that the selection of the initialized cluster heads in the algorithm is not completely random. In order to ensure the difference, the two users with the greatest similarity are first selected as cluster heads. As the number of clusters increases, a target user is selected, and its similarity with the selected cluster head is calculated to obtain the sum of the product values. All target users are traversed, and the user with the smallest product is selected as the cluster head.

[0152] Step 4: Image semantic compression process: The base station semantically encodes the image by using a flexible semantic model to characterize the image fidelity, considering semantic encoders of different sizes, outputting semantic information of different data volumes, and using different semantic encoders and decoders according to the different requirements of each cluster for image fidelity and latency.

[0153] The present invention uses a multi-layer convolutional neural network to extract semantic information. First, a full-size semantic extraction and reconstruction model is constructed, and the parameterized model weight coefficients are represented as ξ, ψ. Next, for the sub-semantic model derived from it, a given proportional compression factor γ is used. k After that, we get its parameterized weight coefficient, expressed as The sub-semantic model is obtained on the basis of the full-size semantic model. That is, when constructing the convolutional neural network framework, the number of convolution kernels in each layer of the full-size model is determined, which is called the layer width. The sub-semantic model does not change the number of layers of the full-size model, but only changes the width of each layer to achieve a small-size model. Therefore, the objective function of learning is defined as:

[0154]

[0155] in represents the semantic reconstruction function under given parameter conditions, |Λ| represents the number of elements in the set Λ, Represents the loss function, and the cross entropy function is used here as the loss function. As can be seen from the above formula, each round of iteration of the training process must sum all samples and enumerate all sub-models. The amount of training is extremely large and cannot be used in practice. The present invention proposes a dynamic sub-model training algorithm. In this algorithm, the total training loss is calculated as the sum of the losses of the randomly selected sub-model and the full-size model. In this way, the training overhead can be reduced while maintaining the performance of different sub-models. However, due to the consideration of a single case, this training method can only guarantee the improvement of the sub-model performance of the selected proportional compression factor in each round of iteration, and cannot improve the performance of the entire model. It will also fall into a local optimum, and even the loss value will fluctuate, and it cannot converge. Inspired by this, the present invention proposes a small-batch dynamic sub-model training method, that is, all proportional compression factors are divided into batches with the same batch size, a batch is randomly selected for training, and its loss and the full-size model are calculated. The gradient is back-propagated to start a new round of iteration. The specific steps are given in Table 2.

[0156] Table 2 Small batch dynamic sub-model training method

[0157]

[0158]

[0159] The present invention uses fidelity as one of the user experience indicators. Considering the cluster k, the proportional compression factor γ is discussed in detail below. k Impact on image reconstruction fidelity. The semantic fidelity of a semantic model is defined as the ability to reconstruct data on a test dataset. Formally, given a proportional compression factor, the semantic fidelity of the corresponding sub-model is calculated as:

[0160]

[0161] Where L represents the number of pixels in the image, X test To measure the number of samples in the test data set, ||·|| represents the L1 norm. After training, the relationship between semantic fidelity and scale factor can be fitted as follows:

[0162]

[0163] Among them, (σ1, σ2, σ3, σ4) are hyperparameters that can be experimentally fitted. k With the increase of, the fidelity of semantic data increases to carry more detailed information.

[0164] Step 5: Semantic communication process: According to the service experience quality requirements of the target users in each cluster, a game model is constructed to solve the problem of maximizing the average experience quality of the target users in the cell. The base station selects an appropriate image semantic compression ratio and transmits the corresponding semantic information to the cluster head. The cluster head reuses the links of cellular users and selects an appropriate link to broadcast the information to the users in the cluster, completing the base station image transmission task.

