Semantic communication system and joint optimization method of user association and resource allocation in the system

Abstract

The application provides a semantic communication system and a user association and resource allocation joint optimization method in the system. The semantic communication system comprises a small base station, which communicates with a user through a traditional communication mode or a semantic communication mode. The small base station is used for buffering a knowledge base and / or hot content in a network. The traditional communication mode is a bit-based information transmission mode. When the user communicates with the small base station through the semantic communication mode, the user and the small base station adopt an attention mechanism-based semantic communication model to communicate, and the knowledge base and semantic information are jointly transmitted. The application can significantly improve the communication utility, the network and utility, and the number of served users.

Description

Technical Field

[0001] This invention relates to the field of semantic communication technology, and in particular to a semantic communication system and a method for joint optimization of user association and resource allocation in the system. Background Technology

[0002] Semantic communication is a task-oriented communication method that follows a "understand first, transmit later" approach. It first selectively extracts, compresses, and transmits features from the original signal, and then uses semantic information for communication. Semantic communication focuses on the transmission, understanding, and retrieval of information at the semantic level. The sender and receiver exchange the semantic meaning of the information; therefore, semantic communication can significantly improve communication performance and is expected to become a key breakthrough technology in the future.

[0003] Because semantic communication transmits semantic information from the source during communication, its resource consumption is difficult to measure accurately compared to traditional bit-based information transmission methods. Balancing utility and resource consumption is a crucial issue. This is especially true in multi-user semantic communication systems, where different users have different needs, leading to varying resource requirements. This can result in some users being allocated excessive resources, while others are unable to meet their needs or even access the service, leading to decreased network utility and wasted resources.

[0004] For multi-user semantic communication network scenarios, the following challenges are faced: how to improve user utility in response to differentiated user needs; and how to improve network efficiency and utility by optimizing resource allocation for multi-user access networks. Summary of the Invention

[0005] This invention provides a semantic communication system and a joint optimization method for user association and resource allocation in the system, in order to solve the defects of low network utility and resource waste in the prior art, and to maximize network utility and the number of users served.

[0006] This invention provides a semantic communication system, comprising: a small base station, wherein the small base station communicates with the user through a traditional communication method or a semantic communication method, and the small base station is used to cache knowledge bases and / or hot content in the network;

[0007] Traditional communication methods are bit-based information transmission methods;

[0008] When a user communicates with the small base station via semantic communication, the user and the small base station communicate using a semantic communication model based on an attention mechanism, and the knowledge base and semantic information are transmitted together.

[0009] According to a semantic communication system provided by the present invention, the semantic communication model based on the attention mechanism includes a sender, a physical channel, and a receiver;

[0010] The transmitting end includes two or more feature extraction modules, an attention module, a semantic coding mask module, and a power normalization module. The output side of the attention module of the transmitting end is equipped with one of the feature extraction modules. The feature extraction module on the input side of the attention module is used to extract semantic features from the input image. The attention module is used to concatenate the input semantic features and channel state feedback information to obtain enhanced semantic features. The feature extraction module on the output side of the attention module is used to extract the semantic information to be transmitted from the enhanced semantic features and scale the dimension of the extracted semantic information to be transmitted to the dimension required by the physical channel. The semantic coding mask module is used to select the semantic information to be transmitted from the semantic information output by the feature extraction module on the output side of the attention module, and dynamically adjust the compression rate of the semantic information to be transmitted according to a preset compression rate. The power normalization module is used to normalize the power of the semantic information output by the semantic coding mask module, and the power-normalized semantic information is transmitted through the physical channel.

[0011] The receiving end includes two or more feature recovery modules, an attention module, and a zero-padding module; wherein, the attention module set in the receiving end has the same structure and performance as the attention module set in the sending end; a feature recovery module is set on the input side of the attention module of the receiving end; the zero-padding module is used to scale the semantic information transmitted by the physical channel to the dimension required by the receiving end, and the feature recovery module on the input side of the attention module extracts and recovers image features from the semantic information output by the zero-padding module; the attention module is used to concatenate the input image features and channel state feedback information to obtain enhanced image features; the feature recovery module on the output side of the attention module is used to recover the image sent by the sending end from the enhanced image features.

[0012] Furthermore, the attention-based semantic communication model is obtained through end-to-end joint training on the general image dataset Mini-ImageNet, and its training process is as follows:

[0013] Train an attention-based semantic communication model on a basic, general-purpose image dataset;

[0014] After the attention-based semantic communication model converges, the parameters of the feature extraction module, semantic coding mask module, power normalization module, zero-padding module, and feature recovery module in the attention-based semantic communication model are frozen. The attention module is then trained on datasets with different image content to obtain attention modules corresponding to different image content datasets.

[0015] This invention also provides a joint optimization method for user association and resource allocation in a semantic communication system, which includes the following steps:

[0016] Construct a multi-objective hierarchical optimization problem; where the multi-objectives include maximizing the number of users accessing the semantic communication system and maximizing network efficiency and utility.

[0017] The multi-objective hierarchical optimization problem is transformed into two sub-problems: a resource allocation strategy optimization problem and a user association strategy optimization problem.

[0018] The block coordinate descent algorithm is used to iteratively solve the two subproblems, and finally the solution to the multi-objective hierarchical optimization problem is obtained. The solution to the multi-objective hierarchical optimization problem is the user association strategy and the resource allocation strategy. The user association strategy includes the user communication method.

[0019] In the joint optimization method for user association and resource allocation in a semantic communication system provided by this invention, the multi-objective hierarchical optimization problem is expressed as:

[0020]

[0021] In the formula, C represents the cache optimization variable, X represents the user association variable, α represents the user communication method variable, and p n Describe the objective function f n The priority, where n represents the objective function f n The number of layers; the priority p of the objective function in the nth layer. n The priority p of the objective function at level n+1 n+1 The relationship between p is: n >p n+1 .

