Edge - end Collaborative Semantic Communication Method and System for Distribution Network

Abstract

An embodiment of the present invention provides a method and system for edge-end collaborative semantic communication for a distribution network, belonging to the field of communication technologies. The method is executed by a terminal-side device, and the method includes: periodically constructing a corresponding image test sample set; sequentially performing semantic encoding and channel encoding on the image test sample set; enabling the edge server to sequentially perform channel decoding and semantic decoding on the transmission sample sets of each terminal-side device received; performing multiple rounds of transmission of the image test sample set with the edge server, the edge server performing iterative operations on each model parameter and establishing a communication connection with each terminal-side device based on each output model parameter; and periodically sending the collected actual image data to the corresponding edge server based on the established communication connection. The solution of the present invention not only improves the accuracy of data transmission but also enhances the anti-interference ability of the system against channel interference based on improved semantic communication in the distribution network.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and particularly to an edge-terminal collaborative semantic communication method for a distribution network and an edge-terminal collaborative semantic communication system for a distribution network. Background Art

[0002] As a key link in power transmission, the stable and efficient operation of the distribution network is directly related to the guarantee of power supply. In modern distribution networks, with the intelligent development of the power system, the number and types of terminal devices are continuously increasing. These terminal devices are widely deployed in various environments to collect a large amount of data such as voltage, current, frequency, device status, fault conditions, and safety conditions. The collection and transmission of this data are of great significance for realizing the real-time monitoring, control, and optimization of the distribution network. However, the existing data transmission solutions face many technical problems in practical applications, seriously affecting the overall performance and reliability of the system.

[0003] With the increase in the number of terminal devices, the amount of data grows exponentially. Traditional data transmission methods need to transmit this data to the edge server for processing through wireless communication. This centralized data processing mode brings a huge communication burden, resulting in congestion of the transmission channel and an increase in transmission delay. In addition, the centralized processing of data also causes the edge server to bear a huge computing pressure, easily forming a processing bottleneck and reducing the response speed of the system. In a complex distribution environment, data transmission faces multiple threats such as electromagnetic interference, network delay, channel interference, and malicious attacks. These factors not only lead to an increase in the bit error rate during data transmission but also may cause packet loss and transmission errors, ultimately affecting the accurate recovery of the original data by the edge server. This difficulty in information recovery directly affects the accuracy of key applications such as fault monitoring and status monitoring, may lead to a lag in fault detection, and even pose a hidden danger to the safe operation of the power system.

[0004] The existing data transmission solutions lack sufficient flexibility and robustness when dealing with dynamic distribution environments. Facing uncertain channel conditions and network situations, traditional transmission methods are difficult to adaptively adjust the transmission strategy, further exacerbating the problems of data transmission delay and reliability. Generally speaking, the technical problems existing in the existing solutions in large-scale distribution networks include: heavy communication burden, large transmission delay, obvious processing bottleneck, and high bit error rate and difficulty in data recovery in complex environments. These problems seriously restrict the real-time performance and accuracy of distribution network monitoring and control, and there is an urgent need for an improved solution that can effectively reduce the communication burden, improve data transmission efficiency, and anti-interference ability. Summary of the Invention

[0005] The objective of the embodiments of the present invention is to provide a side-edge collaborative semantic communication method and system for a distribution network, so as to at least solve the problems of heavy communication burden and large transmission delay existing in the existing side-edge collaborative communication of the distribution network.

[0006] To achieve the above objective, the first aspect of the present invention provides a side-edge collaborative semantic communication method for a distribution network, which is applied to the semantic communication between an edge server and terminal-side devices. The edge server is communicatively connected to a plurality of terminal-side devices. The method is executed by the terminal-side devices and includes: periodically collecting image data at a set position and periodically constructing a corresponding image test sample set; sequentially performing semantic encoding and channel encoding on the image test sample set to obtain a transmission sample set, and sending the transmission sample set to the corresponding edge server; enabling the edge server to sequentially perform channel decoding and semantic decoding on the received transmission sample sets of each terminal-side device, correspondingly obtaining the decoded image sample sets of each terminal-side device, and calculating a transmission loss value based on the comparison relationship between each decoded image sample set and the corresponding image test sample set; taking the minimum transmission loss value as the iteration objective, performing multiple rounds of transmission of the image test sample sets with the edge server, and the edge server performing iterations on the semantic encoding model parameters, channel encoding model parameters, channel decoding model parameters, and semantic decoding model parameters based on the stochastic gradient descent algorithm, and establishing a communication connection with each terminal-side device based on the corresponding output model parameters; periodically sending the collected actual image data to the corresponding edge server based on the established communication connection.