[0165] Consider a single cell in a certain area. There are N e cellular users, represented by Note that cellular users are assigned dedicated link resources, so assuming that the number of available channel links is consistent with the number of cellular users, the total number of available channel links is denoted as M. At the same time, there are N u target users, represented by Receive content recommended by the base station, etc. Assume there are K clusters, and represent the set of cluster heads as Ω = {h1,h2,...,h K}, the set of clusters is represented as Hypothesis Cluster In addition to the cluster head, there is at least one target user in the cluster. The target user within can be specifically represented as Among them, the last user in the cluster is defined as the cluster head, that is, And it exists uniquely. The transmission between the cluster head and the users in the cluster is carried out in D2D transmission mode. The transmission link reuses the cellular link resources, and multiple cluster heads are allowed to reuse the channel resources of the same cellular link. However, while this process improves the spectrum utilization efficiency, it also brings co-channel interference. The cellular network has completed the clustering of user interests, as described in step 3. The cluster head of a cluster selects an appropriate semantic compression ratio to semantically compress the recommended image content according to the situation of the target users in the cluster. The specific process is described in detail later. The extracted semantic information is converted into a form suitable for propagation in the wireless space after channel coding. The next step is the transmission process of the wireless channel. In order to save spectrum resources, the broadcast in the cluster reuses the spectrum resources of cellular users, but at the same time, the performance requirements of cellular users must be guaranteed. Define s={s1,s2,...,s K} is the link selection strategy set of all cluster heads, and the interference tolerance temperature is used to characterize the constraint condition. Assume that the cellular user e m The maximum tolerable interference value is ε m , we can get:

[0166]

[0167] in represents the transmission power of the cluster head, is the cluster head h k The channel gain between the base station and the k ,l) is the indicator function, l is the specified channel, which can be defined as:

[0168]

[0169] At the same time, the throughput of a cluster depends on the minimum achievable rate of the receivers in the cluster, that is, it depends on the maximum interference value received by the receivers in the cluster. Therefore, the interference value received by the cluster is defined as

[0170]

[0171] Therefore, the cluster The achievable rate of the receivers in the cluster is expressed as:

[0172]

[0173] For the target users in the cluster, when receiving content recommendation services, they are often more concerned about the reception delay and image quality. First, the cluster head and the user have completed the training of semantic extraction and recovery in advance, and the encoding and decoding time of semantics in the transmission process is short and can be ignored. Therefore, the user delay is mainly concentrated in the transmission delay. The delay of the target user n in cluster k can be expressed as:

[0174]

[0175] Where K' represents the number of bits of unit semantic information, S k The amount of semantic data of the transmitted image in cluster k under the full-size model. The logical function can well describe the change of user experience quality. When the score of a user's demand continues to increase, the improvement of the user's experience quality brought by this demand will be relatively reduced. The present invention adopts the method of logical function to characterize the user experience quality index, which is specifically expressed as:

[0176]

[0177] Among them, μ k,n ∈[0,1] represents the weight of the fidelity of target user n in cluster k, 1-μ k,n is the weight of user latency. λ1,λ2,λ3 are parameters that determine the range of changes in the logic function, and also ensure that the scores of fidelity and latency are between [0,1]. For users with different needs, their different content recommendation needs can be met. μ k,n =0 means that the target user n in cluster k is more concerned about receiving fast recommended content and only needs to meet the content requirements; μ k,n =1 means that the user is more concerned about the image quality of the recommended content, rather than the timeliness of the recommended content. Therefore, the present invention forms the following optimization problem based on the needs of different users in the community:

[0178] P0:

[0179] stC1:

[0180] C2:

[0181] C3:

[0182] C4:

[0183] The objective function of the above problem (P0) is to maximize the average quality of experience of all target users in the entire cell. Constraint C1 ensures the range of values ​​of the semantic compression ratio, and the minimum value must satisfy that the image reconstructed by the extracted semantic information can express the meaning of the original recommended image; C2 ensures that the semantic compression ratio determined by the cluster head must ensure that the fidelity requirements of all target users in the cluster are met; C3 means that the delay of each target user must meet its minimum delay threshold; C4 limits the interference caused by clustered multiplexing of cellular user channels to not exceed the maximum tolerance value of cellular users. This problem is a typical non-convex problem, and the objective function is highly coupled with the constraints and cannot be solved directly. If the exhaustive search method is used, the computational complexity will increase exponentially with the increase in the number of users and the number of subdivisions of the semantic compression ratio. In order to solve the problem more quickly and efficiently, the present invention introduces a game theory framework to analyze the problem and solves the above problem through a distributed low-complexity method.