[0022] Furthermore, the resource allocation strategy optimization problem is expressed as:

[0023]

[0024] In the formula, Q represents the network and utility; the resource allocation strategy C includes file caching variables and knowledge base caching variables; files are cached according to a preset ratio, and the knowledge base is either fully cached or not cached; U represents the user set, u represents the user identifier, and t u γ represents the user communication delay, and t represents the time delay. uThe weighting coefficients.

[0025] Furthermore, the constraints of the resource allocation strategy optimization problem are as follows:

[0026]

[0027]

[0028]

[0029]

[0030]

[0031] In the formula, This indicates the backhaul link latency, which is related to the small cell buffer and the choice of user communication method. This represents the computational latency for a semantic communication system to extract semantic information, and it is related to the user's communication method. This represents the downlink transmission delay, which is mainly related to the communication method and the user's communication rate; where c kj This represents the proportion of small base station k cache file j, and its value is between 0 and 1; D j Indicates the size of file j, α ku α represents the communication method between small base station k and user u. ku =0 represents the traditional communication method, α ku =1 indicates semantic communication mode; c ks This indicates that the small base station k caches the knowledge base s, c ks The value is a discrete value in the range {0, 1}, c ks =0 indicates no caching, c ks =1 indicates cache; D s R0 represents the size of the knowledge base s; R0 represents the backhaul link communication rate between the macro base station and the core network; SE(D) j S j S represents the computational cost required for the sending end to extract semantic information from the source information. j f represents the amount of semantic information extracted. ku R represents the computing resources used for communication between small base station k and user u; ku C represents the communication rate between small base station k and user u; k This indicates the buffering capacity of the small base station k.

[0032] Furthermore, the optimization problem of the user association strategy after optimizing the resource allocation strategy is expressed as:

[0033]

[0034] In the formula, X represents the user-related variable, α represents the user communication method variable, U represents the user set, u represents the user identifier, Q represents the network and utility, and S represents the user association variable. u β represents the semantic communication performance obtained by user communication, and S represents the semantic communication performance. u The weighting coefficient, t u γ represents the user communication delay, and t represents the time delay. u The weighting coefficients.

[0035] Furthermore, the constraints for optimizing the user association strategy after optimizing the resource allocation strategy are as follows:

[0036]

[0037]

[0038]

[0039]

[0040]

[0041] In the formula, This indicates the backhaul link latency, which is related to the small cell buffer and the choice of user communication method. This represents the computational latency for a semantic communication system to extract semantic information, and it is related to the user's communication method. This represents the downlink transmission delay, which is mainly related to the communication method and the user's communication rate; where c kj This represents the proportion of small base station k cache file j, and its value is between 0 and 1; D j Indicates the size of file j, α ku α represents the communication method between small base station k and user u. ku =0 represents the traditional communication method, α ku =1 indicates semantic communication mode; c ks This indicates that the small base station k caches the knowledge base s, c ks The value is a discrete value in the range {0, 1}, c ks =0 indicates no caching, c ks =1 indicates cache; D s R0 represents the size of the knowledge base s; R0 represents the backhaul link communication rate between the macro base station and the core network; SE(D) ... j S j S represents the computational cost required for the sending end to extract semantic information from the source information. j f represents the amount of semantic information extracted. ku R represents the computing resources used for communication between small base station k and user u; kux represents the communication rate between small base station k and user u; ku ∈{0,1} indicates whether there is a correlation between small base station k and user u; S min N represents the minimum semantic communication performance requirement for each user; k This represents the maximum number of users that a small base station k can connect to.

[0042] Furthermore, the block coordinate descent algorithm is used to iteratively solve the two sub-problems, ultimately obtaining the solution to the multi-objective hierarchical optimization problem, specifically including:

[0043] Determine the distribution of small base stations and users in the semantic communication system, the power, bandwidth, and buffer resources of each small base station, the communication requests of each user, and the popularity of hot content in the semantic communication system; assume that users randomly associate with small base stations to access the network and randomly select communication methods;

[0044] Based on the current user access status, the popularity of trending content in the network, and user requests, a fixed user association strategy is established, and a resource allocation strategy is solved using the branch and bound method.

[0045] A fixed resource allocation strategy is adopted, and a user association strategy is solved through a swap matching algorithm.

[0046] The result of solving one subproblem is used as the input to another subproblem. The two subproblems are iterated alternately until the block coordinate descent algorithm converges, thus obtaining the optimal user association strategy and resource allocation strategy.

[0047] The present invention also provides a joint optimization device for user association and resource allocation in a semantic communication system, which includes an optimization problem construction module, a transformation module and a solution module;

[0048] The optimization problem construction module is used to construct a multi-objective hierarchical optimization problem; wherein, the multi-objective includes maximizing the number of users accessing the semantic communication system and maximizing network efficiency and utility.

[0049] The transformation module is used to transform the multi-objective hierarchical optimization problem into two sub-problems, namely, a resource allocation strategy optimization problem and a user association strategy optimization problem.

[0050] The solution module is used to iteratively solve two sub-problems using a block coordinate descent algorithm, and finally obtain the solution to the multi-objective hierarchical optimization problem.

[0051] This invention provides a semantic communication system that uses small base stations to enable users to communicate with them via traditional or semantic communication methods. The small base stations are used to cache knowledge bases and / or trending content from the network. When users communicate with small base stations via semantic communication, they use an attention-based semantic communication model. This attention-based semantic communication model guides the extraction and generation of semantic information. During transmission, the attention-based knowledge base and semantic information are transmitted together, which can significantly improve communication efficiency.

