An adaptive relay semantic coding method based on deep learning

Abstract

This invention relates to the field of adaptive semantic coding, specifically to an adaptive relay semantic coding method based on deep learning. The scheme includes: constructing a relay semantic communication system based on a deep neural network, the system consisting of a transmitter, relay nodes, and a receiver, including a direct link from the transmitter to the receiver, a relay receiving link from the transmitter to the relay, and a relay forwarding link from the relay to the receiver. A half-duplex DF relay system is adopted, i.e., the relay operates on two orthogonal time slots or frequency bands, such that the relay's reception and decoding of block t corresponds to its forwarding of block t+1. At the transmitter, a bitstream is generated through deep semantic source coding and entropy coding and interfacing with digital channel coding. Simultaneously, half-duplex DF relay is introduced, and cooperative gain is obtained through random binning and two-stage continuous decoding. An end-to-end distortion calculation model is established under finite code length conditions, and finally, joint optimization is performed. This invention is applicable to adaptive semantic coding.

Description

Technical Field

[0001] This invention relates to the field of adaptive semantic coding, and more specifically to an adaptive relay semantic coding method based on deep learning. Background Technology

[0002] Traditional digital communication systems are typically designed and resource-allocated based on core metrics such as bit error rate, throughput, or latency, without explicitly characterizing the semantic contribution of information to downstream tasks (such as classification, detection, and retrieval). While deep learning can be used to extract and transmit low-dimensional semantic features in semantic communication, thereby reducing transmission overhead to some extent, existing technologies generally suffer from the following shortcomings: First, compression ratios and channel coding parameters are mostly fixed settings, making it difficult to adaptively adjust them in real time according to channel conditions; second, most solutions are geared towards point-to-point links, lacking a systematic mechanism design for relay cooperation scenarios; third, under finite code length conditions, end-to-end decoding errors significantly affect the performance of semantic tasks, and existing methods struggle to form a computable end-to-end semantic distortion model, resulting in a lack of quantitative basis for resource allocation; fourth, there is a lack of a complete technical path that is compatible with digital communication systems, can utilize relay cooperation gains, and enables online adaptive parameter tuning. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive relay semantic coding method based on deep learning to improve the performance of semantic tasks and the robustness of the system.

[0004] This invention achieves the above objectives by adopting the following technical solution: This invention provides an adaptive relay semantic coding method based on deep learning, comprising:

[0005] S1. Construct a relay semantic communication system based on deep neural networks;

[0006] The deep neural network-based relay semantic communication system consists of a transmitter, a relay node, and a receiver. It includes a direct link from the transmitter to the receiver, a relay receiving link from the transmitter to the relay, and a relay forwarding link from the relay to the receiver. It adopts a half-duplex DF relay system, that is, the relay operates on two orthogonal time slots or frequency bands, so that the relay's reception and decoding of the t-th block corresponds to its forwarding of the t+1-th block.

[0007] S2. The transmitter generates a bitstream through deep semantic source coding and entropy coding and interfaces it with digital channel coding. At the same time, a half-duplex DF relay is introduced, and cooperative gain is obtained through random binning and two-stage continuous decoding.

[0008] S3. Establish an end-to-end distortion calculation model under the condition of limited code length;

[0009] S4. Jointly optimize the source coding rate, channel coding rate, and power allocation to minimize end-to-end weighted distortion and satisfy time delay and power constraints.

[0010] Furthermore, step S2 specifically includes:

[0011] At the transmitting end, for the input sample pair When performing deep source coding, firstly... Input semantic coding network Feature extraction is performed to obtain low-dimensional continuous semantic features. ,in For neural network parameters, then for Quantization is performed to obtain discrete semantic features Secondly, for Arithmetic coding and entropy coding are performed to generate a bitstream. The source coding rate is defined as Finally, output the bit stream. and the corresponding source rate To proceed to the subsequent channel coding and transmission process;

[0012] When interfacing with a digital communication system, the transmitting end transmits the bit stream. To perform channel coding, firstly... By length Divided into blocks Then, using a block length of The channel encoder for each Encoding generates symbol sequences Secondly, the transmitter channel coding rate is defined as... To compile all the codewords into a codebook Finally, output With bit rate For use in broadcast transmission.