[0007] Optionally, the encoding rule for performing semantic encoding on the image test sample set is:

[0008]

[0009] wherein, is the semantic encoder; α k is the semantic encoding model parameter of the kth terminal-side device; I is the image test sample set; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; E I is the semantic feature after semantic encoding.

[0010] Optionally, the encoding rule for performing channel encoding on the image test sample set:

[0011]

[0012] wherein, is the channel encoder; β k is the channel encoding model parameter of the kth terminal-side device; I is the image test sample set; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; X IIt is the semantic feature after channel coding.

[0013] Optionally, the transmission rule for sending the transmission sample set to the corresponding edge server is: based on an additive white Gaussian noise channel, using the corresponding edge server as the receiving end to perform the transmission of the transmission sample set, and the transmission rule is:

[0014]

[0015] where is the semantic feature after being interfered; X I is the semantic feature after channel coding; N is the noise; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively.

[0016] Optionally, when the edge server decodes the transmission sample sets received from each terminal-side device, it is configured to:

[0017] Perform channel decoding and semantic decoding on the transmission sample sets received from each terminal-side device in sequence to obtain the corresponding decoded image sample sets; among them, the channel decoding rule is: The semantic decoding rule is: where represents the channel encoder; χ represents the channel decoding model parameters deployed on the edge server; represents the channel encoder; δ represents the semantic decoding model parameters deployed on the edge server; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; Y I is the semantic feature after channel decoding; is the decoded image sample set.

[0018] Optionally, when performing the first round of iteration, the edge server is configured to: perform parameter aggregation based on the semantic coding model parameters and channel coding model parameters of each terminal-side device to obtain global model parameters; based on the global model parameters, complete one decoding of the transmission sample set, and use the corresponding channel decoding model parameters and semantic decoding model parameters as the initial channel decoding model parameters and initial semantic decoding model parameters for the iteration.

[0019] Optionally, the edge server performs parameter aggregation based on the semantic coding model parameters and channel coding model parameters of each terminal-side device to obtain global model parameters, including: the edge server respectively compares the transmission loss values corresponding to each terminal-side device, performs weight allocation for each terminal-side device based on the transmission loss values corresponding to each terminal-side device, and the smaller the transmission loss value, the greater the corresponding weight; based on the weight allocation result, perform parameter aggregation on the semantic coding model parameters and channel coding model parameters of each terminal-side device to obtain global model parameters, expressed as:

[0020]

[0021]

[0022] Among them, α and β are global model parameters; K is the number of terminal-side devices; ω k is the allocation weight of the k-th terminal-side device; α k is the semantic coding model parameter of the k-th terminal-side device; β k is the channel coding model parameter of the k-th terminal-side device; N k is the channel noise of the wireless communication between the k-th terminal-side device and the edge server.

[0023] Optionally, in each round of iteration, the edge server is configured to: respectively send the corresponding semantic coding model parameter and channel coding model parameter, and update the semantic coding model parameter and channel coding model parameter of each terminal-side device based on the transmission loss value after this round of iteration, as the initial semantic coding model parameter and initial channel coding model parameter of the corresponding terminal-side device in the next iteration round; perform sequential learning on each terminal-side device based on the semantic coding model parameter and channel coding model parameter of each terminal-side device, and update the channel decoding model parameter and semantic decoding model parameter once every time the learning of a terminal-side device is completed, until this round of iteration is completed, and use the channel decoding model parameter and semantic decoding model parameter updated after the learning of the last terminal-side device as the initial channel decoding model parameter and initial semantic decoding model parameter of the edge server in the next iteration round.