[0052] This invention provides a joint optimization method for user association and resource allocation in a semantic communication system. Based on the semantic communication system provided by this invention, it introduces base station caching resources. These caching resources can be knowledge bases and / or hot content. Caching the knowledge base improves the performance of semantic communication, while caching hot content saves bandwidth or power resources for backhaul transmission from the core network to edge base stations. The base station provides point-to-point semantic communication services to users. When accessing the network, users can select a suitable base station and communication method based on the base station's caching resources and the communication environment. This invention constructs a multi-objective hierarchical optimization problem, transforming it into a resource allocation strategy optimization problem and a user association strategy optimization problem. A block coordinate descent algorithm is used to iteratively solve the two sub-problems, ultimately obtaining the solution to the multi-objective hierarchical optimization problem. The resulting optimized user association and resource allocation strategies significantly improve network efficiency and the number of users served. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0054] Figure 1 This is a schematic diagram of the semantic communication network provided by the present invention;

[0055] Figure 2 This is a schematic diagram of the structure of the semantic communication model based on the attention mechanism in the semantic communication system provided by the present invention;

[0056] Figure 3 This is a schematic diagram of the RCB module in a semantic communication model based on the attention mechanism;

[0057] Figure 4 This is a schematic diagram of the AFB module in a semantic communication model based on the attention mechanism;

[0058] Figure 5 This is a schematic diagram of the RTCB module in a semantic communication model based on the attention mechanism;

[0059] Figure 6 This is a flowchart illustrating the joint optimization method for user association and resource allocation in the semantic communication system provided by the present invention.

[0060] Figure 7 This is a schematic diagram of the structure of the joint optimization device for user association and resource allocation in the semantic communication system provided by the present invention;

[0061] Figure 8 This is the performance curve of the semantic communication system provided by this invention under Rayleigh fading channel conditions;

[0062] Figure 9 This is the convergence curve of the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention;

[0063] Figure 10 This is a simulation diagram showing the performance of the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention as a function of the number of users.

[0064] Figure 11 This is a curve showing the performance of the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention as a function of base station transmit power.

[0065] Figure label:

[0066] 101: Small base station; 102: Macro base station; 103: Core network;

[0067] 701: Optimization problem construction module; 702: Transformation module; 703: Solution module. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0069] In semantic communication systems, the sender and receiver extract semantic information from the source information before transmission, which greatly reduces redundancy, improves coding efficiency, and saves bandwidth resources. When encoding and decoding the source information, both parties rely on a shared knowledge base. Therefore, the extraction and recovery of semantic information depends on the sender and receiver's understanding of the knowledge base. Existing semantic communication systems employ neural networks such as CNNs and Transformers as semantic encoding and decoding structures, leveraging the powerful ability of neural networks to extract deep internal features of the source as semantic information.

[0070] Some existing studies, from the perspective of semantic importance, have focused on regions of higher importance in medical images, using pre-segmentation models as prior information to extract features, thus improving the performance of transmission and reconstruction. From the perspective of the adaptability of semantic communication systems, semantic communication systems that can adapt to changes in signal-to-noise ratio have been proposed. For the transmission of face images, statistically obtained average face values ​​are used as prior information to extract facial features from the image, significantly reducing the amount of data that needs to be transmitted. However, the above studies are all aimed at semantic communication in specific scenarios and tasks, and their applicability is limited.

[0071] The following is combined with Figures 1-6 This invention describes the semantic communication system and the joint optimization method for user association and resource allocation within the system.

[0072] like Figure 1 As shown, this invention provides a semantic communication network, which includes an SBS (Small Base Station 101), an MBS (Macro Base Station 102), and a core network 103. The UE (User) can choose either traditional communication or semantic communication to communicate with the SBS, and then access the core network via a wired connection between the SBS and MBS. The SBS provides access services to the user and has a certain caching capability. Considering the differentiated requests of the UE in the semantic communication network, the SBS can choose to cache knowledge base and / or hot content in the network. The MBS determines which SBS caches knowledge base and / or hot content in the network through base station caching strategies. Cached knowledge base can improve the performance of semantic communication, and cached hot content in the network can save backhaul link latency from MBS to SBS, reducing the overall communication latency for the user. The SBS accesses the core network 103 through the MBS. The MBS communicates directly with servers in the core network 103, providing services such as wide-area coverage and control.

[0073] A UE can form a semantic communication system with an SBS in a semantic communication network. In this system, a UE can choose to access the semantic communication network through one SBS. One SBS can provide communication services to multiple UEs. Due to the limited cache and bandwidth resources of SBSs, the semantic communication network needs to simultaneously select different SBSs to provide access services to the UE based on the UE's communication environment, request content, and SBS cache availability, and choose appropriate communication methods to improve network efficiency and the number of users that can be served. Here, network efficiency refers to the sum of the utilities of all users in the network.

[0074] When the UE selects the semantic communication method to communicate with the SBS, the UE and the SBS use an attention-based semantic communication model to communicate.

[0075] The attention-based semantic communication model was trained on the general image dataset Mini-ImageNet.

[0076] like Figure 2 As shown, the semantic communication model based on the attention mechanism includes a transmitter, a physical channel, and a receiver. The transmitter includes one or more feature extraction modules, an AFB (Attention Fuse Block) module, a semantic coding mask module, and a power normalization module. The AFB module is positioned before the last feature extraction module. That is, a feature extraction module is located at the output of the AFB module, and one or more feature extraction modules are located at the input of the AFB module.

[0077] The AFB module's input-side feature extraction module extracts semantic features from the input image. The AFB module concatenates the input semantic features with channel state feedback information to obtain enhanced semantic features. The AFB module's output-side feature extraction module extracts the semantic information to be transmitted from the enhanced semantic features and scales the dimensions of the extracted semantic information to the dimensions required by the physical channel. The semantic coding mask module selects the semantic information to be transmitted from the semantic information output by the AFB module's output-side feature extraction module and dynamically adjusts the compression rate of the semantic information to be transmitted according to a preset compression rate. The power normalization module normalizes the power of the semantic information output by the semantic coding mask module; the power-normalized semantic information is then transmitted through the physical channel.

