A distributed semantic communication system and bandwidth resource allocation method and device

Abstract

The present invention relates to the field of wireless communications technology, and more particularly to a distributed semantic communication system and a bandwidth resource allocation method and apparatus. The distributed semantic communication system includes D terminal devices and an edge server. The method comprises: deploying a fully connected layer of a deep residual network on the edge server, and deploying neural networks prior to the fully connected layer in the deep residual network on the D terminal devices; establishing a bandwidth resource allocation optimization model based on the distributed semantic communication system; constructing and training a deep deterministic policy gradient (DDPG) algorithm model, and determining an optimal resource allocation strategy network in an offline phase based on the bandwidth resource allocation optimization model; and, when in an online phase, inputting current state information into the optimal resource allocation strategy network to determine an optimal feature compression scheme and bandwidth resource allocation scheme. The present invention can alleviate the problem of scarce available bandwidth resources in current wireless communication networks.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a distributed semantic communication system and a bandwidth resource allocation method and device. Background Art

[0002] Available bandwidth resources in current wireless communication networks are scarce, making it difficult to provide efficient, low-latency data transmission services for the exponentially growing number of IoT devices and data volumes, posing a significant challenge to communication systems. Furthermore, in the current field of communication technology, which focuses on accurate bit recovery, the high redundancy of big data wastes a significant amount of communication resources. Compared to current communication technologies, semantic communication technology extracts useful information at the semantic level and removes redundant information, significantly reducing data volume and latency. This technology is more capable of meeting the needs of complex, diverse, and intelligent information transmission, and holds great potential to address these challenges.

[0003] Semantic communication focuses on useful and valuable information. However, traditional resource allocation methods based on QoS and QoE treat all transmitted content equally, disregarding its semantic information. Therefore, it is necessary to research more efficient resource allocation methods for semantic communication systems, allocating limited communication resources to semantically rich data as much as possible and maximizing the advantages of semantic communication. Reinforcement learning, with its powerful decision-making capabilities, has been widely used for resource management in wireless communications. Among reinforcement learning methods, the DDPG algorithm is suitable for high-dimensional, continuous state and action spaces and has a fast convergence rate. In recent years, it has achieved remarkable results in the field of wireless resource allocation.

[0004] In summary, there is an urgent need for a bandwidth allocation method based on a distributed semantic communication system to alleviate the problem of scarce available bandwidth resources in current wireless communication networks. Summary of the Invention

[0005] To alleviate the problem of scarce available bandwidth resources in current wireless communication networks, embodiments of the present invention provide a distributed semantic communication system and a method and apparatus for allocating bandwidth resources. The technical solution is as follows:

[0006] In one aspect, a method for allocating bandwidth resources based on a distributed semantic communication system is provided. The method is implemented by the distributed semantic communication system and includes:

[0007] Deploying the fully connected layer of the deep residual network on the edge server, and deploying the neural network before the fully connected layer in the deep residual network on the D terminal devices;

[0008] Based on the distributed semantic communication system, establishing a bandwidth resource allocation optimization model;

[0009] Build and train a deep deterministic policy gradient (DDPG) algorithm model, and determine the optimal resource allocation strategy network for the offline phase based on the bandwidth resource allocation optimization model.

[0010] When in the online stage, the current state information is input into the optimal resource allocation strategy network to determine the optimal feature compression scheme and bandwidth resource allocation scheme.

[0011] Optionally, deploying the fully connected layer of the deep residual network on the edge server, and deploying the neural network before the fully connected layer in the deep residual network on the D terminal devices includes:

[0012] Using the convolutional layers in a 50-layer deep residual network as a semantic encoder, and using the fully connected layers in the 50-layer deep residual network as a semantic decoder;

[0013] A distributed architecture is adopted to deploy the neural network before the fully connected layer on the D terminal devices, and the fully connected layer is deployed on the edge server.

[0014] Optionally, establishing a bandwidth resource allocation optimization model based on the distributed semantic communication system includes:

[0015] Activate reverse gradient propagation during neural network training, count the contributions of different features to the correct execution of the target AI task, build a regression model of feature compression rate and target AI task detection accuracy, and form a background knowledge base;

[0016] Based on the background knowledge base, a bandwidth resource allocation optimization model for a distributed semantic communication system is constructed with the goal of maximizing the average image detection accuracy and minimizing the average transmission delay.