[0013] When broadcasting from the transmitting end, the power is first used as the starting point. Send the Block code Then, the relay node and the receiver receive the signal on their respective links, with arbitrary symbol dimensions. The above satisfies ,in For channel parameters, the noise satisfies , representing independent and identically distributed circularly symmetric complex Gaussian noise with mean 0 and variance 1, defines the link signal-to-noise ratio. Finally, the corresponding receive vector is output. and For use in relay decoding and receiver decoding;

[0014] When a relay node performs DF decoding, it first performs... Perform channel decoding to recover the transmitter codeword and obtain Then, the direct link is in block length With bit rate The block error rate is denoted as ,in , This represents the complex Gaussian channel capacity, and finally... This serves as the input for subsequent random binning and collaborative forwarding.

[0015] Random binning and bin index generation: First, the number of bin index bits is set during the initialization phase. and codebook Evenly divided into Each box contains non-overlapping boxes, so that each box contains Each codeword, where the bin index transmission rate is defined as: Then, when the relay decoding is successful, the actual codeword is obtained. At that time, the relay determines the sub-bin mapping based on the prior known mapping. The corresponding box is assigned an index, which is then used as relay forwarding information for transmission through the relay forwarding link.

[0016] When a relay node forwards data, it first performs channel coding on the bin index to generate the length. Forwarding code Then with power In block send Secondly, the receiving end receives the data on the relay link. , This represents the channel parameters of the relay node's forwarding link and defines the signal-to-noise ratio of the relay forwarding link. Finally, the block error rate of binning index decoding is recorded as... and output Used for the first stage of decoding at the receiving end.

[0017] At the receiving end, when decoding using cooperative side information, the signal received via relay forwarding link is first used only to obtain the bin index. Then, within the candidate bins defined by the bin index, the direct link signal is used. Decode to recover And reconstruct the bitstream The equivalent decoding rate of the second stage is The block error rate of this second stage is denoted as Capacity constraints must be met. Finally, output To proceed with deep decoding and semantic inference;

[0018] After the receiving end completes bit recovery, it first... Entropy decoding is performed to obtain Then, the network is reconstructed through observation to obtain... Secondly, the semantic inference network is used to output the posterior. Finally, the maximum a posteriori estimation rule is used to obtain... , These are all parameters of a deep neural network.

[0019] Furthermore, step S3 specifically includes:

[0020] S301, Define the distortion index;

[0021] Observation distortion is defined as mean square error. Then, for classification tasks, semantic distortion is defined as the probability of classification error. .

[0022] S302, Distortion Decomposition;

[0023] End-to-end distortion is decomposed into two parts: transmitter distortion and channel-induced distortion. ,in and By selection The depth is determined by its corresponding depth codec, and and Used to characterize the performance degradation caused by decoding errors with limited code length;

[0024] S303, Calculate the upper bound of channel distortion;

[0025] First, let's look at end-to-end BLER The channel-side error is characterized, and then the end-to-end BLER is represented as a combination of events. And introduce an approximation of the average number of blocks. Then, the sensitivity coefficient is used to upper bound the channel-induced distortion. ,in It reflects the sensitivity of a specific depth semantic codec to BLER, and is obtained through data fitting;

[0026] S304. Build the lookup table offline and call it online;

[0027] First, it is trained offline. Semantic encoding and decoding models with different compression ratios were used, and the sensitivity coefficient of each model was calibrated by regression to construct a lookup table. Then, during online runtime, selection is performed via table lookup. The corresponding parameters and sensitivity coefficients are then substituted into the end-to-end distortion model and joint optimization process to drive adaptive parameter tuning.

[0028] Furthermore, step S4 specifically includes:

[0029] S401. Confirm the optimization objective and establish constraints;

[0030] First, define weighted end-to-end distortion as... Then, a weighted sensitivity coefficient is introduced in conjunction with the upper bound of distortion. The optimization goal is confirmed as follows:

[0031] .