[0024] Optionally, the update rule for the edge server to update the initial coding model parameter and initial channel coding model parameter of each terminal-side device in the next iteration round is:

[0025]

[0026]

[0027] Among them, is the initial semantic coding model parameter of the k-th terminal-side device in the next iteration round; is the initial channel coding model parameter of the k-th terminal-side device in the next iteration round; is the initial semantic coding model parameter of the k-th terminal-side device in this round; is the initial channel coding model parameter of the k-th terminal-side device in this round; η is the learning rate; is the gradient of the transmission loss value corresponding to the k-th terminal-side device in this round.

[0028] Optionally, the edge server performs sequential learning on each terminal device based on the semantic coding model parameters and channel coding model parameters of each terminal device. For each terminal device learned, the update rules for updating the channel decoding model parameters and semantic decoding model parameters are as follows:

[0029]

[0030]

[0031] Among them, is the channel decoding model parameter corresponding to the edge server when performing channel decoding on the k-th terminal device; is the semantic decoding model parameter corresponding to the edge server when performing semantic decoding on the k-th terminal device; is the channel decoding model parameter corresponding to the edge server when performing channel decoding on the (k - 1)-th terminal device; is the semantic decoding model parameter corresponding to the edge server when performing semantic decoding on the (k - 1)-th terminal device; η is the learning rate; is the gradient of the transmission loss value corresponding to the (k - 1)-th terminal device.

[0032] Optionally, the rules for the edge server to use the channel decoding model parameters and semantic decoding model parameters updated after learning the last terminal device as the initial channel decoding model parameters and initial semantic decoding model parameters of the edge server in the next iteration round are as follows:

[0033]

[0034]

[0035] Among them, χ (t+1) is the initial channel decoding model parameter of the edge server in the next iteration round; δ (t+1) is the initial semantic decoding model parameter of the edge server in the next iteration round; is the signal decoding model parameter for the edge server to perform channel decoding on the last terminal device in this round; is the semantic decoding model parameter for the edge server to perform semantic decoding on the last terminal device in this round; η is the learning rate; is the gradient of the transmission loss value corresponding to the last terminal device in this round.

[0036] The second aspect of the present invention provides a side - end collaborative semantic communication method for a distribution network, which is applied to semantic communication between an edge server and terminal - side devices. The edge server is communicatively connected to multiple terminal - side devices. The method is executed by the edge server and includes: periodically receiving the transmission sample sets uploaded by each terminal - side device; wherein, the transmission sample sets are obtained by each terminal - side device based on the following rules: periodically collecting image data at a set position and periodically constructing a corresponding image test sample set; sequentially performing semantic encoding and channel encoding on the image test sample set to obtain a transmission sample set; sequentially performing channel decoding and semantic decoding on the received transmission sample sets of each terminal - side device to correspondingly obtain the decoded image sample sets of each terminal - side device, and calculating the transmission loss value based on the comparison relationship between each decoded image sample set and the corresponding image test sample set; taking the minimum transmission loss value as the iteration target, performing multiple rounds of transmission of the image test sample sets with each terminal - side device, performing iteration of the semantic encoding model parameters, channel encoding model parameters, channel decoding model parameters, and semantic decoding model parameters based on the stochastic gradient descent algorithm, and establishing communication connections with each edge - side device based on the correspondingly output model parameters; periodically recovering the actual image data collected by each edge - side device based on the established communication connections.

[0037] Optionally, in each iteration process, the method includes: respectively sending the corresponding semantic encoding model parameters and channel encoding model parameters, and updating the semantic encoding model parameters and channel encoding model parameters of each terminal - side device based on the transmission loss value after this iteration as the initial semantic encoding model parameters and initial channel encoding model parameters of the corresponding terminal - side device in the next iteration round; performing sequential learning on each terminal - side device based on the semantic encoding model parameters and channel encoding model parameters of each terminal - side device, and updating the channel decoding model parameters and semantic decoding model parameters once every time the learning of a terminal - side device is completed until this iteration is completed, and taking the updated channel decoding model parameters and semantic decoding model parameters after the learning of the last terminal - side device as the initial channel decoding model parameters and initial semantic decoding model parameters of the edge server in the next iteration round.

