A semantic communication method for semantic source-channel adaptive coding under parallel channels

Abstract

The present invention discloses a semantic communication method for semantic source-channel adaptive coding under parallel channels, comprising the following steps: S1. constructing a communication model for semantic source-channel adaptive coding under parallel channels; S2. at the sending end, encoding the semantic source samples in ( X, S ), and distributing the bitstream after encoding each semantic source sample to an independent channel for transmission. K channel encoders with different channel rates are adopted to protect their respective bitstreams against channel errors; S3. at the receiving end, using K channel decoders to decode the received signal and combining it into a single serial stream, which is decoded by a semantic source decoder based on a deep neural network; S4. determining the source rate and prior information under any given deep neural network model; S5. constructing an optimization problem to jointly optimize the deep neural network parameter Φ and the resource allocation variables. The present invention realizes the optimization of resource allocation in the parallel Gaussian channel scenario.

Description

Technical Field

[0001] The present invention relates to the field of semantic communication, and particularly to a semantic communication method for semantic source-channel adaptive coding under parallel channels. Background Art

[0002] With the rise of intelligent services and the deep integration of artificial intelligence in wireless communication, task-oriented semantic communication technology is regarded as an important development direction for future 6G wireless networks. Different from traditional communication systems that focus on bit-level precise transmission, semantic communication extracts and transmits the intrinsic semantic information of data sources related to specific tasks (such as object detection, edge inference, etc.), significantly improving the transmission efficiency under limited communication resources.

[0003] However, existing semantic communication technologies face the following key challenges in practical applications: 1. Compatibility issues with digital hardware. Existing joint source-channel coding schemes based on deep neural networks mainly generate analog symbols for transmission, making it difficult to be compatible with modern digital communication systems and restricting their practical deployment; 2. Insufficient adaptability. Current deep learning methods rely on end-to-end training for specific tasks and channel conditions. When task requirements or channel conditions change, the entire network needs to be redesigned and retrained, which is difficult to achieve in actual communication scenarios; 3. Lack of complex channel adaptability. Existing research mainly focuses on additive white Gaussian noise channels, and there is insufficient research on resource allocation and rate adaptation strategies in more complex channel scenarios such as frequency-selective channels and MIMO channels. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a semantic communication method for semantic source-channel adaptive coding under parallel channels. By designing an architecture that combines deep neural network semantic source coding with parallel digital channel coding, the problems of poor compatibility between existing semantic communication schemes and digital systems and insufficient channel adaptability are solved, and at the same time, the optimization of resource allocation in parallel Gaussian channel scenarios is achieved.

[0005] The purpose of the present invention is realized through the following technical solutions: A semantic communication method for semantic source-channel adaptive coding under parallel channels, comprising the following steps:

[0006] S1. Construct a communication model for semantic source-channel adaptive coding under parallel channels;

[0007] Let the semantic source be characterized by the joint probability distribution where and respectively represent the external observation and internal semantic state of the source, ( X, S ) represents the semantic source pair, represents ( X, S)'s joint probability distribution, the sender transmits the observations through K parallel channels to X enable the receiver to obtain the restored semantic source pair , where is the extrinsic observation of the restored source, is the semantic state of the restored source, where K parallel Gaussian channels are represented by the discrete set ;

[0008] S2. At the sender, use the semantic source encoder to encode the semantic source samples in the semantic source pair ([[]] X, S ). Each encoded bitstream of the semantic source sample is assigned to an independent channel for transmission, and K channel encoders with different channel rates are used to protect their respective bitstreams against channel errors;

[0009] S3. At the receiver, use K channel decoders to decode the received signal and merge it into a single serial stream, which is decoded by the semantic source decoder based on a deep neural network;

[0010] S4. Determine the source rate and prior information according to an arbitrarily given deep neural network model;

[0011] S5. Construct an optimization problem and perform joint optimization of the deep neural network parameters Φ and the resource allocation variable .