[0032] The constraints include discrete model selection constraints. The maximum number of times the channel can be used. Given delay constraints Total power budget constraint and the two-stage decoding capacity constraint .

[0033] S402, Two-stage solution;

[0034] First, construct a finite set of configurations. Then for each Determined by lookup table Its corresponding parameters and sensitivity coefficients are obtained from the power budget. Next, calculate the signal-to-noise ratio of the three links. Further, the capacity constant is obtained. Finally, in the fixed Under the condition only for Perform continuous optimization and obtain the The optimal target value is obtained, thus enabling... Select the globally optimal configuration above;

[0035] S403, Rate Optimization Subproblem under Fixed Configuration;

[0036] In fixed Then, the continuous subproblems are written as:

[0037] ;

[0038] S404. Solve the subproblem using the hybrid projection gradient-Newton method;

[0039] First, define the iteration variable. And select feasible initial values. With stop threshold Then in the Calculate in the next iteration The step size is obtained by searching through backtracking lines. Next, construct candidate points for the projection gradient. Further in Hessian calculation And construct the Newton direction Then, by searching through backtracking lines, we can obtain... And form projected Newton candidate points Secondly, compare and And select the better one as Finally, the gradient residual is projected. Determine whether the condition is met. If satisfied, output Otherwise, continue iterating;

[0040] S405, Calculate the projection operator;

[0041] Using sequential projection, the constraint structure is first determined. feasible range Then Truncation projection is Secondly, in the given Determine under the conditions feasible range ,Will Truncation projection is Finally, we got .

[0042] S406, Online Output and Closed-Loop Execution;

[0043] When the system is running online, firstly, for each Solve for its corresponding And calculate Then Select the globally optimal configuration Secondly, output the adaptive parameter set. Finally, this parameter set is used to drive the end-to-end transmission, cooperative forwarding, and recovery inference process, realizing a closed-loop adaptive parameter selection-transmission-recovery or inference process.

[0044] The beneficial effects of this invention are as follows:

[0045] This invention enables the transmitter to generate a bitstream through deep semantic source coding and entropy coding and interface with digital channel coding, thereby achieving compatibility with existing digital communication systems. At the same time, it introduces half-duplex DF (Decode-Forward) relay and obtains cooperative gain through random binning and two-stage continuous decoding. Furthermore, it establishes an end-to-end distortion computable model under finite code length conditions and jointly optimizes the source coding rate, channel coding rate, and power allocation to minimize end-to-end weighted distortion and meet time delay and power constraints, thereby improving semantic task performance and system robustness. Attached Figure Description

[0046] Figure 1 This is a flowchart of an adaptive relay semantic coding method based on deep learning provided by the present invention;

[0047] Figure 2 This is a block diagram of a relay semantic communication system based on a deep neural network provided by the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0049] This invention provides an adaptive relay semantic coding method based on deep learning, such as... Figure 1 As shown, it includes:

[0050] S1. Construct a relay semantic communication system based on deep neural networks;

[0051] like Figure 2 As shown, the relay semantic communication system based on deep neural networks consists of a transmitter, a relay node, and a receiver. It includes a direct link from the transmitter to the receiver, a relay receiving link from the transmitter to the relay, and a relay forwarding link from the relay to the receiver. It adopts a half-duplex DF relay system, that is, the relay operates on two orthogonal time slots or frequency bands, so that the relay's reception and decoding of the t-th block corresponds to its forwarding of the t+1-th block.

[0052] S2. At the transmitting end, a bitstream is generated through deep semantic source coding and entropy coding and then interfaced with digital channel coding. Simultaneously, a half-duplex DF relay is introduced, and cooperative gain is obtained through random binning and two-stage continuous decoding. The specific steps are as follows:

[0053] Semantic source model and task objective:

[0054] For each sample to be transmitted, construct a semantic source pair. ,in Represents observation data (e.g., images). Represents task-related semantic states (e.g., category labels). This represents the dimension of the image. The system goal is to enable the receiver to obtain [image data] through relay cooperation while satisfying time delay and power constraints. And infer .