[0038] Optionally, the update rule for the edge server to update the initial encoding model parameters and initial channel encoding model parameters of each terminal - side device in the next iteration round is:

[0039]

[0040]

[0041] wherein, is the initial semantic encoding model parameter of the k - th terminal - side device in the next iteration round; is the initial channel coding model parameter for the next iteration of the k-th terminal device; is the initial semantic coding model parameter of the k-th terminal device in this round; is the initial channel coding model parameter of the k-th terminal device in this round; η is the learning rate; is the gradient of the transmission loss value corresponding to the k-th terminal device in this round.

[0042] Optionally, the edge server performs sequential learning on each terminal device based on the semantic coding model parameters and channel coding model parameters of each terminal device. For each terminal device's learning completion, the update rules for updating the channel decoding model parameters and semantic decoding model parameters are as follows:

[0043]

[0044]

[0045] Among them, is the channel decoding model parameter corresponding to the edge server when performing channel decoding on the k-th terminal device; is the semantic decoding model parameter corresponding to the edge server when performing semantic decoding on the k-th terminal device; is the channel decoding model parameter corresponding to the edge server when performing channel decoding on the (k - 1)-th terminal device; is the semantic decoding model parameter corresponding to the edge server when performing semantic decoding on the (k - 1)-th terminal device; η is the learning rate; is the gradient of the transmission loss value corresponding to the (k - 1)-th terminal device.

[0046] Optionally, the rules for the edge server to use the updated channel decoding model parameters and semantic decoding model parameters after the learning of the last terminal device as the initial channel decoding model parameters and initial semantic decoding model parameters for the next iteration are as follows:

[0047]

[0048]

[0049] Among them, χ (t+1) is the initial channel decoding model parameter of the edge server for the next iteration; δ (t+1) is the initial semantic decoding model parameter of the edge server for the next iteration; is the signal decoding model parameter for the edge server to perform channel decoding on the last terminal device in this round; is the semantic decoding model parameter for the edge server to perform semantic decoding on the last terminal device in this round; η is the learning rate; is the gradient of the transmission loss value corresponding to the last terminal device in this round.

[0050] The third aspect of the present invention provides an edge-end collaborative semantic communication system for a distribution network, which is applied to semantic communication between an edge server and terminal devices. The edge server is communicatively connected to multiple terminal devices. The system includes: a terminal device for executing the above-mentioned edge-end collaborative semantic communication method for a distribution network, and an edge server for executing the above-mentioned edge-end collaborative semantic communication method for a distribution network.

[0051] The fourth aspect of the present invention provides a computer-readable storage medium. Instructions are stored on the computer-readable storage medium, and when running on a computer, the computer executes the above-mentioned edge-end collaborative semantic communication method for a distribution network.

[0052] The fifth aspect of the present invention provides a computer program product, including a computer program, and the computer program implements the above-mentioned edge-end collaborative semantic communication method when executed by a processor.

[0053] Through the above technical solutions, the edge-end collaborative semantic communication method proposed by the present invention significantly improves the efficiency and reliability of distribution network data transmission by applying semantic communication technology between an edge server and multiple terminal devices. First, the terminal device periodically collects image data and performs semantic encoding and channel encoding locally, greatly reducing the data dimension and thus reducing the communication burden of data transmission. Subsequently, the encoded data is sent to the edge server. The edge server performs channel decoding and semantic decoding on the received sample set, restores the original image data, and calculates the transmission loss value. Through multiple rounds of iterative optimization, the edge server uses the stochastic gradient descent algorithm to continuously adjust and optimize the model parameters of semantic encoding, channel encoding, channel decoding, and semantic decoding, enabling the system to adaptively reduce the transmission loss value. This collaborative optimization process not only improves the accuracy of data transmission but also enhances the anti-interference ability of the system to channel interference.