[0012] The beneficial effects of the present invention are: Through the architecture design that combines deep neural network semantic source coding and parallel digital channel coding, the problems of poor compatibility between existing semantic communication schemes and digital systems and insufficient channel adaptability are solved, and at the same time, resource allocation optimization in the parallel Gaussian channel scenario is achieved; at the same time, the key network parameters are determined, prior information is given for each model, a look-up table is formed, and the key performance indicators of each model are characterized, reducing the implementation difficulty. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a schematic diagram of the principle of the present invention;

[0014] Figure 2 is a schematic diagram of the principle of the semantic source encoder;

[0015] Figure 3 is a schematic diagram of the principle of the semantic source decoder;

[0016] Figure 4 is a schematic diagram of the relationship between the maximum transmission power and the signal-to-noise ratio in the embodiment;

[0017] Figure 5Schematic diagram of the relationship between the maximum transmission power and the accuracy rate in the embodiment. Detailed implementation manners

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0019] As Figure 1 shown, a semantic communication method for semantic source-channel adaptive coding in a parallel channel includes the following steps:

[0020] S1. Construct a communication model for semantic source-channel adaptive coding in a parallel channel;

[0021] Consisting of K parallel Gaussian channels, represented by the discrete set . These parallel channels can be equivalently implemented through different communication technologies. For example, in an OFDM system, the total bandwidth is divided into multiple orthogonal subcarriers, and each subcarrier can be regarded as a parallel channel; in a multi-antenna system, the MIMO channel can be decomposed into parallel channels through the technique of singular value decomposition. The semantic source is characterized by the joint probability distribution , where and represent the external observation and the internal semantic state of the source respectively. The purpose of the system is to transmit the observation K through X parallel channels so that the receiver can jointly recover , where is the external observation of the recovered source, and is the semantic state of the recovered source. For this purpose, the present invention proposes an adaptive digital source-channel coding scheme, which includes a general semantic source encoder and K parallel channel encoders.

[0022] S2. At the transmitter, let ([[]]END]] x, s ) represent the sample of the semantic source for ([[]]END]] X, S ). The specific steps of the semantic source encoder are as Figure 2 shown. The semantic source encoder can access the observation sample x , where x is related to the semantic state through the conditional probability s . The specific implementation steps are as follows:

[0023] Step 1: Deep neural network feature extraction: Map through the deep neural network with parameters x to a W -dimensional continuous value vector y , that is: , where,W much smaller than x the original dimension ; denote the feature extraction function of the deep neural network with parameter ;

[0024] Step 2: Quantization. Perform uniform scalar quantization on y to obtain: where denotes the uniform scalar quantization operation.

[0025] Step 3: Arithmetic coding. Based on the probability mass function of the quantization vector , use lossless entropy coding to encode into a bitstream b , where denotes the random variable corresponding to . The expected source rate for encoding each source sample is defined as the entropy of : .

[0026] After semantic source coding, the bitstream after encoding each semantic source sample is assigned to an independent channel for transmission. Considering that the conditions of these channels are different, K channel encoders with different channel rates are used to protect their respective bitstreams against channel errors. For the k th channel ( ), the assigned bitstream, defined as , is encoded by block channel codes, where denotes the number of data bits in each block, L is the code block length. Specifically, the bitstream with a total length of is divided into blocks, where denotes the ceiling operation. Each block is encoded into a symbol sequence L with a length of , denotes the k th symbol in the th block channel codeword in the k th channel. The channel rate of the th channel is calculated as

[0027] (10)

[0028] where is the kThe proportion of semantic source samples transmitted by each channel. To evaluate the transmission efficiency, the average bandwidth ratio is defined as , which is used to measure the average number of channel uses required for each element of the observed state.

[0029] Next, in the channel transmission part, for the k th Gaussian channel, the input-output relationship of the t th block is:

[0030] (11)

[0031] where l = 1,..., L , , ; represents the transmitted symbol, with zero mean and average power of 1; is the corresponding received symbol; is independent and identically distributed zero-mean circularly symmetric complex Gaussian noise with variance ; is the power allocated to channel k , satisfying the total power constraint: , where is the maximum transmit power of the system. The received signal-to-noise ratio of the k th channel is defined as .

[0032] S3. At the receiver, the received signal is decoded by K channel decoders and merged into a single serial stream, which is then decoded by a semantic source decoder based on a deep neural network;

[0033] At the receiver, for the k th channel, the received symbol sequence is first decoded by the channel decoder k into the recovered bit stream . Due to transmission errors, the recovered bit stream may not be equal to the transmitted bit stream. Then, the bit streams K from the channel decoders are merged into a single serial stream through a parallel-to-serial converter, waiting to be decoded by a common semantic source decoder based on a deep neural network. The semantic source decoding process is as shown in Figure 3 , where the recovered bit stream corresponding to the semantic source sample ( x, s ) is denoted as . Note that may correspond to any one of the decoded bit streams , depending on which parallel channel the semantic source sample ( x, s ) is transmitted through. Specifically, First, it is decoded by an arithmetic decoder into a recovered feature vector . Then, it is input into a deep neural network recovery function θ with parameters to recover the observed sample x , and the recovered observed sample is . Finally, the semantic state is recovered through a maximum a posteriori probability scheme s , that is: , where is the recovered semantic state, P ( s|x̂, ψ ) is the posterior probability estimated by a deep neural network with parameters ψ .