[0184] First, consider the selection of semantic compression ratios for each cluster. Given the selection of cluster channels, the delay of the target user in each cluster is also fixed, so maximizing the quality of experience of all users in the cell is equivalent to maximizing the quality of experience of users in each cluster. Therefore, the original problem P0 is transformed into:

[0185] P1:

[0186]

[0187] From the training method of semantic encoding and decoding, we can know that the selection space of the semantic compression ratio is a finite set, where the optimal solution can be calculated by enumeration method.

[0188] The next step is to optimize the channel selection of each cluster under a given semantic compression ratio, so the original problem P0 is transformed into the following problem:

[0189] P2:

[0190]

[0191] It can be seen from problem P2 that the channel selection of each cluster will affect the user experience quality of all clusters in the entire cell, so problem P2 cannot be solved directly. The present invention introduces the method of potential game to solve this problem. The optimization problem is modeled as a cluster head link selection game Where Ω is the set of cluster heads, S = s1×...s K is the strategy space of cluster heads, U k is the utility function of the cluster head. In order to design a reasonable utility function, the benefit of clustering k is first defined as:

[0192]

[0193] On this basis, the utility function of cluster k is designed as:

[0194]

[0195] Among them, s -k Indicates the link selection strategy of other clusters except cluster k. k′ (s k′ ,s -k′\k ) represents the benefit of cluster k′ when cluster k does not adopt any strategy. Therefore, V k′ (s k′ ,s -k′ )-V k′ (s k′ ,s -k\k ) represents the change in the benefit of cluster k′ before and after cluster k takes action. Specifically, the designed utility function has two specific meanings. The first term represents the benefit of the cluster itself, and the second term represents the impact of the action on the benefits of other clusters in the cell cellular network. According to the definition of Nash equilibrium, if and only if no cluster head can unilaterally change its link selection strategy to improve its own utility function, that is:

[0196]

[0197] at this time, is considered to be the Nash equilibrium of the proposed cluster head link selection game. According to the definition of the exact potential energy game, if there exists a potential energy function Φ(s k ,s -k ), so that the following formula:

[0198]

[0199] If it holds, then the proposed cluster head link selection game is an exact potential energy game.

[0200] On the other hand, the potential energy function Φ(s k ,s -k ), expressed as:

[0201]

[0202] The designed potential energy function is the objective function of problem P2, which means the sum of the benefits of all target users in the cell. Now consider the impact of the change of clustering k strategy on its own utility function:

[0203]

[0204] Although the strategy of cluster k has changed, when cluster k does not take action, the strategies of other clusters remain unchanged. At this time, the impact of cluster k on other clusters before and after taking action is the same, which is specifically expressed as

[0205]

[0206] Substitution can be simplified to:

[0207]

[0208] It can be seen that the change in the utility function brought about by the change in the clustering k-link strategy is the same as the change in the potential energy function in the system. It can be determined that the proposed game model is an exact potential energy game. According to its intrinsic characteristics, the game has at least one pure strategy Nash equilibrium.

[0209] Through the above analysis, although the existence of Nash equilibrium has been proved, how to reach Nash equilibrium quickly and effectively is still a thorny problem. Taking into account both computational complexity and optimality of the solution, the present invention proposes a cluster head link selection algorithm based on parameter correction, as shown in Table 3. Common algorithms that can be used to achieve Nash equilibrium include optimal response algorithm, better response algorithm and random learning algorithm. Compared with the above traditional algorithms, the proposed algorithm has the following advantages:

[0210] 1. The algorithm can adaptively change its learning parameters according to the number of iterations. When the number of iterations is large, the learning step size increases, which can converge faster; when the number of iterations is small, the step size decreases, which can fully explore unknown areas.