[0017] Optionally, the detection accuracy is expressed using the following formula (1):

[0018]

[0019] Where j represents the jth execution task; i represents the i-th terminal device, and the value range of i is [1,n j ], where n j represents the number of devices executing task j, J represents the number of task categories, and D represents the number of terminal devices; is the parameter obtained by the logistic regression method, and the loss function is the mean square error; represents the feature compression rate of the i-th device corresponding to the j-th execution task;

[0020] Transmission delay is the total delay of transmitting information from the terminal device to the edge server, which is expressed by the following formula (2):

[0021]

[0022] in, Indicates terminal device Processing delay for feature extraction and compression;

[0023] In formula (2) Indicates terminal device transmission The time delay is expressed by the following formula (3):

[0024]

[0025] in, Represents each time slot interval t Δ Internal terminal equipment The amount of data generated, Indicates the actual amount of data to be transmitted after semantic compression;

[0026] terminal equipment Data transfer rate for:

[0027] It is a terminal device The allocated bandwidth, P represents the terminal device The transmission power, Indicates terminal device Channel gain to edge server, Indicates terminal device Noise power to the edge server.

[0028] Optionally, the constructing and training of a deep deterministic policy gradient (DDPG) algorithm model to determine an optimal resource allocation strategy network in an offline phase based on the bandwidth resource allocation optimization model includes:

[0029] Define the state space, action space, optimization problem, and reward function in the initial DDPG algorithm model;

[0030] In the offline stage, the feature compression scheme and bandwidth resource allocation scheme of the terminal device are taken as actions, and the system average detection accuracy, average delay, and the number of devices executing each target task are taken as states. By training the intelligent agent, the optimal resource allocation strategy network is obtained.

[0031] Optionally, in the offline phase, the feature compression scheme and bandwidth resource allocation scheme of the terminal device are used as actions, and the system average detection accuracy, average latency, and the number of devices performing each target task are used as states. An optimal resource allocation strategy network is obtained by training an intelligent agent, including:

[0032] Initialize the neural network parameters and clear the experience replay pool;

[0033] The action network is based on the current state s k and exploration noise N to select action a k ;

[0034] The agent performs action a k Get instant rewards k , generating the next state s k+1 , and the tuple (s k ,a k ,r k ,s k+1 ) is stored in the experience replay pool;

[0035] Repeat steps (2) and (3) until the experience replay pool overflows, and sample N samples from it;

[0036] Calculate the current target Q value based on the sample and update the parameters of the evaluation network;

[0037] Update the parameters of the action network according to the sampled policy gradient;

[0038] Each iteration, the parameters of the action target network and the evaluation target network are soft updated;

[0039] When the number of iterations reaches a preset threshold, offline training is completed and a policy network that can achieve optimal resource allocation is obtained.

[0040] On the other hand, a distributed semantic communication system is provided, which is applied to a bandwidth resource allocation method based on a distributed semantic communication system. The system includes D terminal devices and an edge server; wherein:

[0041] A neural network in a deep residual network is deployed on the terminal device to detect specific targets in the image;

[0042] The edge server is deployed with a fully connected layer of a deep residual network for intelligently processing and calculating images uploaded by the terminal device and returning the processing results to the terminal device.

[0043] On the other hand, a bandwidth resource allocation device based on a distributed semantic communication system is provided, wherein the device is used to implement a bandwidth resource allocation method based on a distributed semantic communication system, and the device includes:

[0044] A deployment module, configured to deploy the fully connected layer of the deep residual network on the edge server, and deploy the neural network before the fully connected layer in the deep residual network on the D terminal devices;

[0045] An establishment module is used to establish a bandwidth resource allocation optimization model based on the distributed semantic communication system;

[0046] A construction module is used to build and train a deep deterministic policy gradient (DDPG) algorithm model, and determine the optimal resource allocation strategy network in the offline phase based on the bandwidth resource allocation optimization model.

[0047] The determination module is used to input the current state information into the optimal resource allocation strategy network when in the online stage, and determine the optimal feature compression scheme and bandwidth resource allocation scheme.