[0055] Transmitter semantic source encoder:

[0056] Input sample pairs via transmitter When performing deep source coding, firstly... Input semantic coding network Feature extraction is performed to obtain low-dimensional continuous semantic features. ,in For neural network parameters, then for Quantization is performed to obtain discrete semantic features Secondly, for Arithmetic coding and isentropic coding are performed to generate a bitstream. Furthermore, the source coding rate is defined as Finally, output the bit stream. and the corresponding source rate This allows for subsequent channel coding and transmission processes.

[0057] Transmitter digital channel encoder:

[0058] When interfacing with a digital communication system, the transmitting end transmits the bit stream. To perform channel coding, firstly... By length Divided into blocks Then, using a block length of The channel encoder for each Encoding generates symbol sequences Secondly, the transmitter channel coding rate is defined as... Further, all codewords are combined into a codebook. Finally, output With bit rate For use in broadcast transmission.

[0059] Transmitter-side broadcast transmission and reception signal model:

[0060] When the transmitter performs broadcast transmission, it first uses power Send the Block code Then, the relay node and the receiver receive the signal on their respective links, with arbitrary symbol dimensions. The above satisfies ,in For channel parameters, the noise satisfies This represents independent, identically distributed, circularly symmetric complex Gaussian noise with a mean of 0 and a variance of 1. The link signal-to-noise ratio is then defined based on this. Finally, the corresponding receive vector is output. and It is used for relay decoding and receiver decoding.

[0061] When a relay node performs DF decoding, it first performs... Perform channel decoding to recover the transmitter codeword and obtain Then, the direct link is in block length With bit rate The block error rate is denoted as ,in , This represents the complex Gaussian channel capacity, and finally... This serves as the input for subsequent random binning and collaborative forwarding.

[0062] Random binning and binning index generation:

[0063] To achieve collaborative compression and side information forwarding, the system first sets the number of bin index bits during the initialization phase. and codebook Evenly divided into Each box contains non-overlapping boxes, so that each box contains Each codeword, where the bin index transmission rate is defined as: Then, when the relay decoding is successful, the actual codeword is obtained. At that time, the relay determines the sub-bin mapping based on the prior known mapping. The corresponding box is assigned an index, which is then used as relay forwarding information for transmission through the relay forwarding link.

[0064] Relay nodes encode and forward the bin index channel:

[0065] When a relay node forwards data, it first performs channel coding on the bin index to generate the length. Forwarding code Then with power In block send Secondly, the receiving end receives the data on the relay link. , This represents the channel parameters of the relay forwarding link, and further defines the signal-to-noise ratio of the relay forwarding link. Finally, the block error rate of binning index decoding is recorded as... and output Used for the first stage of decoding at the receiving end.

[0066] Two-stage continuous decoding at the receiving end:

[0067] When the receiving end decodes using cooperative side information, it first uses only the relay forwarding link to receive signals for decoding to obtain the bin index, and then uses the direct link signal within the candidate bins defined by the bin index. Decode to recover And reconstruct the bitstream The equivalent decoding rate of the second stage is The block error rate at this stage is denoted as To ensure reliable two-stage decoding, capacity constraints must be met. Finally, output To proceed with in-depth decoding and semantic inference.

[0068] Receiver-side deep decoding and semantic inference:

[0069] After the receiving end completes bit recovery, it first... Entropy decoding is performed to obtain Then, the network is reconstructed through observation to obtain... Secondly, the semantic inference network is used to output the posterior. Finally, the maximum a posteriori estimation rule is used to obtain... , These are all parameters of a deep neural network.

[0070] S3. Establish an end-to-end distortion calculation model under the condition of limited code length;

[0071] Distortion index definition:

[0072] To simultaneously characterize observation recovery quality and semantic task performance, this invention first defines observation distortion as mean squared error (MSE). Then, for classification tasks, semantic distortion is defined as the probability of classification error. .