[0211] 2. The decision variables are expanded from the traditional one-dimensional to two-dimensional. The common optimal response and better response algorithms are both one-dimensional link selection or power optimization, while the proposed algorithm combines the semantic compression ratio within the cluster with the cluster head link selection, and treats the two independent variables as a combined variable.

[0212] Table 3 Adaptive cluster head link selection optimization algorithm based on parameter modification

[0213]

[0214] Next, we will explain that the proposed cluster head link selection game model can converge to the Nash equilibrium. It is an exact potential energy game, that is, the change in the utility function of the cluster head is equal to the variable of the potential energy function. During the algorithm cycle, each strategy update of the cluster head ensures that its utility function increases while its potential energy function value continues to increase. At the same time, the strategy space of the cluster head is limited, and the increase of the utility function is also limited, so it can converge to the local optimal solution or the global optimal solution in a limited number of cycles.

[0215] For the solution of step 1 in Table 3, the semantic compression model is completed by pre-training. Therefore, based on the given link selection, the optimal semantic compression ratio is selected by enumeration. The complexity of enumeration is the number of semantic compression ratio factors, that is, Where L1 is a constant. The link selection at this time is the optimal solution under a given compression scale factor. Through link selection, the average user experience quality of the entire cell is increased. The complexity is the value of the utility function calculated by the selected cluster head, which is expressed as Where I1 is the number of iterations of the cluster head link selection game algorithm, and L2 is a constant. From the above analysis, it can be seen that in each round of iteration, the solution of semantic compression ratio and link selection ensures the non-decreasing property of the objective function, and the strategy space is finite, thus ensuring the convergence of the iterative process and the boundedness of the objective function. The alternating algorithm has a lower complexity, which is It can be solved in polynomial time, where I2 is the number of iterations.

[0216] In one embodiment, a semantic communication-oriented D2D multicast communication system is provided, the system comprising:

[0217] The first module is used to implement: the base station broadcasts semantic information of the image to be transmitted, wherein the semantic information is a triple structure;

[0218] The second module is used to define a search list: the target user receives the semantic information of the image sent by the base station, selects the interest preference for the image according to the semantic information, and sets the interest list, i.e., the search list;

[0219] The third module is used to implement: user clustering based on the similarity between target users;

[0220] The fourth module is used to achieve: building image semantic compression models with different compression ratios;

[0221] The fifth module is used for semantic communication: according to the service experience quality requirements of the target users in each cluster, a game model is constructed to solve the problem of maximizing the average experience quality of the target users within the area; the base station selects the image semantic compression ratio, inputs it into the image semantic compression model, and transmits the output corresponding semantic information to the cluster head. The cluster head reuses the transmission link of the cellular user and selects the transmission link to broadcast the information to the users in the cluster, completing the base station image transmission task.

[0222] For the specific limitations of the D2D multicast communication system for semantic communication, please refer to the limitations of the D2D multicast communication method for semantic communication mentioned above, which will not be repeated here. Each module in the above-mentioned D2D multicast communication system for semantic communication can be implemented in whole or in part through software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0223] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the following is achieved:

[0224] Step 1, the base station broadcasts semantic information of the image to be transmitted, wherein the semantic information is a triple structure;

[0225] Step 2, defining a search list: the target user receives the semantic information of the image sent by the base station, selects an interest preference for the image according to the semantic information, and sets an interest list, i.e., a search list;

[0226] Step 3: cluster users based on the similarity between target users;

[0227] Step 4, constructing image semantic compression models with different compression ratios;

[0228] Step 5, perform semantic communication: construct a game model based on the service experience quality requirements of the target users in each cluster to solve the problem of maximizing the average experience quality of the target users within the area; the base station selects the image semantic compression ratio, inputs it into the image semantic compression model, and transmits the output corresponding semantic information to the cluster head. The cluster head reuses the transmission link of the cellular user and selects the transmission link to broadcast the information to the users in the cluster, completing the base station image transmission task.