[0034] Correspondingly, the recovered semantic source sample is affected by lossy compression and transmission errors; define the end-to-end observation distortion and semantic distortion of the k th channel as:

[0035] (12)

[0036] (13)

[0037] where and represent the observation distortion metric and the semantic distortion metric respectively, represents the noise under channel k . and represent the mathematical expectations of the end-to-end observation distortion k and the semantic distortion of the th channel respectively. To evaluate the average performance of the considered system, we define as the expectation of the weighted sum of the two distortions on K parallel channels, that is:

[0038] (14)

[0039] where α ∈[0,1] is the weight coefficient between the observation distortion and the semantic distortion .

[0040] S4. Determine the source rate and prior information according to an arbitrarily given deep neural network model;

[0041] First, for any given deep neural network model, by using the data fitting method, it can be obtained that the logarithm of the observation distortion and the semantic distortion can be approximated by the logistic function respectively, that is:

[0042] (15)

[0043] (16)

[0044] where represents the neural network parameters in the semantic source codec, represents the channel k in which the bit error probability, and respectively represent the logarithm of the observation distortion and the semantic distortion under the fixed deep neural network model. These two parameters correspond to the distortion caused by source compression. and are the parameters representing the maximum distortion increment caused by channel errors. and are the logistic function growth rate parameters, which are used to describe the rate of change of distortion with the bit error rate. and are the logistic function midpoint parameters, indicating the bit error rate corresponding to when the distortion reaches the intermediate value. The parameters , , , , , , and can all be obtained by data fitting (previously through conventional algorithms, such as the least mean square error, etc.).

[0045] Then, for the k th Gaussian channel, under the conditions of signal-to-noise ratio , code block length L and channel rate , the bit error probability under random channel block coding can be approximated as:

[0046] (17)

[0047] where satisfies , the function is defined as , is the channel capacity expression, which satisfies .

[0048] S5. Construct an optimization problem and jointly optimize the deep neural network parameters Φ and resource allocation variables .

[0049] Based on the above analysis and modeling, the objective of the present invention is to minimize the average distortion by optimizing the following two aspects : the parameters of the semantic source encoder and decoder and K the power allocation and channel rate adaptation strategies of multiple parallel channels. The corresponding optimization problem is

[0050] (18)

[0051] To solve the above optimization problem, the present invention proposes an optimization algorithm based on model selection to decouple the joint optimization of the deep neural network parameters Φ and resource allocation variables .

[0052] Step 1: Construct a pre-trained model. First, construct a lookup table containing N pre-trained deep neural network models

[0053] 1. The pre-training process refers to training the deep neural network with parameters in the semantic source codec. By changing some adjustable hyperparameters of the neural network, N deep neural networks with different performances can be trained. The specific training process is as follows: First, consider a semantic data set composed of a large number of semantic source samples ( x, s ), and preset the hyperparameters of the neural network. Since the semantic source encoder in this patent only inputs the observed samples x , the input of the neural network is only a large number of observed samples x . The observed sample x is first input into the deep neural network with parameters in the semantic encoder; then, the obtained output is directly input into the deep neural network with parameters θ , and the output is the restored observed sample ; finally, is input into the deep neural network with parameters , and the output is the restored semantic state . Correspondingly, the loss function of the neural network is the weighted sum of the distortion between s and and the distortion between x and . The training objective is to make the loss function as small as possible. Use the backpropagation algorithm to calculate the gradient, and adopt the gradient descent method to update the network parameters Φ, and iterate and optimize until convergence or reach the preset number of training rounds. Here, convergence means that the loss function is less than the set threshold value.

[0054] 2. For each pre-trained model, perform distortion characteristic fitting. The specific steps are as follows:

[0055] (1) Set a set of discrete value sets of the bit error rate covering the error situations under actual channel conditions;

[0056] (2) For each value: Input a large number of source samples ( x, s ) into the semantic source encoder, and each sample corresponds to an output bitstream b . Randomly introduce bit errors corresponding to the error probability b artificially in . Use this bitstream with introduced bit errors as the recovered bitstream . Then decode through the semantic source decoder to obtain the corresponding and . Correspondingly, the calculation method of the source rate is: Calculate the average length of the bitstream b , which can be approximately expressed as the source rate . Finally, obtain the actual values of the corresponding observed distortion and semantic distortion according to equations (3) and (4).