[0048] Optionally, the establishing module is used to:

[0049] Activate reverse gradient propagation during neural network training, count the contributions of different features to the correct execution of the target AI task, build a regression model of feature compression rate and target AI task detection accuracy, and form a background knowledge base;

[0050] Based on the background knowledge base, a bandwidth resource allocation optimization model for a distributed semantic communication system is constructed with the goal of maximizing the average image detection accuracy and minimizing the average transmission delay.

[0051] Optionally, the building block is used to:

[0052] Define the state space, action space, optimization problem, and reward function in the initial DDPG algorithm model;

[0053] In the offline stage, the feature compression scheme and bandwidth resource allocation scheme of the terminal device are taken as actions, and the system average detection accuracy, average delay, and the number of devices executing each target task are taken as states. By training the intelligent agent, the optimal resource allocation strategy network is obtained.

[0054] On the other hand, an electronic device is provided, wherein the bandwidth resource allocation system based on the distributed semantic communication system includes a processor and a memory, wherein at least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned bandwidth resource allocation method based on the distributed semantic communication system.

[0055] On the other hand, a computer-readable storage medium is provided, wherein the storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned bandwidth resource allocation method based on the distributed semantic communication system.

[0056] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0057] By using the DDPG algorithm in reinforcement learning, the average detection accuracy and average transmission delay are jointly optimized according to the device's target AI tasks and the total amount of bandwidth resources to determine the optimal feature compression scheme and the optimal bandwidth allocation scheme. Bandwidth resources are allocated based on the optimal feature compression scheme and the optimal bandwidth allocation scheme, which can alleviate the problem of scarce available bandwidth resources in current wireless communication networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0059] Figure 1 This is a flow chart of a bandwidth resource allocation method based on a distributed semantic communication system provided by an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of a DDPG model provided by an embodiment of the present invention;

[0061] Figure 3 is a structural diagram of a distributed semantic communication system provided by an embodiment of the present invention;

[0062] Figure 4 This is a structural diagram of a bandwidth resource allocation device based on a distributed semantic communication system provided by an embodiment of the present invention;

[0063] Figure 5 It is a structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0065] The embodiment of the present invention provides a bandwidth resource allocation system based on a distributed semantic communication system. The method is implemented by the distributed semantic communication system, which includes D terminal devices and an edge server. Figure 1 The flowchart of a bandwidth resource allocation method based on a distributed semantic communication system is shown. The processing flow of the method may include the following steps:

[0066] S1. Deploy the fully connected layer of the deep residual network on the edge server, and deploy the neural network before the fully connected layer in the deep residual network on D terminal devices.

[0067] In one feasible implementation, the system uses a neural network to fit the semantic encoding and decoding process of image transmission. Using a distributed architecture, the neural network is deployed on both the terminal device and the edge server. Specifically, the input to the semantic encoder is the original image, of size (B, C, H, W), where B is the batch size of the image; C is the number of channels; and H and W are the height and width of the image, respectively. The output size of the semantic encoder is (B, C1, H1, W1), where C1 is the number of channels in the feature map, and H1 and W1 are the height and width of the output feature map. The input size of the semantic decoder is (B, C1), and the output size is (B, 1).

[0068] Alternatively, the convolutional layers of a 50-layer deep residual network can be used as semantic encoders, and the fully connected layers of a 50-layer deep residual network can be used as semantic decoders. This system uses a neural network to fit the encoding and decoding process of image transmission. Using a distributed architecture, the neural network is deployed on both the terminal device and the edge server, alleviating device storage pressure.

[0069] S2. Based on the distributed semantic communication system, a bandwidth resource allocation optimization model is established.

[0070] In a feasible implementation, S2 may include the following steps S21-S22:

[0071] S21. Activate reverse gradient propagation during neural network training, count the contributions of different features to the correct execution of the target AI task, build a regression model of feature compression rate and target AI task detection accuracy, and form a background knowledge base;

[0072] S22. Based on the background knowledge base, a bandwidth resource allocation optimization model for the distributed semantic communication system is constructed with the goal of maximizing the average image detection accuracy and minimizing the average transmission delay.