[0229] For the specific definition of each step, please refer to the definition of the D2D multicast communication method for semantic communication in the above text, which will not be repeated here.

[0230] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the computer program implements:

[0231] Step 1, the base station broadcasts semantic information of the image to be transmitted, wherein the semantic information is a triple structure;

[0232] Step 2, defining a search list: the target user receives the semantic information of the image sent by the base station, selects an interest preference for the image according to the semantic information, and sets an interest list, i.e., a search list;

[0233] Step 3: cluster users based on the similarity between target users;

[0234] Step 4, constructing image semantic compression models with different compression ratios;

[0235] Step 5, perform semantic communication: construct a game model based on the service experience quality requirements of the target users in each cluster to solve the problem of maximizing the average experience quality of the target users within the area; the base station selects the image semantic compression ratio, inputs it into the image semantic compression model, and transmits the output corresponding semantic information to the cluster head. The cluster head reuses the transmission link of the cellular user and selects the transmission link to broadcast the information to the users in the cluster, completing the base station image transmission task.

[0236] For the specific definition of each step, please refer to the definition of the D2D multicast communication method for semantic communication in the above text, which will not be repeated here.

[0237] Figure 2 This is a D2D multicast system model diagram adapted by the present invention. The base station transmits semantic information to the cluster head in a directional manner, and the cluster head then broadcasts the received semantic information to users in the cluster.

[0238] Figure 3 The performance of user clustering is shown. The cluster head random selection algorithm indicates that the initial cluster head selection is completely random. It can be seen from the figure that the algorithm proposed in the present invention has the best clustering effect when the threshold value is 2, indicating that the clustering effect can be fully improved by adjusting the similarity threshold; on the other hand, compared with the k-means algorithm, the algorithm proposed in the present invention can achieve better results than the k-means algorithm regardless of whether the cluster head is randomly selected or selected according to the proposed rules.

[0239] Figure 4 The relationship between the loss value and the training cycle of the proposed small batch training method is given. The "FAST" training method randomly extracts the sum of the proportional compression factor of a sub-model and the full-size model as the loss function. It can be seen from the figure that the model can realize fast sub-model training, and the loss value shows a downward trend as the training cycle increases. However, in each training cycle, the gradient value of the objective function changes greatly, the gradient estimation is inaccurate, and the direction of the previous descent cannot be inherited in the next iteration, so it cannot converge to the optimal value of the objective function, and the model will produce large fluctuations in different training cycles. The small batch sub-model training method proposed in the present invention randomly selects multiple proportional compression factors from all batches as the loss function, and randomly selects a batch of sub-models to train together with the full-size model each time, estimates the gradient of the objective function to the greatest extent, and avoids the problem of large amount of computational complexity of the gradient calculation of the complex objective function. On the other hand, the loss value of the proposed algorithm is significantly lower than that of the "FAST" algorithm, which proves the superiority of the proposed algorithm.

[0240] Figure 5 The changes in user delay scores and fidelity scores with the change of delay weight values ​​are given. It can be seen that the weight of delay in the similarity index is increased from 0.2 to 1. The delay experience score and the content service score show different trends. Specifically, as the weight of physical location in the similarity index increases, the delay score increases significantly. This is because the clustering process of target users focuses more on clustering users who are physically closer, and does not emphasize content similarity. This results in lower latency within the formed clusters, but relatively poor content experience in image recommendations. On the contrary, when the weight of delay is reduced, users receive content that is more in line with their interests, although at the cost of delay. Therefore, the experimental results verify that the user clustering method proposed in the present invention can effectively adapt to a variety of scenarios.

[0241] Figure 6 The figure shows the change of the average user experience of the target user as the number of clusters increases. In this scenario, considering the number of cellular users to be 3, it can be observed that the target user's quality of experience gradually decreases as the number of clusters increases. When the number of clusters increases further, the channel occupancy will also increase, resulting in a significant increase in interference to the same channel. As a result, the target user experiences a significant decrease in latency, resulting in an overall decrease in the user's quality of experience. On the other hand, the algorithm proposed in the present invention can be closer to the optimal solution. Especially in complex scenarios with a large number of clusters, the algorithm shows its superiority in terms of optimality.