[0078] In the attention-based semantic communication model, the receiver and transmitter are symmetrically configured. The receiver includes two or more feature recovery modules, an AFB (Automatic Feature Extraction) module, and a zero-padding module. The AFB module in the receiver has the same structure and performance as the AFB module in the transmitter. The AFB module is placed after the first feature recovery module. That is, there is one feature recovery module on the input side of the AFB module, and one or more feature extraction modules on the output side of the AFB module.

[0079] The zero-padding module scales the semantic information transmitted through the physical channel to the dimension required by the receiver. The feature recovery module on the input side of the AFB module recovers image features from the semantic information output by the zero-padding module. The AFB module concatenates the input image features and channel state feedback information to obtain enhanced image features. The feature recovery module on the output side of the AFB module recovers the image transmitted by the transmitter from the enhanced image features.

[0080] In one specific embodiment, the feature extraction module uses the RCB (Residual Convolutional Block) module.

[0081] like Figure 3 As shown, the RCB module includes convolutional layers, a generalized divisive normalization (GDN) layer, and an activation function, which is used to extract semantic features from the input image. Specifically, the input to the RCB module is an H*W*K dimensional feature vector. After convolution, normalization, activation, and other operations, a higher-dimensional feature vector (i.e., semantic information) H′*W′*K′ is generated and fed into the subsequent RCB module for further compression and extraction of semantic information.

[0082] like Figure 4 As shown, the AFB module comprises a first fully connected layer (FC), a ReLU activation function layer, a second fully connected layer, and a Sigmoid activation function layer connected in sequence, which is used to improve the performance of semantic communication. The input of the AFB module is the features output by the previous RCB module and the channel state feedback information (mainly the signal-to-noise ratio SNR). The fully connected layer in the AFB module compresses the input features into a one-dimensional vector and concatenates it with the SNR. After passing through the ReLU and Sigmoid activation functions, a set of scaling factors with the same dimension as the input features is generated, and multiplied with the input features before being sent as the output to the subsequent RCB module, thereby scaling the image feature dimension to improve the performance of semantic communication.

[0083] The AFB module is lightweight and, once trained, can be modularly integrated into existing networks for plug-and-play functionality.

[0084] In one specific embodiment, the feature recovery module adopts the RTCB (Residual Transpose Convolutional Block) module.

[0085] like Figure 5 As shown, the RTCB module includes deconvolution layers, inverse normalization layers, and activation functions, which are used to recover the original image from the received semantic information. The input of the RTCB module is a high-dimensional feature vector, which is output as a low-dimensional feature vector after deconvolution, inverse normalization, and other operations, thereby recovering the original image step by step from the high-dimensional features (semantic information).

[0086] The semantic communication system provided by this invention can be mainly used in image transmission and recovery communication scenarios. By learning the feature distribution of images with different content through the AFB module, the semantic communication performance and model applicability are improved.

[0087] In one possible embodiment, the training method for the attention-based semantic communication model is as follows:

[0088] Training an attention-based semantic communication model on a basic general image dataset enables the RCB module to learn how to better extract image features, i.e., semantic communication.

[0089] After the attention-based semantic communication model converges, the parameters of the RCB module, semantic coding mask module, power normalization module, zero-padding module, and RTCB module in the attention-based semantic communication model are frozen. The attention modules are then trained on datasets with different image content to learn the feature distribution of images in different image content, thus obtaining attention modules corresponding to different image content datasets.

[0090] When transmitting corresponding content, the semantic communication model based on the attention mechanism can choose to use the corresponding pre-trained attention module, thereby improving the performance of semantic communication.

[0091] In this embodiment, the general image dataset is Mini-ImageNet, which contains images in 100 categories and is a subset of the large ImageNet dataset. Thirty image categories are selected from the Mini-ImageNet dataset as the general dataset (mainly including images with significant differences in structure or scene, such as birds, boats, cars, road signs, and fish), with 500 images per category. For each image category in the general dataset, more images are selected from the ImageNet dataset to construct specialized datasets with different image content; that is, each image category constitutes a separate dataset, with 300 images per category, and these datasets do not overlap with the general dataset.

[0092] This attention-based semantic communication model is trained using an end-to-end joint training approach (i.e., the sender and receiver train together) and is built on the PyTorch framework. The loss function is calculated using the original image Y from the sender and the reconstructed image from the receiver. The minimum mean square error (MSE) between them It is obtained by calculating the difference at each pixel point between the restored image and the original image.

[0093] The attention-based semantic communication model trained on a general dataset is frozen, and the parameters of other modules except the AFB module are frozen. The model is then trained on different specialized datasets to enable the AFB module to learn the feature distribution of different images. In other words, there are several specialized datasets, and finally several different AFB modules are trained (such as AFB modules for bird image feature distribution, AFB modules for ship image feature distribution, and AFB modules for road sign image feature distribution).

[0094] When transmitting the corresponding image, select the corresponding AFB module for replacement.

[0095] In semantic communication, a knowledge base typically refers to shared model parameters between the sender and receiver, used to extract and generate semantic information. In this invention, a pre-trained attention module is used as the knowledge base. By training on datasets with different image content, it learns the feature distributions of different image content, thereby improving the effectiveness of semantic communication. Considering the information asymmetry that may arise from knowledge base updates, a method of jointly transmitting both the knowledge base and semantic information is adopted during communication.

[0096] like Figure 6 As shown, based on the semantic communication system architecture proposed in this invention, the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention includes the following steps:

[0097] S1. Construct a multi-objective hierarchical optimization problem;

[0098] The multi-objective hierarchical optimization problem is an optimization problem that maximizes network efficiency and utility while ensuring that as many users as possible access the semantic communication system. In other words, the multi-objective nature of the problem manifests as maximizing the number of users accessing the semantic communication system, as well as maximizing network efficiency and utility.

[0099] Specifically, by jointly optimizing user association and resource allocation in a multi-user downlink semantic communication system, the network and utility are maximized, thus constructing a multi-objective hierarchical optimization problem.