[0073] Optionally, the detection accuracy is expressed using the following formula (1):

[0074]

[0075] Where j represents the jth execution task; i represents the i-th terminal device, and the value range of i is [1,n j ], where n j represents the number of devices executing task j, J represents the number of task categories, and D represents the number of terminal devices; is the parameter obtained by the logistic regression method, and the loss function is the mean square error; represents the feature compression rate of the i-th device corresponding to the j-th execution task;

[0076] Transmission delay is the total delay of transmitting information from the terminal device to the edge server, which is expressed by the following formula (2):

[0077]

[0078] in, Indicates terminal device Processing delay for feature extraction and compression;

[0079] In formula (2) Indicates terminal device transmission The time delay is expressed by the following formula (3):

[0080]

[0081] in, Represents each time slot interval t Δ Internal terminal equipment The amount of data generated, Indicates the actual amount of data to be transmitted after semantic compression.

[0082] In a feasible implementation, suppose there are J task categories in the system model and the number of devices that execute task j is n. j ,So Therefore, for the i-th device executing task j In terms of detection accuracy, and feature compression ratio The relationship can be expressed as:

[0083]

[0084] Among them, the value range of i is [1,n j ], It is the parameter solved by the logistic regression method, and the loss function is the mean square error.

[0085] terminal equipment The data transmission rate is:

[0086]

[0087] in, It is a device The allocated bandwidth, P represents the terminal device The transmission power, Indicates terminal device Channel gain to edge server, Indicates terminal device Noise power to the edge server.

[0088] equipment The channel gain to the base station can be expressed as:

[0089]

[0090] Among them, the large-scale fading part Can be expressed as:

[0091]

[0092] in, is the path loss constant, Obeying log-normal distribution shadow fading, for The distance to the base station, Represents the path loss exponent.

[0093] Small scale fading part is time-varying and obeys Rayleigh distribution with unit variance and mean 0. Assume that in the time slot interval t Δ The inner channel will remain in a stable state, then can be modeled as independent first-order complex Gauss-Markov processes:

[0094]

[0095] in, is the channel autocorrelation function, J0(.) depends on the maximum Doppler frequency f d The zero-order Bessel function of the first kind, Obeys a circularly symmetric complex Gaussian distribution.

[0096] Assuming the device side The processing delay for feature extraction and compression is equipment transmission The delay is:

[0097]

[0098] From the device The total delay of transmitting information to the edge server is:

[0099] The goal of the resource allocation method described in the present invention is to determine the optimal feature compression rate and bandwidth share for each device under constraints, thereby maximizing the average target detection accuracy and minimizing the average delay. With the goal of jointly optimizing the average detection accuracy and average delay of the system, the optimization problem P1 can be expressed as:

[0100]

[0101] st

[0102] C1:

[0103] C2:

[0104] C3:

[0105] C4:

[0106] C5:

[0107] Among them, the constraint C1 means that the sum of the bandwidth allocated to all devices is not greater than the total bandwidth resource B max , the constraint C2 defines the maximum degree η of semantic feature compression for each device max , constraint C3 represents the device The data sent should make the detection accuracy greater than the threshold T required by the task min , constraint C4 indicates that the data transmission delay of each device is no greater than the delay threshold t max ,Constraint C5 indicates that the sum of the number of devices that execute different target AI tasks is equal to the total number of devices,D.

[0108] S3. Build and train a deep deterministic policy gradient (DDPG) algorithm model. Based on the bandwidth resource allocation optimization model, determine the optimal resource allocation strategy network in the offline phase.

[0109] In a feasible implementation, the above step S3 may include the following steps S31-S32:

[0110] S31. Define the state space, action space, optimization problem, and reward function in the initial DDPG algorithm model.

[0111] In one feasible implementation, the state space is determined by the performance and environment of the distributed semantic communication system. The state space at the kth time slot can be defined as:

[0112] s k ={T(k),t(k),n1,...,n J}

[0113] Where T(k) represents the average detection accuracy of the kth time slot, t(k) represents the average delay of the kth time slot, J is the number of task categories, and n jis the number of devices executing task j. T(k), t(k) are continuous variables that reflect the optimization degree of the current objective function and system performance, n1,...,n J It is a discrete variable determined by the environment.