[0054] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent specific embodiment part. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the embodiments of the present invention, but do not constitute a limitation to the embodiments of the present invention. In the drawings:

[0056] Figure 1 is a flowchart of the steps of the edge-end collaborative semantic communication method for a distribution network provided by an embodiment of the present invention;

[0057] Figure 2 It is the system structure diagram of the edge - end collaborative semantic communication system for the distribution network provided by an embodiment of the present invention. Specific embodiments

[0058] The following will detail the specific embodiments of the present invention with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not used to limit the present invention.

[0059] The core of the semantic communication method lies in extracting the semantic information in the original information through the deployed semantic extraction model at the sending end. After transmission through the wireless channel, the receiving end uses the semantic restoration model to restore it. This method can not only retain the core meaning of the information, but also significantly reduce the data volume and maintain high transmission performance in an environment with low signal - to - noise ratio. In addition, the receiving end can directly apply the received semantic information to specific tasks, such as fault detection, etc., improving the efficiency and adaptability of the communication system.

[0060] In recent years, significant progress has been made in the research of semantic communication. For example, the end - to - end semantic communication system proposed by Bourtsoulatze et al. in 2019 demonstrated that semantic communication based on deep learning can improve the transmission performance of image data under low signal - to - noise ratio conditions, overcoming the "cliff effect" in traditional digital communication. Xie et al. further utilized the Transformer encoder - decoder in 2021 to achieve an end - to - end semantic communication system for text data transmission, showing the great potential of deep learning in semantic communication. In addition, the semantic communication system for image transmission based on neural networks proposed by Zhang et al. in 2022 can adapt to dynamic data sets under the conditions that the tasks and channel environments at the sending end are unknown, while maintaining high performance in data recovery and task execution.

[0061] However, although semantic communication shows great potential in reducing data volume and anti - interference performance, existing research mainly focuses on single - task or centralized scenarios. In the actual application of the distribution network, terminal sensing devices are usually distributedly deployed at various locations, collecting different types of data and sending them to the central server through wireless transmission. This distributed scenario poses new challenges to semantic communication, and existing methods are difficult to effectively support edge - end collaborative semantic communication. Therefore, developing an edge - end collaborative semantic communication method for the distribution network applicable to the distribution network scenario has become an urgent problem to be solved.

[0062] Figure 1 It is the method flow chart of the edge - end collaborative semantic communication method for the distribution network provided by an embodiment of the present invention. As Figure 1As shown in the figure, an edge - end collaborative semantic communication method for a distribution network provided by an embodiment of the present invention includes the following steps:

[0063] Step S10: Each terminal - side device periodically collects image data at the set position and periodically constructs a corresponding image test sample set.

[0064] Specifically, multiple terminal - side devices are distributed at different physical locations and are responsible for periodically collecting image data at their respective locations. These image data are usually used to monitor and evaluate the operating status of the distribution network, such as the working conditions of equipment, environmental changes, and possible fault signs. To ensure the timeliness and accuracy of the data, the terminal devices will periodically construct corresponding image test sample sets. These sample sets not only contain the image data of the current environment but also the historical data collected over a period of time, which are used for further pattern recognition and predictive analysis. After the collection of the image data, the terminal devices will pre - process these data. This step includes denoising the original images, removing redundant information, and standardizing the format to ensure that the data occupies the minimum bandwidth resources during transmission.

[0065] In a possible implementation manner, the construction of the image test sample set not only involves simple data collection but also the pre - processing and screening of the data. The terminal device will perform preliminary processing on the collected original image data, such as removing noise, enhancing image clarity, and converting the image format for subsequent transmission and processing. These operations can significantly reduce the redundant information in the data, ensuring that the key information that best reflects the current environment is retained when constructing the image test sample set.

[0066] Step S20: Each terminal - side device sequentially performs semantic encoding and channel encoding on the image test sample set to obtain a transmission sample set, and sends the transmission sample set to the corresponding edge server.

[0067] Specifically, the encoding rule for performing semantic encoding on the image test sample set is as follows:

[0068]

[0069] Where, is the semantic encoder; α k is the semantic encoding model parameter of the k - th terminal - side device; I is the image test sample set; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; E I is the semantic feature after semantic encoding.