[0057] (3) For each pre-trained model, according to the experimental data obtained in (1) and (2), use the least mean square error criterion to fit the parameters , , , , , , and in the generalized Logistic functions (6) and (7).

[0058] Save each pre-trained deep neural network model and the corresponding source rate , as well as the parameters , , , , , , and in the lookup table;

[0059] Step 2: Resource allocation optimization. For each given pre-trained model, that is, fix the deep neural network parameters Φ, and use classical optimization techniques, such as the successive convex approximation algorithm, to solve the optimization problem (P1), so as to obtain the power allocation and channel rate control scheme under the given pre-trained model.

[0060] Step 3: Optimal solution search. Solve the above resource allocation optimization problem for each pre-trained model according to Step 2 to obtain the corresponding minimum objective function value. Search in the lookup table for the model that can achieve the minimum objective function value. This model and its corresponding resource allocation scheme are the optimal solutions to the original optimization problem (P1).

[0061] In this section, through image restoration and classification tasks, the effectiveness of the proposed parallel Gaussian channel task-oriented semantic source-channel adaptive coding (ADSCC) scheme is verified. Among them, the image is used as the sample of the external observation X while its label corresponds to the sample of the internal semantic state S ; the observation distortion is the mean square error distortion, and the semantic distortion is the Hamming distortion. The CUB-200-2011 dataset is used for experiments. This dataset contains 11,788 images of 200 bird species: training set: 5,994 images; test set: 5,794 images; all images are randomly cropped to 256×256 pixels.

[0062] Model architecture of the neural network: Feature extraction function and feature recovery function : A deep neural network model adopting the classical hyper-prior model; for the image classification task: the Resnet-152 network is adopted; the lookup table contains 23 deep neural network models of pre-trained hyper-prior models, and the range of each bit pixel value is from 0.02 to 1.41

[0063] Channel setting: Parallel Gaussian channel K = 8; different channel conditions are represented by different noise variances Specific noise variance values (in ascending order): (0.4819, 0.6354, 0.8098, 1.4888, 1.6768, 2.8978, 8.8339, 9.1324)

[0064] Comparison scheme: Deep joint source-channel coding scheme; Separate source-channel coding scheme: Its source coding scheme is BPG, and the channel coding scheme is rate of ( , 256) Polar code and rate of ( , 4096) LDPC code. The power allocation strategy of the comparison scheme is the truncated channel inversion strategy, that is, allocate the total power to the first channels:

[0065] (19)

[0066] InFigure 4 In the experiment, we compared the proposed method with existing separate source-channel coding schemes: BPG+LDPC or BPG+Polar, and deep joint source-channel coding schemes. The abscissa is the maximum transmission power, and the ordinate is the peak signal-to-noise ratio, which is used to measure the recovery quality of the image. From the figure, we can see that our method has significant performance gains compared with existing separate source-channel coding schemes and deep joint source-channel coding schemes at different maximum transmission powers.

[0067] In Figure 5 the experiment, the abscissa is the maximum transmission power, and the ordinate is the accuracy rate, which is used to measure the quality of the image classification task. Our method also has significant performance gains compared with existing separate source-channel coding schemes and deep joint source-channel coding schemes at different maximum transmission powers.

[0068] The above are the preferred embodiments of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in other combinations, modifications, and environments, and can be changed within the scope of the concept described herein through the above teachings or the techniques or knowledge in related fields. And the changes and modifications made by those skilled in the art without departing from the spirit and scope of the present invention should all be within the protection scope of the appended claims of the present invention.

Claims

1. A semantic communication method for semantic source channel adaptive coding under parallel channels, characterized by: The following steps are involved: S1. Construct a communication model for semantic source-channel adaptive coding under parallel channels; Assume that the semantic information source is composed of the joint probability distribution Characterization, where and Represent the external observation and internal semantic state of the source respectively, ( X,S ) represents a semantic source pair, express( X,S ), the sender passes K Parallel channel transmission observation X , so that the receiver can obtain the restored semantic source pair ,in, is the external observation of the recovered source, is the semantic state of the recovered source, where K discrete set express; S2. At the sending end, the semantic source encoder is used to encode the semantic source pair ( X,S ), and the bit stream after each semantic source sample is encoded is assigned to an independent channel for transmission. K channel encoders with different channel rates to protect their respective bit streams against channel errors; S3. At the receiving end, the received signals are decoded by K channel decoders and merged into a single serial stream, which is decoded by a semantic source decoder based on a deep neural network; S4. Determine the source rate and prior information according to any given deep neural network model; S5. Construct optimization problems and perform deep neural network parameters Φ and resource allocation variables Joint optimization of Assign to channel k The power, For the k The channel rate of the channel.