[0114] The action space includes the feature compression rate and allocated bandwidth share of each device at time slot k, which can be expressed as:

[0115] a k ={η1(k),...,η D (k),B1(k),...,B D (k)}

[0116] Since the output of the action network in DDPG is a continuous variable, the number of features actually compressed at time slot k for device i is The actual allocated bandwidth is in For the rounding operation, is the floor operation, and F is the total number of features of each image. In order to satisfy the constraints C1 and C2 in the optimization problem P1, the output actions η1(k),...,η D (k) linearly scale them so that their range is [0,η max ], and output actions B1(k),...,B D (k) Perform softmax operation.

[0117] Assuming that each device generates the same amount of data in any time slot, and satisfying the constraints C3 and C4 in the optimization problem P1, the immediate reward is defined as the difference between the current time slot objective function value Z(k) and the previous time slot objective function value Z(k-1). If the constraints C3 and C4 are not satisfied, a penalty is incurred. The reward function can be expressed as:

[0118]

[0119] Among them, Z(k)-Z(k-1) can be further expressed as:

[0120]

[0121] Before the state is input into the neural network, it is batch normalized to unify the value range of the input state. Two scaling factors are introduced For t(k), n1,...,n in the state set, J Zoom in, where

[0122] The training hyperparameters of the DDPG algorithm model are as follows:

[0123] Action network learning rate 0.0001 Evaluation network learning rate 0.0002 Discount Factor 0.9 Sample size 32 Experience replay pool capacity 20000

[0124] S32. In the offline stage, the feature compression scheme and bandwidth resource allocation scheme of the terminal device are used as actions, and the system average detection accuracy, average delay, and the number of devices executing each target task are used as states. By training the intelligent agent, the optimal resource allocation strategy network is obtained.

[0125] In a feasible implementation, Figure 2 As shown in the figure, the DDPG algorithm model has a total of 4 neural networks, namely the action network, the evaluation network, the action target network and the evaluation target network, which are represented by μ(s|θ μ ), Q(s,a|θ Q ), μ'(s|θ μ '), Q'(s,a|θ Q ') indicates that the corresponding neural network parameter is θ μ ,θ Q ,θ μ ',θ Q '. The specific training process is as follows: First, the action network is trained according to the current state s k and exploration noise N to select action a k , where N is the mean μ e , the variance is Gaussian distribution of . The agent performs action a k Get instant rewards k , generating the next state s k+1 , and the tuple (s k ,a k ,r k ,s k+1 ) is stored in the experience replay pool. If the experience replay pool is full, N samples are sampled from it and the current target Q value y is calculated i And update θ through the gradient back propagation of the neural network Q ,y i With the loss function L(θ Q ) can be expressed as follows:

[0126] y k =r k +γQ'(s k+1 ,μ'(s k+1 |θ μ ')|θ Q ')

[0127] L(θ Q )=E μ' [(y k -Q(s k ,a k |θ Q)) 2 ]

[0128] where γ is the discount factor.

[0129] Then, θ μ According to the sampling policy gradient Update the parameters of the action network:

[0130]

[0131] If the number of iterations reaches T, the parameters θ of the evaluation network and the evaluation target network are adjusted. μ ',θ Q 'Perform soft update according to the update coefficient τ:

[0132] θ μ '=τθ μ +(1-τ)θ μ '

[0133] θ Q '=τθ Q +(1-τ)θ Q '

[0134] In a feasible implementation, S32 may include the following steps S321-S328, including:

[0135] S321, initialize the neural network parameters and clear the experience replay pool;

[0136] S322, the action network is based on the current state s k and exploration noise N to select action a k ;

[0137] S323, the agent performs action a k Get instant rewards k , generating the next state s k+1 , and the tuple (s k ,a k ,r k ,s k+1 ) is stored in the experience replay pool;

[0138] S324, repeat steps S322 and S323 until the experience replay pool overflows, and sample N samples from it;

[0139] S325, calculate the current target Q value based on the sample and update the parameters of the evaluation network;

[0140] S326, updating the parameters of the action network according to the sampling strategy gradient;

[0141] S327, each iteration, soft-update the parameters of the action target network and the evaluation target network;

[0142] S328. When the number of iterations reaches a preset threshold, the offline training is completed to obtain a policy network that can achieve optimal resource allocation.

[0143] S4. When in the online stage, the current state information is input into the optimal resource allocation strategy network to determine the optimal feature compression scheme and bandwidth resource allocation scheme.