[0242] In this method, the semantic codec is trained through convolutional neural networks, deep learning, etc., and game theory is introduced as the research framework. The utility function is calculated in a distributed manner through continuous iteration until convergence. This optimization method can be widely used in the fields of D2D communication and semantic communication.

[0243] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A D2D multicast communication method for semantic communication, characterized in that: The method comprises the following steps: Step 1, the base station broadcasts semantic information of the image to be transmitted, wherein the semantic information is a triple structure; Step 2, defining a search list: the target user receives the semantic information of the image sent by the base station, selects an interest preference for the image according to the semantic information, and sets an interest list, i.e., a search list; Step 3: cluster users based on the similarity between target users; Step 4: construct image semantic compression models with different compression ratios; specifically, it includes: Step 4-1, construct the objective function of the image semantic codec with different compression ratios: Construct a full-scale semantic extraction and reconstruction model, and the parameterized model weight coefficient is expressed as (ξ,ψ); For the sub-semantic model derived from it, under a given ratio compression factor γ k After that, the parameterized model weight coefficient is expressed as The objective function is defined as: in, represents the semantic reconstruction function under given parameter conditions, |Λ| represents the number of elements in the set Λ, represents the loss function; Indicates that the parameter set is The semantic encoding of Indicates that the parameter set is Semantic decoding of Step 4-2, based on the objective function, using a small batch dynamic sub-model training method to train image semantic codecs with different compression ratios to generate image semantic compression models with different compression ratios; The specific process includes: Step 4-2-1, input training set X sample ,(ξ,ψ), set the maximum number of cycles; Step 4-2-2, divide the proportional compression factor into N equal batches, and randomly select |Λ n | proportional compression factors and divided into n groups, where n = 1, 2, ..., N, and randomly select one of the batches Λ n ; Step 4-2-3, for each batch of training data x batch ∈X sample ; Loop, iterating the following process: Step 4-2-4, given the proportional compression factor γ, calculate the decoding result of the size width under γ Step 4-2-5, use the full-size model to calculate the decoding result under the full-size model Step 4-2-6, calculate the loss function: Step 4-2-7, back propagation, update (ξ,ψ); Step 4-2-8, output the weight coefficients (ξ, ψ) of the full-size semantic extraction and reconstruction model, and generate image semantic compression models with different compression ratios; Step 5, semantic communication: construct a game model according to the service experience quality requirements of the target users in each cluster, and solve the problem of maximizing the average experience quality of the target users within the area; the base station selects the image semantic compression ratio, inputs it into the image semantic compression model, and transmits the output corresponding semantic information to the cluster head. The cluster head reuses the transmission link of the cellular user, selects the transmission link to broadcast the information to the users in the cluster, and completes the base station image transmission task; specifically includes: Step 5-1, construct the following optimization problem P0 according to different user needs: in, In the formula, is the experience quality indicator of target user n in cluster k. C1, C2, C3, and C4 are all constraints. is the delay of target user n in cluster k, yes The threshold of k,n ∈[0,1] represents the weight of the fidelity of target user n in cluster k, 1-μ k,n is the weight of the target user n’s delay; λ1,λ2,λ3 are the parameters that determine the range of the logic function change, ensuring that the scores of fidelity and delay are between [0,1]; μ k,n =0 means that the target user n in cluster k is more concerned about receiving fast recommended content and only needs to meet the content requirements; μ k,n =1 means that the user is more concerned about the image quality of the recommended content, rather than the timeliness of the recommended content; N u Indicates that there are N u target users; min Represents γ k The lower limit of the value range; ρ(s k ,l) represents the indicator function, l is the designated channel, sk represents the cluster head h k The link selection strategy set, represents the cluster head h k The transmission power, is the cluster head h k The channel gain between the base station and the m For cellular users m The maximum tolerable interference value; K' represents the number of bits of unit semantic information; S k is the amount of semantic data of the transmitted image in cluster k under the full-size model; Cluster The achievable rate of the receivers in the cluster; is the compression factor γ of cluster k at a given ratio k The semantic fidelity of the corresponding sub-model; (σ1, σ2, σ3, σ4) are hyperparameters; For target users The semantic fidelity threshold of Step 5-2, using the cluster head link selection algorithm based on parameter correction to solve the optimization problem P0, the specific process includes: Step 5-2-1, generating a cluster transmission link selection strategy and an image semantic compression ratio that meet the above C1, C2, C3, and C4 constraints as initial variables; The following process is executed in a loop: Step 5-2-2, randomly select cluster k, and use one-dimensional enumeration method to update the optimal image semantic compression ratio that meets the constraints; Step 5-2-3, update the transmission link selection strategy of cluster k, calculate the utility function of all optional transmission link selection strategies of the cluster head corresponding to cluster k, and then follow the hybrid strategy π k (i+1) selection strategy, updated using the Boltzmann exploration strategy criterion: Among them, U k (l,s -k ) represents the utility of the cluster head corresponding to cluster k when selecting the transmission link l', σ(i+1)≥0 is the learning parameter, which is positively correlated with the number of iterations. is the probability that the cluster head corresponding to cluster k selects the transmission link l', s -k represents the transmission link selection strategy of clusters other than cluster k, s k represents the transmission link selection strategy corresponding to cluster k, U k (s k ,s -k ) represents the utility function of cluster k, represents the set of transmission link selection strategies corresponding to cluster k; The utility function calculation formula is: in, Where V k (s k ,s -k ) is the benefit of cluster k, V k′ (s k′ ,s -k ) is the benefit of cluster k' that does not take the transmission link selection strategy action, represents the clusters in the cluster set C except cluster k, s k′ represents the transmission link selection strategy of cluster k', s -k′ represents the transmission link selection strategy of other clusters except cluster k', V k′ (s k′ ,s -k′\k ) is the benefit of cluster k' after cluster k adopts the transmission link selection strategy; Step 5-2-4, update the transmission link selection strategy s k ; The loop ends until all target users no longer change their transmission link selection strategies and semantic compression ratios.