[0100] Specifically, the multi-objective hierarchical optimization problem can be expressed as:

[0101]

[0102] In equation (1), C represents the cache optimization variable, X represents the user association variable, α represents the user communication method variable, and p n Describe the objective function f n The priority, where n represents the objective function f n The number of layers. The priority p of the objective function in the nth layer. n The priority p of the objective function at level n+1 n+1 The relationship between p is: n >p n+1 .

[0103] S2. Transform the multi-objective hierarchical optimization problem into two sub-problems;

[0104] The two sub-problems are the resource allocation strategy optimization problem (i.e., the optimization problem of SBS cache resources) and the optimization problem of user association strategy after the resource allocation strategy is optimized.

[0105] S3. The block coordinate descent algorithm is used to iteratively solve the two sub-problems, and finally the solution to the multi-objective hierarchical optimization problem is obtained.

[0106] Among them, the solution to the multi-objective hierarchical optimization problem is the user association strategy, resource allocation strategy, and user communication method.

[0107] By decoupling the multi-objective hierarchical optimization problem, it is transformed into solving two subproblems. After obtaining the optimal solution of one subproblem using the block coordinate descent algorithm, the optimized variables are substituted into the other subproblem to obtain the optimal solution of the other subproblem. By iteratively solving the two problems until convergence, the optimal solution of the multi-objective hierarchical optimization problem is obtained, that is, the optimal user association strategy and resource allocation strategy are obtained.

[0108] In one possible embodiment, the objective function f n It includes the first-level optimization function and the second-level optimization function.

[0109] The first-level optimization function is:

[0110]

[0111] In equation (2), f1(C, X, α) represents the number of users that meet the access conditions; This represents the number of users who meet the QoE (Quality of Experience) criteria, and it is an indicator function. If Q... u ≥Q min ,but The value is 1, otherwise, The value is 0; U represents the user set, u represents the user identifier, and Q... uQ represents the utility of user u. u =βS u -γt u S u t represents the semantic communication performance obtained by user communication. u β represents the user communication delay, and S represents the latency. u The weighting coefficient, γ represents t u Weighting coefficients; Q min This represents the minimum utility required by each user.

[0112] The second-level optimization function is:

[0113]

[0114] In equation (3), f2(C, X, α) represents the sum of utility for all users.

[0115] The constraints for the first-level optimization function and the second-level optimization function are:

[0116]

[0117]

[0118]

[0119]

[0120]

[0121]

[0122]

[0123]

[0124] In equations (4) to (11), equation (4) represents the communication delay for each user, which consists of three parts. This indicates the return link latency, which is related to the SBS buffer and the choice of user communication method. This represents the computational latency for a semantic communication system to extract semantic information, and it is related to the user's communication method. This represents downlink transmission latency, which is mainly related to the communication method and the user's communication rate. Each user has an upper limit on the required communication latency; where c... kj This represents the proportion of hotspot content (i.e., file j) cached by small base station k, and its value ranges from 0 to 1; D j Indicates the size of file j, α ku α represents the communication method between small base station k and user u. ku=0 represents the traditional communication method, α ku =1 indicates semantic communication mode; c ks This indicates that the small base station k caches the knowledge base s, c ks The value is a discrete value in the range {0, 1}, c ks =0 indicates no caching, c ks =1 indicates cache; D s R0 represents the size of the knowledge base s; R0 represents the backhaul link communication rate between macro base station 102 and core network 103. SE(D) j S j S represents the computational cost required for the sending end to extract semantic information from the source information (i.e., file j). j f represents the amount of semantic information extracted. ku R represents the computing resources used for communication between small base station k and user u. uu This represents the communication rate between small base station k and user u. (C) k This represents the buffering capacity of small base station k. ku ∈{0,1} indicates whether there is a correlation between small base station k and user u. min This represents the minimum semantic communication performance requirement for each user. N k This indicates the maximum number of users that a small base station k can connect to.

[0125] Since the variables in the multi-objective hierarchical optimization problem include both integers and non-integers, the problem is difficult to solve. Therefore, it is necessary to transform the multi-objective hierarchical optimization problem into two sub-problems for solving.

[0126] In one possible embodiment, the resource allocation strategy optimization problem (i.e., the optimization problem of SBS cache resources) is expressed as:

[0127]

[0128] In equation (12), Q represents the network and utility. The resource allocation strategy C includes file caching variables and knowledge base caching variables. Hot content (i.e., files) is cached according to a preset ratio, while the knowledge base needs to be fully cached or not cached due to reasons such as irregular updates. The problem represented by equation (12) is a mixed integer programming problem, which is solved using the branch and bound method.

[0129] The constraints of equation (12) are equations (4) to (8).

[0130] In one possible embodiment, the optimization problem of user association strategy after optimizing SBS cache resources is expressed as:

[0131]

[0132] Since the user association variable X is a discrete value of 0 and 1, the problem represented by equation (13) is non-convex and cannot be solved by using methods for solving convex optimization. This invention designs a heuristic exchange matching algorithm based on Gay-Shapley matching theory to solve the user association strategy, and selects the optimal communication method for each user based on a greedy strategy.

[0133] The constraints of equation (13) are equation (4) and equations (8) to (11).

[0134] In one possible implementation, a block coordinate descent algorithm is used to iteratively solve the two subproblems, ultimately obtaining the solution to the multi-objective hierarchical optimization problem, specifically including:

[0135] S31. Initialization:

[0136] Determine the distribution of SBSs and UEs in the semantic communication system, the communication resources of each SBS such as power, bandwidth and cache resources, the communication requests of each UE, and the popularity of hot content in the network; assume that UEs randomly associate with SBSs to access the network and randomly select communication methods.