[0070] Furthermore, the encoding rule for performing channel encoding on the image test sample set is:

[0071]

[0072] Among them, is the channel encoder; β k is the channel coding model parameter of the k-th terminal-side device; I is the image test sample set; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; X I is the semantic feature after channel coding.

[0073] In the embodiments of the present invention, the collected image data is processed by the semantic encoder of the terminal device. The semantic encoder extracts features from the image data through a deep learning model, generating the semantic features of the image. The process of semantic coding greatly reduces the data dimension, only retaining the semantic features that can represent the core information of the image, making the data more concise and efficient during transmission. In addition, the application of the deep learning model enhances the accuracy and generalization ability of feature extraction, enabling the extracted semantic features to better adapt to different image data.

[0074] Furthermore, the semantic features are further channel-coded. The channel encoder is based on a convolutional neural network architecture, further mapping the semantic features to low-dimensional features, thereby reducing the bandwidth required for data transmission and enhancing the anti-noise interference ability of the data during transmission. Through channel coding, the data is further compressed and has stronger anti-interference ability, providing guarantee for subsequent data transmission.

[0075] Based on the solution of the present invention, through semantic coding and channel coding on the terminal device side, the data has been greatly compressed before transmission, retaining the key semantic information. This enables more effective information to be transmitted under limited bandwidth conditions, reducing the communication burden. The application of channel coding not only reduces the data dimension but also increases the anti-interference ability of the data during transmission. Even in a complex communication environment (such as the existence of electromagnetic interference, signal attenuation, etc.), the data can still maintain high integrity and accuracy. The introduction of the deep learning model makes the semantic coding process more flexible and can adapt to different image types and features. Whether it is the complexity of the image or the detailed information it contains, the encoder can effectively extract the features useful for the task, thereby enhancing the adaptability of the system.

[0076] Furthermore, the transmission rule for sending the transmission sample set to the corresponding edge server is:

[0077] Based on an additive white Gaussian noise channel, using the corresponding edge server as the receiving end to perform the transmission of the transmission sample set, the transmission rule is:

[0078]

[0079] Among them, is the semantic feature after being interfered; X I is the semantic feature after channel coding; N is the noise; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively.

[0080] In the embodiments of the present invention, considering that wireless communication is usually adopted between the terminal device and the edge server, an Additive White Gaussian Noise (AWGN) channel is selected as the channel model. The AWGN channel model is a commonly used basic channel model in wireless communication. It simulates the interference situation in the wireless channel through additive noise, and its characteristics are that the noise is uniformly distributed across the entire spectrum and has zero mean and constant power spectral density. In this model, the edge server acts as the receiving end and receives the semantic feature signals from the terminal device. These signals will inevitably be interfered by noise during transmission, resulting in signal distortion or distortion. The introduction of this noise model enables the system to consider the interference situation in the actual communication environment when designing and optimizing the channel decoder, so that the decoder can better process the received signals affected by noise. The use of the AWGN channel model, although an idealized model, can effectively simulate the basic interference characteristics in the wireless channel and is an important basis for designing a highly robust decoding algorithm.

[0081] Step S30: The edge server sequentially performs channel decoding and semantic decoding on the received transmission sample sets of each terminal device, obtains the decoded image sample sets of each terminal device correspondingly, and calculates the transmission loss value based on the comparison relationship between each decoded image sample set and the corresponding image test sample set.

[0082] Specifically, sequentially perform channel decoding and semantic decoding on the received transmission sample sets of each terminal device to obtain the corresponding decoded image sample sets; among them, the channel decoding rule is: The semantic decoding rule is: Among them, represents the channel encoder; χ represents the channel decoding model parameters deployed on the edge server; represents the channel encoder; δ represents the semantic decoding model parameters deployed on the edge server; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; Y I is the semantic feature after channel decoding; is the decoded image sample set.

[0083] In the embodiments of the present invention, after the edge server receives the transmission sample sets from each terminal-side device, it will perform channel decoding and semantic decoding in sequence to obtain the corresponding decoded image sample sets. The edge server performs channel decoding on the received transmission samples. The channel decoder is used to eliminate the interference introduced during data transmission due to reasons such as noise and channel fading, and restore the encoded semantic features to a signal close to the original transmission features. The success of channel decoding directly affects the accuracy of subsequent semantic decoding. Therefore, the design of the channel decoder must have a strong anti-interference ability to restore the original semantic features to the greatest extent.