2. The D2D multicast communication method for semantic communication according to claim 1, characterized in that: Step 1 specifically includes: Semantically encode the image to be transmitted; Based on the semantic segmentation technology of relational transformer and deep neural network, the triple structure of the image to be transmitted is obtained, which is expressed as follows: t n =D(x n ) Among them, τ n represents the set of triple structures, represents the semantic content of image n, m n represents the total number of triplets in image n, represents the first A triplet, x n Denotes the input image n, and D denotes a function of the semantic segmentation technique based on the relational transformer and deep neural network.

3. The D2D multicast communication method for semantic communication according to claim 1, characterized in that: In step 2, the interest list of target user i for a set of images is expressed as: q i =[q i,1 ,...,q i,2 ,q i,N ] T Where N represents the total number of images pushed by the base station to the target user i, q i,n represents the interest level of target user i in image n, q i,n ∈{0,1}, q i,n =1 means that target user i is interested in image n, q i,n =0 means that target user i has no interest in image n.

4. The D2D multicast communication method for semantic communication according to claim 1, characterized in that: Step 3 specifically includes: Step 3-1, calculating the similarity between target users based on their interest lists and physical locations; Step 3-2: Based on the similarity, target users are clustered using an improved k-center algorithm.

5. The D2D multicast communication method for semantic communication according to claim 3 or 4, characterized in that: Step 3-1 specifically includes: Assume that the physical location of target user i is represented as (x i ,y i ), the normalized position coordinates are R is the radius of the target user’s area; similarly, the normalized interest vector is Then the similarity between any two target users i and j is expressed as: Among them, F sim {u i ,u j ) represents the similarity between target users i and j, μ1 and μ2 represent the weights of interest similarity and physical location similarity, respectively. Respectively represent the normalized position coordinates of target users i and j, They represent the normalized interest vectors corresponding to target users i and j respectively; α is the large-scale fading coefficient.