[0137] S32. Optimize resource allocation strategy:

[0138] Based on the current user access status, the popularity of trending content on the network, and user requests, a fixed user association strategy X is established. The resource allocation strategy C is then solved using the branch and bound method. The process is as follows:

[0139] For k = 1: K, determine the caching strategy C for each small base station k. k ={C kj C ks};

[0140] S321. Configure the knowledge base cache variable C. ks The relaxation is a continuous variable of 0 to 1. The CVX convex optimization toolbox is used to solve equation (12) to obtain the optimal solution, which is denoted as C0.

[0141] S322. Update the global upper and lower bound information, with the upper bound being ∞ and the lower bound being Q(C0), and put C0 into the branch exploration queue Queue.

[0142] S323, Determine the C obtained by solving equation (12). ks If all integers are 0 or 1, then the optimal solution of equation (12) has been obtained, and the solution process is exited; otherwise, proceed to step S324 to perform branch solution.

[0143] S324. When performing branching solutions, determine the knowledge base cache variable C. ks The upper bound is 1 and the lower bound is 0, for C ks ={Cks0 C ks1 C ksn In the algorithm, one variable is fixed at 0 or 1, while other variables are relaxed. The current optimal solution C is obtained by solving the problem using the convex optimization toolbox. Temp If C Temp The solution satisfies the constraints and Q(C) Temp If Q(C0) > Q(C0), then update the global upper bound, perform optimal pruning, and return to step S323; otherwise, proceed to step S325.

[0144] S325. Update the upper and lower bound information of the current node, i.e., current lower bound = Q(C Temp The current upper bound equals the global upper bound. For Q(C) Temp Compare Q(C0) and Q(C0). If Q(C0) = Q(C0) Temp If Q(C0) > Q(C0), then cut the branch (boundary pruning); otherwise, continue to perform branching operations on the node.

[0145] S326, Regarding C Temp C Tempks For variables that do not meet the integer constraints, continue to determine them as 0 or 1, continue branch exploration, and put them into the exploration queue (Queue) to become new nodes for branch exploration.

[0146] S327. Traverse the solutions in the exploration queue (Queue), determine if a better caching strategy can be obtained, and make the judgment in step S323. Continuously branch and explore, update the optimal solution and global upper and lower bounds, and finally obtain the optimal cache resource allocation strategy C. kout ={C kjout C ksout}

[0147] S33. Optimize user association strategy:

[0148] Fixed small base station 101 caching strategy C k ={C kj C ks The user association strategy X is solved by the exchange matching algorithm.

[0149] For t = 1: T, where T represents the number of attempts of the swap matching algorithm. When the algorithm exceeds the limit number of attempts or the utility no longer increases, the algorithm is considered to have converged and obtained the optimal value. t represents the number of rounds of swap selection iterations.

[0150] Fixed cache resource allocation strategy C k ={C kj C ks The user association strategy X is solved by the exchange matching algorithm.

[0151] The process of optimizing user association strategies is as follows:

[0152] S331. Calculate the network and utility Q before the exchange, and from the user set U = {U1, U2, ..., U...} u Two users, U1 and U2, are randomly selected from the list.

[0153] S332. Exchange the small base stations 101 associated with two users U1 and U2, and calculate the network and utility Q' after the exchange.

[0154] S333. Determine whether the network and utility Q' after the exchange are greater than the network and utility Q before the exchange. If so, retain the exchange and update the user association policy, Q = Q', X = X'; otherwise, keep the user association policy unchanged.

[0155] S334. For u = 1: U, select the optimal communication method α for each user u based on a greedy strategy. u .

[0156] S34. Repeat steps S32 and S33, using the optimal resource allocation strategy obtained in each step as the initial value for solving the user association strategy in the next step, until the network and utility no longer increase, indicating that the algorithm has converged and obtained the optimal user association strategy and resource allocation strategy.

[0157] In step S32 above, the caching strategy C of the small base station k k This includes file cache variable C. k j and knowledge base cache variable C ks .

[0158] Among them, the file cache variable C kj This can be represented as: {C kj0 C kj1 C kjm}, m represents the number of files, C kj element C kji The buffer ratio of small base station k for file i is a continuous variable between 0 and 1.

[0159] Knowledge base cache variable C ks This can be represented as: {C ks0 C ks C ksn}, where n represents the number of knowledge bases, and C ks element C ksl This represents the cache of knowledge base l by small base station k, and is an integer variable of 0 or 1.

[0160] In step S32 above, the constraint condition for the solution is:

[0161] The obtained Ck ={C kj C ks In}, C kj Each element in C is a continuous variable between 0 and 1. ks Each element in the variable is an integer variable that is either 0 or 1.

[0162] Based on the joint optimization method for user association and resource allocation in a semantic communication system provided by the present invention, the present invention also provides a joint optimization device for user association and resource allocation in a semantic communication system.

[0163] The following describes the joint optimization device for user association and resource allocation in a semantic communication system provided by the present invention. The joint optimization device for user association and resource allocation in a semantic communication system described below can be referred to in correspondence with the joint optimization method for user association and resource allocation in a semantic communication system described above.

[0164] like Figure 7 As shown, the joint optimization device for user association and resource allocation in the semantic communication system provided by the present invention includes an optimization problem construction module 701, a transformation module 702, and a solution module 703.

[0165] Among them, the optimization problem construction module 701 is used to construct multi-objective hierarchical optimization problems.

[0166] The transformation module 702 is used to transform the multi-objective hierarchical optimization problem into two sub-problems, namely, the resource allocation strategy optimization problem and the user association strategy optimization problem.

[0167] The solver module 703 is used to solve the two subproblems alternately and iteratively using the block coordinate descent algorithm, and finally obtain the solution to the multi-objective hierarchical optimization problem.

[0168] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the joint optimization method for user association and resource allocation in the semantic communication system provided by the above methods. The method includes: constructing a multi-objective hierarchical optimization problem; transforming the multi-objective hierarchical optimization problem into two sub-problems; using a block coordinate descent algorithm to alternately and iteratively solve the two sub-problems, and finally obtaining the solution to the multi-objective hierarchical optimization problem.