[0073] Distortion decomposition:

[0074] To facilitate analysis and optimization, this invention further decomposes end-to-end distortion into two parts: transmitter distortion and channel-induced distortion. ,in and By selection The depth is determined by its corresponding depth codec, and and Used to characterize the performance degradation caused by decoding errors with limited code length.

[0075] Calculate the upper bound of channel distortion:

[0076] This invention first uses end-to-end BLER To characterize channel-side errors and to delineate the overall block error probability of the cooperative system, this invention further represents end-to-end BLER as a combination of events. And introduce an approximation of the average number of blocks. Then, the sensitivity coefficient is used to upper bound the channel-induced distortion. ,in The sensitivity of a specific depth semantic codec to BLER can be obtained through data fitting.

[0077] Lookup table Offline build and online invocation:

[0078] In order to make and the distortion at the transmitter varies with the source compression rate. Controllable, this invention first obtains the result during the offline training phase. Semantic encoding and decoding models with different compression ratios were used, and the sensitivity coefficient of each model was calibrated by regression to construct a lookup table. Then, during online runtime, selection is performed via table lookup. The corresponding parameters and sensitivity coefficients are then substituted into the end-to-end distortion model and joint optimization process to drive adaptive parameter tuning.

[0079] S4. Jointly optimize the source coding rate, channel coding rate, and power allocation to minimize end-to-end weighted distortion and satisfy time delay and power constraints.

[0080] Optimization goals and constraint construction:

[0081] To achieve online adaptive parameter tuning, this invention first defines weighted end-to-end distortion as... Then, a weighted sensitivity coefficient is introduced in conjunction with the upper bound of distortion. The optimization objective is written in a computable approximation form:

[0082] .

[0083] Secondly, the constraints include discrete model selection constraints. It further includes the maximum allowed number of channel uses. Given delay constraints Including total power budget constraints Finally, it includes a two-stage decoding capacity constraint. .

[0084] Two-stage solution framework:

[0085] To balance feasibility and online complexity, this invention employs a two-stage solution. First, a finite set of configurations is constructed. Then for each Determined by lookup table Its corresponding parameters and sensitivity coefficients are obtained from the power budget. Next, calculate the signal-to-noise ratio of the three links. Further, the capacity constant is obtained. Finally, in the fixed Under the condition only for Perform continuous optimization and obtain the The optimal target value is obtained, thus enabling... Select the globally optimal configuration.

[0086] In fixed Subsequently, this invention rewrites the continuous subproblem as follows:

[0087] ;

[0088] To solve the above subproblems, this invention employs the Hybrid Projected Gradient-Newton (HPGN) method. First, iterative variables are defined... And select feasible initial values. With stop threshold Then in the Calculate in the next iteration The step size is obtained by searching through backtracking lines. Next, construct candidate points for the projection gradient. Further in Hessian calculation And construct the Newton direction Then, by searching through backtracking lines, we can obtain... And form projected Newton candidate points Secondly, compare and And select the better one as Finally, the gradient residual is projected. Determine whether the condition is met. If satisfied, output Otherwise, continue iterating.

[0089] Due to constraints Due to strong coupling, this invention employs sequential projection. First, it determines the constraint structure... feasible range Then Truncation projection is Secondly, in the given Determine under the conditions feasible range Further Truncation projection is Finally, we got .

[0090] Online output and closed-loop execution:

[0091] When the system is running online, firstly, for each Solve for its corresponding And calculate Then Select the globally optimal configuration Secondly, output an adaptive parameter set. Finally, this parameter set is used to drive the end-to-end transmission, cooperative forwarding, and recovery inference process described in this disclosure, achieving closed-loop adaptive parameter selection, transmission, and recovery / inference.