6. The D2D multicast communication method for semantic communication according to claim 5, characterized in that: Step 3-2 specifically includes: (1) Assume that the target users are divided into K clusters, and the set of cluster heads is represented as Ω = {h1,h2,...,h K }, h K represents the cluster head of the Kth cluster, and the cluster is represented as represents the Kth cluster, and the users in the cluster are represented as represents the first target users, represents the total number of target users in cluster k; In order to ensure the similarity of users in the cluster, the following constraints are given: in, is the preset similarity threshold, k=1,2,...,K; The silhouette coefficient is introduced to measure the clustering effect: Among them, the intra-cluster similarity is defined as: The inter-cluster similarity is defined as the maximum similarity between the target user and other clusters, expressed as: In the formula, SC k,i represents the silhouette coefficient corresponding to the target user i in cluster k, a k,i represents the similarity of target user i in cluster k, b k,i represents the maximum similarity between the target user i in cluster k and other clusters, represents the similarity between target users i and j in cluster k, represents the similarity between target user i in cluster k and target user j in cluster k′, represents the k′th cluster, k′=1,2,...,K; represents the total number of target users in cluster k′; The purpose of clustering is to maximize the following objectives: Then the clustering problem in step 3-2 is expressed as: (2) Based on the definition in (1) above, the target users are clustered using the improved k-center algorithm. The specific process includes: Step 3-2-1, enter the similarity values ​​of all target users and set the maximum number of cycles; Step 3-2-2, set the number of clusters K, where N u Indicates that there are N u target users; Step 3-2-3, for each given number of clusters K, loop, set the number of iterations t = 0, and iterate the following process; Step 3-2-4, for each iteration t, initialize the cluster head to represents the cluster head of the Kth cluster in the tth iteration; for each cluster k, Cycle the following process; Step 3-2-5, Settings Step 3-2-6, for each cluster k∈[1,K], select the target user u ω , where u ω is the non-cluster head user in the cluster; Step 3-2-7, if but Among them, F sim (u ω ,h k ) represents user u ω With cluster head h k similarity; Step 3-2-8, if The target user u ω Enter cluster k * , Otherwise, user u ω are considered isolated points; Step 3-2-9, traverse all users u ω , calculate the corresponding silhouette coefficient as SC0; randomly select a target user in cluster k as the cluster head, and loop through steps 3-2-4 to 3-2-8 to calculate a new silhouette coefficient SC1; Then calculate: ΔSC=SC1-SC0 If ΔSC is a positive number, the cluster head of cluster k is changed to the randomly selected user, otherwise, it remains unchanged; The iteration ends until the maximum number of cycles is reached; Step 3-2-10, compare the silhouette coefficients under all clustering numbers, where the maximum value is the optimal allocation number K; Step 3-2-11, output the optimal number of clusters and the clustering and cluster head results.

7. A D2D multicast communication system for semantic communication based on the method according to any one of claims 1 to 6, characterized in that: The system comprises: The first module is used to implement: the base station broadcasts semantic information of the image to be transmitted, wherein the semantic information is a triple structure; The second module is used to define a search list: the target user receives the semantic information of the image sent by the base station, selects the interest preference for the image according to the semantic information, and sets the interest list, i.e., the search list; The third module is used to implement: user clustering based on the similarity between target users; The fourth module is used to achieve: building image semantic compression models with different compression ratios; The fifth module is used for semantic communication: according to the service experience quality requirements of the target users in each cluster, a game model is constructed to solve the problem of maximizing the average experience quality of the target users within the area; the base station selects the image semantic compression ratio, inputs it into the image semantic compression model, and transmits the output corresponding semantic information to the cluster head. The cluster head reuses the transmission link of the cellular user and selects the transmission link to broadcast the information to the users in the cluster, completing the base station image transmission task.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images