2. The semantic communication method of semantic source channel adaptive coding under parallel channels according to claim 1, characterized in that: In step S2, at the transmitting end, a semantic source encoder is used to encode the semantic source pair ( X,S ) to encode the semantic source samples in, including: set up( x,s ) represents the semantic source pair ( X,S ), the semantic source encoder can access the observed samples x ,in x Through conditional probability and semantic state s The specific implementation steps are as follows: Step 1: Deep neural network feature extraction: Through the parameters Deep Neural Network x Mapping W dimensional continuous-valued vector y ,Right now: ,in, W Much smaller than x The original dimension ; Indicates that the parameter is Feature extraction function of deep neural network; Step 2: Quantification: y Perform uniform scalar quantization to obtain: ,in represents a uniform scalar quantization operation; Step 3: Arithmetic coding: based on quantized vector The probability mass function of , using lossless entropy coding Encoding as bitstream b ,in Representation and The corresponding random variable is The expected source rate at which each source sample is encoded Defined as Entropy of: ,in Express expectations.

3. The semantic communication method of semantic source channel adaptive coding under parallel channels according to claim 2, characterized in that: In step S2, the bit stream after encoding each semantic source sample is allocated to an independent channel for transmission, and the bit stream is transmitted by K Channel encoders with different channel rates to protect their respective bit streams against channel errors include: After semantic source coding, the encoded bit stream of each semantic source sample is assigned to an independent channel for transmission; considering the different conditions of these channels, K Channel encoders with different channel rates protect their respective bit streams against channel errors: For k channels, , whose allocated bit stream is defined as ,Depend on Block channel code is used for encoding, where Indicates the number of data bits in each block, L is the code block length, specifically: The total length is Bitstream Divided into blocks, of which Indicates the rounding operation, each block is encoded as a length of L The symbol sequence , Indicates k Channel The first Symbol, k The channel rate of a channel is calculated as ; The average channel usage times for transmitting one source sample is calculated as: （1） in It is through k The ratio of semantic source samples transmitted by each channel is defined as the average bandwidth ratio: , used to measure the average number of channel usages required to observe each element of the state; Channel transmission part, k In a Gaussian channel t The input and output relationship of each block is: （2） in, l =1,..., L , , ; Indicates the transmitted symbol, the mean is zero, and the average power is 1; is the corresponding received symbol; is an independent and identically distributed zero-mean cyclically symmetric complex Gaussian noise with a variance of ; Assign to channel k The power meets the total power constraint: ,in, is the maximum transmit power of the system, k The received signal-to-noise ratio of a channel is defined as .

4. The semantic communication method of semantic source channel adaptive coding in parallel channels according to claim 3, characterized in that: The step S3 comprises: At the receiving end, set k The received symbol sequence of the channel , by the channel decoder k Decode to recovered bitstream ; When a transmission error occurs, the recovered bit stream Not equal to the transmitted bit stream; Then, from K The bit stream of the channel decoder Merge into a single serial stream via a parallel-to-serial converter, waiting to be decoded by a common deep neural network-based semantic source decoder; The decoding process of the semantic source is as follows: Assume that the semantic source sample ( x,s ) is represented as , May correspond to any decoded bitstream , which depends on the semantic source sample ( x,s ) through which parallel channel it is transmitted; It is first decoded by the arithmetic decoder into a recovered feature vector ;then, Input parameters are θ The deep neural network recovery function Restore observation samples x , the restored observation sample is ; Finally, the semantic state is restored through the maximum a posteriori probability scheme s ,Right now: ,in is the restored semantic state, P ( s|x̂,ψ ) is composed of parameters ψ The posterior probability estimated by the deep neural network; Recovered semantic source samples subject to lossy compression and transmission errors; define k The end-to-end observed distortion of the channel and semantic distortion They are: （3） （4） in and denote the observation distortion measure and the semantic distortion measure respectively, Indicates channel k Noise below; and Respectively represent k The end-to-end observed distortion of the channel and semantic distortion The mathematical expectation of for K The expectation of the weighted sum of two distortions on parallel channels is: （5） in α ∈[0,1] is the observation distortion and semantic distortion The weight coefficient between .