[0169] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the joint optimization method for user association and resource allocation in the semantic communication system provided by the methods described above. The method includes: constructing a multi-objective hierarchical optimization problem; transforming the multi-objective hierarchical optimization problem into two sub-problems; using a block coordinate descent algorithm to iteratively solve the two sub-problems alternately, and finally obtaining the solution to the multi-objective hierarchical optimization problem.

[0170] Figure 8 The figure shows the performance curve of the semantic communication system provided by the present invention under Rayleigh fading channel conditions.

[0171] Figure 9 The convergence curve of the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention is shown. The multi-objective hierarchical optimization problem is decoupled into two sub-problems and solved iteratively using the block coordinate descent algorithm. The optimal solution can be obtained in a finite number of iterations, proving the feasibility of the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention.

[0172] Figure 10 The simulation graph shows the performance of the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention as a function of the number of users. With a fixed number of base stations and maximum base station transmit power, the graph compares the performance curves of different user communication methods as the number of users changes. It can be seen that compared to a system where users can only choose between semantic communication and traditional communication methods, the algorithm proposed in this invention, which allows users to choose between semantic communication and traditional communication methods, shows a significant performance improvement. Figure 10 As can be seen, network performance and utility are lowest when users only use traditional communication methods. The user association and resource allocation joint optimization method in the semantic communication system provided by this invention has significant advantages over the semantic communication method, with a performance improvement of approximately 61%. Furthermore, the user association and resource allocation joint optimization method in the semantic communication system provided by this invention also improves network performance and utility compared to using only semantic communication methods. This is because some users have better communication environments in the network and can directly use traditional communication methods, saving computational latency and improving network performance and utility, resulting in a performance improvement of approximately 6%. This demonstrates the superiority of the user association and resource allocation joint optimization method in the semantic communication system provided by this invention.

[0173] Figure 11The simulation graph shows the performance of the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention as a function of the maximum base station transmit power. The curves showing the changes in users and utility with varying base station transmit power are presented. From top to bottom, the three curves represent semantic communication and traditional communication coexisting, semantic communication only, and traditional communication only. It can be seen that as the base station transmit power increases, the transmission rate per user increases, latency decreases, and network efficiency and utility increase. Furthermore, the algorithm proposed in this invention exhibits optimal performance when both communication methods coexist.

[0174] in, Figures 9-11 The JUCA (Joint User Association and Cache Resource Optimization Algorithms) algorithm described in this paper is the joint optimization method for user association and resource allocation in the semantic communication system provided by this invention.

[0175] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A joint optimization method for user association and resource allocation in a semantic communication system, characterized in that, Includes the following steps: Construct a multi-objective hierarchical optimization problem; where the multi-objectives include maximizing the number of users accessing the semantic communication system and maximizing network efficiency and utility. The multi-objective hierarchical optimization problem is transformed into two sub-problems: a resource allocation strategy optimization problem and a user association strategy optimization problem. The block coordinate descent algorithm is used to iteratively solve the two subproblems, and finally the solution to the multi-objective hierarchical optimization problem is obtained. The solution to the multi-objective hierarchical optimization problem is the user association strategy and the resource allocation strategy. The user association strategy includes the user communication method. The multi-objective hierarchical optimization problem is expressed as: ， In the formula, This represents a cache optimization variable. Represents user-related variables. This represents a variable indicating the user's communication method. Describe the objective function priority, n Describe the objective function The number of layers; the first n Priority of layer objective function With the n Priority of the +1 level objective function The relationship between them is: ; The resource allocation strategy optimization problem is expressed as: ， In the formula, Q Representing network and utility; resource allocation strategies C This includes file cache variables and knowledge base cache variables; files are cached according to a preset ratio, while the knowledge base is either fully cached or not cached at all. U Represents a set of users. u Indicates user identifier, Indicates user communication latency. express Weighting coefficients; The user association strategy optimization problem is expressed as: ， In the formula, This indicates the semantic communication performance obtained by the user. express Weighting coefficients; The constraints for optimizing the user association strategy after optimizing the resource allocation strategy are as follows: ， ， ， ， In the formula, This indicates the backhaul link latency, which is related to the small cell buffer and the choice of user communication method. This represents the computational latency for a semantic communication system to extract semantic information, and it is related to the user's communication method. This represents downlink transmission delay, which is mainly related to the communication method and the user's communication rate; among which, Indicates small base station k cache files j The proportion, whose value is between 0 and 1; Represents a file j Size, Indicates small base station k and users u The communication methods between them =0 indicates the traditional communication method. =1 indicates semantic communication mode; Indicates small base station k Caching knowledge base s , The value is a discrete value in the range {0,1}. =0 means no caching. =1 indicates caching; Representation of knowledge base s Size; This indicates the backhaul link communication rate between the macro base station and the core network; This represents the computational cost required for the sending end to extract semantic information from the source information. S j This indicates the amount of semantic information extracted. Indicates small base station k and users u Computing resources for communication between them; Indicates small base station k and users u Communication rate between them; Indicates small base station k and users u Are they related? This represents the minimum semantic communication performance requirements for each user; Indicates small base station k The maximum number of users that can be connected.

2. The joint optimization method for user association and resource allocation in a semantic communication system according to claim 1, characterized in that, The constraints of the resource allocation strategy optimization problem are: ， ， , ， ， In the formula, Indicates small base station k caching capabilities.

3. The joint optimization method for user association and resource allocation in a semantic communication system according to claim 1, characterized in that, The block coordinate descent algorithm is used to iteratively solve the two sub-problems, ultimately yielding the solution to the multi-objective hierarchical optimization problem, specifically including: Determine the distribution of small base stations and users in the semantic communication system, the power, bandwidth, and buffer resources of each small base station, the communication requests of each user, and the popularity of hot content in the semantic communication system; assume that users randomly associate with small base stations to access the network and randomly select communication methods; Based on the current user access status, the popularity of trending content in the network, and user requests, a fixed user association strategy is established, and a resource allocation strategy is solved using the branch and bound method. A fixed resource allocation strategy is adopted, and a user association strategy is solved through a swap matching algorithm. The result of solving one subproblem is used as the input to another subproblem. The two subproblems are iterated alternately until the block coordinate descent algorithm converges, thus obtaining the optimal user association strategy and resource allocation strategy.