[0144] In an embodiment of the present invention, the DDPG algorithm in reinforcement learning is adopted to jointly optimize the average detection accuracy and the average transmission delay according to the device's target AI task and the total amount of bandwidth resources to determine the optimal feature compression scheme and the optimal bandwidth allocation scheme. Bandwidth resources are allocated according to the optimal feature compression scheme and the optimal bandwidth allocation scheme, which can alleviate the problem of scarce available bandwidth resources in current wireless communication networks.

[0145] The embodiment of the present invention provides a distributed semantic communication system, which is applied to a bandwidth resource allocation method based on a distributed semantic communication system. The system includes D terminal devices and an edge server; Figure 3 The structural diagram of the distributed semantic communication system shown in FIG.

[0146] A neural network in a deep residual network is deployed on the terminal device to detect specific targets in the image;

[0147] The edge server is deployed with a fully connected layer of a deep residual network for intelligently processing and calculating images uploaded by the terminal device and returning the processing results to the terminal device.

[0148] The embodiment of the present invention provides a bandwidth resource allocation device 400 based on a distributed semantic communication system, which is used to implement a bandwidth resource allocation method based on a distributed semantic communication system. Figure 4 The device structure diagram of the device 400 includes:

[0149] A deployment module 410 is configured to deploy a fully connected layer of the deep residual network on the edge server, and to deploy a neural network before the fully connected layer in the deep residual network on the D terminal devices;

[0150] Establishing module 420, for establishing a bandwidth resource allocation optimization model based on the distributed semantic communication system;

[0151] A construction module 430 is used to construct and train a deep deterministic policy gradient (DDPG) algorithm model, and determine the optimal resource allocation strategy network in the offline phase based on the bandwidth resource allocation optimization model;

[0152] The determination module 440 is configured to input current state information into the optimal resource allocation strategy network when in the online stage, and determine the optimal feature compression scheme and bandwidth resource allocation scheme.

[0153] Optionally, the establishing module 420 is configured to:

[0154] Activate reverse gradient propagation during neural network training, count the contributions of different features to the correct execution of the target AI task, build a regression model of feature compression rate and target AI task detection accuracy, and form a background knowledge base;

[0155] Based on the background knowledge base, a bandwidth resource allocation optimization model for a distributed semantic communication system is constructed with the goal of maximizing the average image detection accuracy and minimizing the average transmission delay.

[0156] Optionally, the building module 430 is configured to:

[0157] Define the state space, action space, optimization problem, and reward function in the initial DDPG algorithm model;

[0158] In the offline stage, the feature compression scheme and bandwidth resource allocation scheme of the terminal device are taken as actions, and the system average detection accuracy, average delay, and the number of devices executing each target task are taken as states. By training the intelligent agent, the optimal resource allocation strategy network is obtained.

[0159] In an embodiment of the present invention, the DDPG algorithm in reinforcement learning is adopted to jointly optimize the average detection accuracy and the average transmission delay according to the device's target AI task and the total amount of bandwidth resources to determine the optimal feature compression scheme and the optimal bandwidth allocation scheme. Bandwidth resources are allocated according to the optimal feature compression scheme and the optimal bandwidth allocation scheme, which can alleviate the problem of scarce available bandwidth resources in current wireless communication networks.

[0160] Figure 5 This is a structural diagram of a bandwidth resource allocation system 500 based on a distributed semantic communication system provided by an embodiment of the present invention. The bandwidth resource allocation system 500 based on a distributed semantic communication system may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 501 and one or more memories 502, wherein at least one instruction is stored in the memory 502, and the at least one instruction is loaded and executed by the processor 501 to implement the steps of the above-mentioned bandwidth resource allocation method based on the distributed semantic communication system.