5. The semantic communication method of semantic source channel adaptive coding in parallel channels according to claim 4, characterized in that: The step S4 comprises: For any given deep neural network model, the logarithmic value of the observation distortion and the semantic distortion are approximated by the logistic function, namely: （6） （7） in, Describe the neural network parameters in the semantic source codec, Indicates channel k The bit error probability in and They represent the logarithm of the observed distortion under a fixed deep neural network model. and semantic distortion The lower asymptote of ; these two parameters correspond to the distortion caused by source compression; and It is a parameter that characterizes the maximum distortion increment caused by channel error; and is the growth rate parameter of the logistic function, which is used to describe the rate at which the distortion changes with the bit error rate; and is the midpoint parameter of the logistic function, indicating the bit error rate corresponding to the intermediate value of the distortion; parameter , , , , , , and The method is obtained by data fitting, wherein the data fitting method includes minimum mean square error; Then, for the k Gaussian channel, at a signal-to-noise ratio , code block length L and channel rate Under the condition of , the bit error probability under random channel block coding is approximately: （8） in, satisfy , The function is defined as , is the channel capacity expression, satisfying .

6. The semantic communication method of semantic source channel adaptive coding in parallel channels according to claim 1, characterized in that: In step S5, the optimization problem objective is to minimize the average distortion by optimizing the following two aspects: : Parameters of semantic source encoder and decoder and K The power allocation and channel rate adaptation strategy of parallel channels, the corresponding optimization problem is: （9） Through the optimization algorithm based on model selection, the deep neural network parameters Φ and resource allocation variables are Joint optimization.

7. The semantic communication method of semantic source channel adaptive coding in parallel channels according to claim 6, characterized in that: The deep neural network parameters Φ and resource allocation variables are optimized by the model selection-based optimization algorithm. Joint optimization of A1. Pre-trained model construction: Build a lookup table containing N pre-trained deep neural network models: A101. The pre-training process refers to training the semantic source codec with parameters By changing the hyperparameters of the neural network, N deep neural networks with different performances are obtained. For any deep neural network, the training process is as follows: Consider a semantic dataset consisting of a large number of semantic source samples ( x,s ) and pre-set the hyperparameters of the neural network; the semantic source encoder only inputs the observed samples x , the input of the neural network is just a large number of observation samples x ; Observation sample x First, the parameters are input into the semantic encoder. The output is then directly fed into a deep neural network with parameters θ The deep neural network output is the restored observation sample ;at last, Input parameters are The deep neural network output is the restored semantic state ; Correspondingly, the loss function of the neural network is s and Distortion between x and The weighted sum between distortions, the training goal is to make the loss function as small as possible; use the back propagation algorithm to calculate the gradient, use the gradient descent method to update the network parameter Φ, and iterate the optimization until convergence or reach the preset number of training rounds; After training N deep neural networks with different performances, N pre-trained deep neural network models are obtained; A102. Fit the distortion characteristics of each pre-trained deep neural network model. The specific steps are as follows: (1) Setting the bit error rate A set of discrete values ​​of , covering the error conditions under actual channel conditions; (2) For each Value: A large number of source samples ( x,s ) Input semantic source encoder, each sample corresponds to an output bit stream b ,exist b The probability of error caused by random introduction bit error, and use this bit stream with bit errors as the recovery bit stream ,Then After being decoded by the semantic source decoder, the corresponding and ; Correspondingly, the source rate The calculation method is: calculate the bit stream b The average length of ; Finally, according to equations (3) and (4), the corresponding observation distortion is obtained and semantic distortion The actual value of (3) For each pre-trained model, the parameters in the generalized logistic function (6) and (7) are fitted using the minimum mean square error criterion based on the data obtained from (1) and (2). , , , , , , and ; A103. Save each pre-trained deep neural network model and the corresponding source rate in the lookup table , and parameters , , , , , , and ; A2. Resource allocation optimization: For each given pre-trained model, i.e., fixed deep neural network parameter Φ, classical optimization techniques, including continuous convex approximation algorithm, are used to solve the optimization problem P1, thereby obtaining the power allocation and channel rate control scheme under the given pre-trained model; A3. Optimal solution search: Solve the resource allocation optimization problem for each pre-trained model to obtain the corresponding minimum objective function value; search the lookup table for a model that can achieve the minimum objective function value. This model and its corresponding resource allocation solution are the optimal solutions to the original optimization problem P1.

Images