4. A semantic communication system, employing the joint optimization method for user association and resource allocation in the semantic communication system according to any one of claims 1 to 3, characterized in that, include: Small base stations communicate with users through traditional communication methods or semantic communication methods, and are used to cache knowledge bases and / or trending content in the network; Traditional communication methods are bit-based information transmission methods; When a user communicates with the small base station via semantic communication, the user and the small base station communicate using a semantic communication model based on an attention mechanism, and the knowledge base and semantic information are transmitted together.

5. The semantic communication system according to claim 4, characterized in that, The attention-based semantic communication model includes a sender, a physical channel, and a receiver. The transmitting end includes two or more feature extraction modules, an attention module, a semantic coding mask module, and a power normalization module. The output side of the attention module of the transmitting end is equipped with one of the feature extraction modules. The feature extraction module on the input side of the attention module is used to extract semantic features from the input image. The attention module is used to concatenate the input semantic features and channel state feedback information to obtain enhanced semantic features. The feature extraction module on the output side of the attention module is used to extract the semantic information to be transmitted from the enhanced semantic features and scale the dimension of the extracted semantic information to be transmitted to the dimension required by the physical channel. The semantic coding mask module is used to select the semantic information to be transmitted from the semantic information output by the feature extraction module on the output side of the attention module, and dynamically adjust the compression rate of the semantic information to be transmitted according to a preset compression rate. The power normalization module is used to normalize the power of the semantic information output by the semantic coding mask module, and the power-normalized semantic information is transmitted through the physical channel. The receiving end includes two or more feature recovery modules, an attention module, and a zero-padding module; wherein, the attention module set in the receiving end has the same structure and performance as the attention module set in the sending end; a feature recovery module is set on the input side of the attention module of the receiving end; the zero-padding module is used to scale the semantic information transmitted by the physical channel to the dimension required by the receiving end, and the feature recovery module on the input side of the attention module recovers the image features from the semantic information output by the zero-padding module; the attention module is used to concatenate the input image features and channel state feedback information to obtain enhanced image features; the feature recovery module on the output side of the attention module is used to recover the image sent by the sending end from the enhanced image features.

6. The semantic communication system according to claim 5, characterized in that, The attention-based semantic communication model was obtained through end-to-end joint training on the general image dataset Mini-ImageNet. The training process is as follows: Train an attention-based semantic communication model on a basic, general-purpose image dataset; After the attention-based semantic communication model converges, the parameters of the feature extraction module, semantic coding mask module, power normalization module, zero-padding module, and feature recovery module in the attention-based semantic communication model are frozen. The attention module is then trained on datasets with different image content to obtain attention modules corresponding to different image content datasets.

7. A joint optimization apparatus for user association and resource allocation in a semantic communication system, employing the joint optimization method for user association and resource allocation in a semantic communication system as described in any one of claims 1 to 3, characterized in that, It includes an optimization problem construction module, a transformation module, and a solution module; The optimization problem construction module is used to construct a multi-objective hierarchical optimization problem; wherein, the multi-objective includes maximizing the number of users accessing the semantic communication system and maximizing network efficiency and utility. The transformation module is used to transform the multi-objective hierarchical optimization problem into two sub-problems, namely, a resource allocation strategy optimization problem and a user association strategy optimization problem. The solution module is used to iteratively solve two sub-problems using a block coordinate descent algorithm to finally obtain the solution to the multi-objective hierarchical optimization problem; wherein, the solution to the multi-objective hierarchical optimization problem is a user association strategy and a resource allocation strategy, and the user association strategy includes user communication methods; The multi-objective hierarchical optimization problem is expressed as: ， In the formula, This represents a cache optimization variable. Represents user-related variables. This represents a variable indicating the user's communication method. Describe the objective function priority, n Describe the objective function The number of layers; the first n Priority of layer objective function With the n Priority of the +1 level objective function The relationship between them is: ; The resource allocation strategy optimization problem is expressed as: ， In the formula, Q Representing network and utility; resource allocation strategies C This includes file cache variables and knowledge base cache variables; files are cached according to a preset ratio, while the knowledge base is either fully cached or not cached at all. U Represents a set of users. u Indicates user identifier, Indicates user communication latency. express Weighting coefficients; The user association strategy optimization problem is expressed as: ， In the formula, This indicates the semantic communication performance obtained by the user. express Weighting coefficients; The constraints for optimizing the user association strategy after optimizing the resource allocation strategy are as follows: ， ， ， ， In the formula, This indicates the backhaul link latency, which is related to the small cell buffer and the choice of user communication method. This represents the computational latency for a semantic communication system to extract semantic information, and it is related to the user's communication method. This represents downlink transmission delay, which is mainly related to the communication method and the user's communication rate; among which, Indicates small base station k cache files j The proportion, whose value is between 0 and 1; Represents a file j Size, Indicates small base station k and users u The communication methods between them =0 indicates the traditional communication method. =1 indicates semantic communication mode; Indicates small base station k Caching knowledge base s , The value is a discrete value in the range {0,1}. =0 means no caching. =1 indicates caching; Representation of knowledge base s Size; This indicates the backhaul link communication rate between the macro base station and the core network; This represents the computational cost required for the sending end to extract semantic information from the source information. S j This indicates the amount of semantic information extracted. Indicates small base station k and users u Computing resources for communication between them; Indicates small base station k and users u Communication rate between them; Indicates small base station k and users u Are they related? This represents the minimum semantic communication performance requirements for each user; Indicates small base station k The maximum number of users that can be connected.

Images