[0161] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions. The instructions are executable by a processor in a terminal to implement the above-described method for allocating bandwidth resources based on a distributed semantic communication system. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0162] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0163] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A bandwidth resource allocation method based on a distributed semantic communication system, characterized in that: The method is implemented by a distributed semantic communication system, which includes D terminal devices and an edge server; The method comprises: Deploying the fully connected layer of the deep residual network on the edge server, and deploying the neural network before the fully connected layer in the deep residual network on the D terminal devices; Based on the distributed semantic communication system, establishing a bandwidth resource allocation optimization model; Build and train a deep deterministic policy gradient (DDPG) algorithm model, and determine the optimal resource allocation strategy network for the offline phase based on the bandwidth resource allocation optimization model. When in the online stage, the current state information is input into the optimal resource allocation strategy network to determine the optimal feature compression scheme and bandwidth resource allocation scheme; The construction and training of the deep deterministic policy gradient (DDPG) algorithm model, based on the bandwidth resource allocation optimization model, determines the optimal resource allocation strategy network in the offline phase, including: S31. Define the state space, action space, optimization problem, and reward function in the initial DDPG algorithm model. Among them, the state space at the kth time slot is defined as: s k ={T(k),t(k),n1,...,n J } Where T(k) represents the average detection accuracy of the kth time slot, t(k) represents the average delay of the kth time slot, J is the number of task categories, and n j is the number of devices executing task j; The action space includes the feature compression rate and allocated bandwidth ratio of each device at time slot k, expressed as: and k ={η1(k),...,η D (k),B1(k),...,B D (to)} For device i, the number of features actually compressed at time slot k is The actual allocated bandwidth is B max Represents the total bandwidth resources, where For the rounding operation, is the floor operation, F is the total number of features of each image, in order to meet the constraints C1 and C2 in the optimization problem P1, the output actions η1(k),...,η D (k) linearly scale them so that their range is [0,η max ], and output actions B1(k),...,B D (k) performs softmax operation, The reward function is expressed as: Among them, Z(k)-Z(k-1) is further expressed as: Among them, the optimization problem P1 is expressed as: Among them, the constraint C1 means that the sum of the bandwidth allocated to all devices is not greater than the total bandwidth resource B max , the constraint C2 defines the maximum degree η of semantic feature compression for each device max , constraint C3 represents the device The data sent should make the detection accuracy greater than the threshold T required by the task min , constraint C4 indicates that the data transmission delay of each device is no greater than the delay threshold t max ,Constraint C5 indicates that the sum of the number of devices that perform different target AI tasks is equal to the total number of devices D; S32. In the offline phase, the feature compression scheme and bandwidth resource allocation scheme of the terminal device are used as actions, and the average detection accuracy, average latency, and number of devices executing each target task of the system are used as states. An optimal resource allocation strategy network is obtained by training the intelligent agent. The specific training process is as follows: First, the action network is trained according to the current state s k and explore noise Select action a k ,in is subject to the mean μ e , the variance is Gaussian distribution of; the agent performs action a k Get instant rewards k , generating the next state s k+1 , and the tuple (s k ,a k ,r k ,s k+1 ) is stored in the experience replay pool; if the experience replay pool is full, N samples are sampled from it and the current target Q value y is calculated i And update θ through the gradient back propagation of the neural network Q ,y i With the loss function L(θ Q ) is represented as follows: y k =r k +γQ'(s k+1 ,m'(s k+1 |θ μ' )|θ Q' ) Where γ is the discount factor; Then, θ μ According to the sampling policy gradient Update the parameters of the action network: If the number of iterations reaches Then the parameters θ of the evaluation network and the evaluation target network are μ' ,θ Q' Soft update is performed according to the update coefficient τ: i μ' =tθ μ +(1-τ)θ μ' i Q' =tθ Q +(1-τ)θ Q' 。 2. The method according to claim 1, characterized in that The method of deploying the fully connected layer of the deep residual network on the edge server and deploying the neural network before the fully connected layer in the deep residual network on the D terminal devices includes: Using the convolutional layers in a 50-layer deep residual network as a semantic encoder, and using the fully connected layers in the 50-layer deep residual network as a semantic decoder; A distributed architecture is adopted to deploy the neural network before the fully connected layer on the D terminal devices, and the fully connected layer is deployed on the edge server.

3. The method according to claim 1, characterized in that The step of establishing a bandwidth resource allocation optimization model based on the distributed semantic communication system includes: Activate reverse gradient propagation during neural network training, count the contributions of different features to the correct execution of the target AI task, build a regression model of feature compression rate and target AI task detection accuracy, and form a background knowledge base; Based on the background knowledge base, a bandwidth resource allocation optimization model for a distributed semantic communication system is constructed with the goal of maximizing the average image detection accuracy and minimizing the average transmission delay.