[0092] Simulation: This invention employs the classic hyper-prior model as the encoder and decoder for data users, and the semantic information recovery network for semantic users adopts the classic ResNet architecture. The proposed method was tested on the CUB-200-2011 image dataset. The training dataset for the source coding model is the CUB-200-2011 training dataset. When creating the lookup table, this invention uses the Adam optimizer and [unclear] to train the neural network model. The batch size is 16, with a total of 200 rounds, and the initial learning rate is set to [unclear]. The loss function is decayed by a factor of 0.1 while remaining constant. The established lookup table contains 16 models, with an average compressed bit value per pixel ranging from 0.012 to 1.36.

[0093] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. An adaptive relay semantic coding method based on deep learning, characterized in that, The encoding method includes: S1. Construct a relay semantic communication system based on deep neural networks; The deep neural network-based relay semantic communication system consists of a transmitter, a relay node, and a receiver. It includes a direct link from the transmitter to the receiver, a relay receiving link from the transmitter to the relay, and a relay forwarding link from the relay to the receiver. It adopts a half-duplex DF relay system, that is, the relay operates on two orthogonal time slots or frequency bands, so that the relay's reception and decoding of the t-th block corresponds to its forwarding of the t+1-th block. S2. At the transmitting end, a bit stream is generated by deep semantic source coding and entropy coding and then connected to digital channel coding. At the same time, a half-duplex DF relay is introduced, and cooperative gain is obtained by random binning and two-stage continuous decoding. At the transmitting end, for the input sample pair When performing deep source coding, firstly... Input semantic coding network Feature extraction is performed to obtain low-dimensional continuous semantic features. ,in For neural network parameters, then for Quantization is performed to obtain discrete semantic features Secondly, for Arithmetic coding and entropy coding are performed to generate a bitstream. The source coding rate is defined as Finally, output the bit stream. and the corresponding source rate To proceed to the subsequent channel coding and transmission process; When interfacing with a digital communication system, the transmitting end transmits the bit stream. To perform channel coding, firstly... By length Divided into blocks Then, using a block length of The channel encoder for each Encoding generates symbol sequences Secondly, the transmitter channel coding rate is defined as... To compile all the codewords into a codebook Finally, output With bit rate For use in broadcast transmission; When broadcasting from the transmitting end, the power is first used as the starting point. Send the Block code Then, the relay node and the receiver receive the signal on their respective links, with arbitrary symbol dimensions. The above satisfies ,in For channel parameters, the noise satisfies , representing independent and identically distributed circularly symmetric complex Gaussian noise with mean 0 and variance 1, defines the link signal-to-noise ratio. Finally, the corresponding receive vector is output. and For use in relay decoding and receiver decoding; When a relay node performs DF decoding, it first performs... Perform channel decoding to recover the transmitter codeword and obtain Then, the direct link is in block length With bit rate The block error rate is denoted as ,in , This represents the complex Gaussian channel capacity, and finally... This serves as the input for subsequent random binning and collaborative forwarding; Random binning and bin index generation: First, the number of bin index bits is set during the initialization phase. and codebook Evenly divided into Each box contains non-overlapping boxes, so that each box contains Each codeword, where the bin index transmission rate is defined as: Then, when the relay decoding is successful, the real codeword is obtained. At that time, the relay determines the sub-bin mapping based on the prior known mapping. The corresponding box is assigned an index, which is then used as relay forwarding information for transmission through the relay forwarding link. When a relay node forwards data, it first performs channel coding on the bin index to generate the length. Forwarding code Then with power In block send Secondly, the receiving end receives the data on the relay link. , This represents the channel parameters of the relay node's forwarding link and defines the signal-to-noise ratio of the relay forwarding link. Finally, the block error rate of binning index decoding is recorded as... and output Used for the first stage of decoding at the receiving end; S3. Establish an end-to-end distortion calculation model under the condition of limited code length; S4. Jointly optimize the source coding rate, channel coding rate, and power allocation to minimize end-to-end weighted distortion and satisfy time delay and power constraints.