4. The method according to claim 3, characterized in that The detection accuracy is expressed by the following formula (1): Where j represents the jth execution task; i represents the i-th terminal device, and the value range of i is [1,n j ], where n j represents the number of devices executing task j, J represents the number of task categories, and D represents the number of terminal devices; is the parameter obtained by the logistic regression method, and the loss function is the mean square error; represents the feature compression rate of the i-th device corresponding to the j-th execution task; Transmission delay is the total delay of transmitting information from the terminal device to the edge server, which is expressed by the following formula (2): in, Indicates terminal device Processing delay for feature extraction and compression; In formula (2) Indicates terminal device transmission The time delay is expressed by the following formula (3): in, Represents each time slot interval t Δ Internal terminal equipment The amount of data generated Indicates the actual amount of data to be transmitted after semantic compression; terminal equipment Data transfer rate for: It is a terminal device The allocated bandwidth, P represents the terminal device The transmission power, Indicates terminal device Channel gain to edge server, Indicates terminal device Noise power to the edge server.

5. The method according to claim 1, wherein In the offline phase, the terminal device's feature compression scheme and bandwidth resource allocation scheme are used as actions, and the system's average detection accuracy, average latency, and the number of devices executing each target task are used as states. By training the agent, an optimal resource allocation strategy network is obtained, including: Initialize the neural network parameters and clear the experience replay pool; The action network is based on the current state s k and explore noise Select action a k ; The agent performs action a k Get instant rewards k , generating the next state s k+1 , and the tuple (s k ,a k ,r k ,s +k1 ) is stored in the experience replay pool; Repeat steps (2) and (3) until the experience replay pool overflows, and sample N samples from it; Calculate the current target Q value based on the sample and update the parameters of the evaluation network; Update the parameters of the action network according to the sampled policy gradient; Each iteration, the parameters of the action target network and the evaluation target network are soft updated; When the number of iterations reaches a preset threshold, offline training is completed and a policy network that can achieve optimal resource allocation is obtained.

6. A distributed semantic communication system, characterized in that: The distributed semantic communication system is used to implement the bandwidth resource allocation method based on the distributed semantic communication system according to any one of claims 1 to 5, wherein the distributed semantic communication system includes D terminal devices and an edge server; wherein: A neural network in a deep residual network is deployed on the terminal device to detect specific targets in the image; The edge server is deployed with a fully connected layer of a deep residual network for intelligently processing and calculating images uploaded by the terminal device, and returning the processing results to the terminal device.

7. A bandwidth resource allocation device based on a distributed semantic communication system, characterized in that: The bandwidth resource allocation device based on the distributed semantic communication system is used to implement the bandwidth resource allocation method based on the distributed semantic communication system according to any one of claims 1 to 5, and the device includes: A deployment module, configured to deploy the fully connected layer of the deep residual network on the edge server, and deploy the neural network before the fully connected layer in the deep residual network on the D terminal devices; An establishment module is used to establish a bandwidth resource allocation optimization model based on the distributed semantic communication system; A construction module is used to build and train a deep deterministic policy gradient (DDPG) algorithm model, and determine the optimal resource allocation strategy network in the offline phase based on the bandwidth resource allocation optimization model. The determination module is used to input the current state information into the optimal resource allocation strategy network when in the online stage, and determine the optimal feature compression scheme and bandwidth resource allocation scheme.

8. The device according to claim 7, characterized in that The establishment module is used to: Activate reverse gradient propagation during neural network training, count the contributions of different features to the correct execution of the target AI task, build a regression model of feature compression rate and target AI task detection accuracy, and form a background knowledge base; Based on the background knowledge base, a bandwidth resource allocation optimization model for a distributed semantic communication system is constructed with the goal of maximizing the average image detection accuracy and minimizing the average transmission delay.

9. The device according to claim 7, characterized in that The building blocks are used to: Define the state space, action space, optimization problem, and reward function in the initial DDPG algorithm model; In the offline stage, the feature compression scheme and bandwidth resource allocation scheme of the terminal device are taken as actions, and the system average detection accuracy, average delay, and the number of devices executing each target task are taken as states. By training the intelligent agent, the optimal resource allocation strategy network is obtained.

Images