2. The deep learning-based adaptive relay semantic coding method according to claim 1, characterized in that, Step S2 specifically includes: At the receiving end, when decoding using cooperative side information, the signal received via relay forwarding link is first used to obtain the bin index, and then the direct link signal is used within the candidate bins defined by the bin index. Decode to recover And reconstruct the bitstream The equivalent decoding rate of the second stage is The block error rate of this second stage is denoted as Capacity constraints must be met. Finally, output To proceed with deep decoding and semantic inference; After the receiving end completes bit recovery, it first... Entropy decoding is performed to obtain Then, the network is reconstructed through observation to obtain... Secondly, the semantic inference network is used to output the posterior. Finally, the maximum a posteriori estimation rule is used to obtain... , These are all parameters of a deep neural network.

3. The deep learning-based adaptive relay semantic coding method according to claim 2, characterized in that, Step S3 specifically includes: S301, Define the distortion index; Observation distortion is defined as mean square error. Then, for classification tasks, semantic distortion is defined as the probability of classification error. ; S302, Distortion Decomposition; End-to-end distortion is decomposed into two parts: transmitter distortion and channel-induced distortion. ,in and By selection The depth is determined by its corresponding depth codec, and and Used to characterize the performance degradation caused by decoding errors with limited code length; S303, Calculate the upper bound of channel distortion; First, let's look at end-to-end BLER The channel-side error is characterized, and then the end-to-end BLER is represented as a combination of events. And introduce an approximation of the average number of blocks. Then, the sensitivity coefficient is used to upper bound the channel-induced distortion. ,in It reflects the sensitivity of a specific depth semantic codec to BLER, and is obtained through data fitting; S304. Build the lookup table offline and call it online; First, it is trained offline. Semantic encoding and decoding models with different compression rates were used, and the sensitivity coefficient of each model was calibrated by regression to construct a lookup table. Then, during online runtime, selection is made by looking up a table. The corresponding parameters and sensitivity coefficients are then substituted into the end-to-end distortion model and joint optimization process to drive adaptive parameter tuning.

4. The deep learning-based adaptive relay semantic coding method according to claim 3, characterized in that, Step S4 specifically includes: S401. Confirm the optimization objective and establish constraints; First, define weighted end-to-end distortion as... Then, a weighted sensitivity coefficient is introduced in conjunction with the upper bound of distortion. The optimization goal is confirmed as follows: ； The constraints include discrete model selection constraints. The maximum number of times the channel can be used. Given delay constraints Total power budget constraint and the two-stage decoding capacity constraint ; S402, Two-stage solution; First, construct a finite set of configurations. Then for each Determined by lookup table Its corresponding parameters and sensitivity coefficients are obtained from the power budget. Next, calculate the signal-to-noise ratio of the three links. To obtain the capacity constant Finally, in the fixed Under the condition only for Perform continuous optimization and obtain the The optimal target value is obtained, thus enabling... Select the globally optimal configuration above; S403, Rate Optimization Subproblem under Fixed Configuration; In fixed Then, the continuous subproblems are written as: ； S404. Solve the subproblem using the hybrid projection gradient-Newton method; First, define the iteration variable. And select feasible initial values. With stop threshold Then in the Calculate in the next iteration The step size is obtained by searching through backtracking lines. Secondly, construct candidate points for the projection gradient. Further in Hessian calculation And construct the Newton direction Then, by searching through backtracking lines, we can obtain... And form projected Newton candidate points Secondly, compare and And select the better one as Finally, the gradient residual is projected. Determine whether the condition is met. If satisfied, output Otherwise, continue iterating; S405, Calculate the projection operator; Using sequential projection, the constraint structure is first determined. feasible range Then Truncation projection is Secondly, in the given Determine under the conditions feasible range ,Will Truncation projection is Finally, we got ; S406, Online Output and Closed-Loop Execution; When the system is running online, firstly, for each Solve for its corresponding And calculate Then Select the globally optimal configuration Secondly, output an adaptive parameter set. Finally, this parameter set is used to drive the end-to-end transmission, cooperative forwarding, and recovery inference process, realizing a closed-loop adaptive parameter selection-transmission-recovery or inference process.

Images