[0084] Furthermore, the edge server performs semantic decoding on the decoded signal. The semantic decoder uses the model parameters deployed on the edge server to convert the features after channel decoding into the final decoded image sample sets. The process of semantic decoding is actually to reconstruct the semantic information in the data to restore it into image data similar to the original image samples. This step not only requires the decoder to have strong feature extraction and restoration capabilities, but also needs to be trained to ensure that the decoded image sample sets are highly consistent with the original data in terms of visual effects and semantic expressions.

[0085] Based on the solution of the present invention, through channel decoding, the edge server can effectively eliminate the noise and interference during transmission and ensure the integrity of the data. The efficiency of channel decoding directly determines whether the transmitted data can be accurately restored, thus providing a solid foundation for subsequent semantic decoding. When the semantic decoder processes the decoded data, it can highly accurately reconstruct the semantic features of the original image according to the model parameters. This efficient semantic restoration ensures that the finally decoded image sample sets not only retain the original information, but also have good visual effects and semantic consistency. The dual processing of channel decoding and semantic decoding not only enhances the system's adaptability to noise and signal fading, but also improves the robustness and reliability of data transmission. This dual decoding strategy enables the system to still maintain a high decoding accuracy and data recovery ability in a complex communication environment.

[0086] Furthermore, after completing channel decoding and semantic decoding, the edge server will obtain a set of decoded image samples. To evaluate the fidelity of information during data transmission, the system needs to compare the decoded image sample set with the original image test sample set to calculate the transmission loss value. The calculation of the transmission loss value usually adopts a loss function that measures the difference between images, such as the Mean Squared Error (MSE) or the Structural Similarity Index (SSIM). These loss functions can accurately capture the subtle differences between the decoded image and the original image at the pixel level or the structural level. The smaller the calculated loss value, the closer the decoded image is to the original image, indicating a higher quality of the data transmission and decoding process; on the contrary, the larger the loss value, the more information loss or distortion occurs during transmission, and further optimization of the encoding and decoding processes is required.

[0087] Furthermore, by regularly calculating the transmission loss value, the edge server can monitor the transmission quality in real time and adjust the model parameters of the channel encoder, channel decoder, semantic encoder, and semantic decoder according to the change of the loss value. This feedback-based optimization mechanism can significantly improve the adaptive ability of the system, enabling the system to continuously maintain high-efficient data transmission and recovery performance in a dynamically changing communication environment.

[0088] Based on the solution of the present invention, by comparing the decoded image with the original image, the transmission loss value provides an objective measure for the data recovery accuracy. This quantified index enables the system to accurately identify problems in the transmission process and make targeted improvements, thereby improving the data restoration accuracy of the entire system. Based on the feedback mechanism of the transmission loss value, the system can continuously adjust the encoding and decoding parameters for iterative optimization. This dynamic adjustment process enables the system to quickly adapt to different communication conditions and environmental changes, thus maintaining stable performance in various complex scenarios. The transmission loss value not only helps the system identify errors and noise effects in the transmission process but also promotes the adaptability and recovery ability of the system in a noisy and interference environment. This process enhances the robustness of the system, enabling the system to provide higher reliability in practical applications.

[0089] Step S30: The edge server takes the minimum transmission loss value as the iteration target, performs multiple rounds of transmission of the image test sample set with the terminal-side device, the edge server performs iterative operations on the semantic encoding model parameters, channel encoding model parameters, channel decoding model parameters, and semantic decoding model parameters based on the stochastic gradient descent algorithm, and establishes communication connections with each terminal-side device based on the corresponding output model parameters.

[0090] Specifically, when performing the first round of iteration, the edge server is configured to: perform parameter aggregation based on the semantic coding model parameters and channel coding model parameters of each terminal-side device to obtain global model parameters; based on the global model parameters, complete the decoding of a transmission sample set, and use the corresponding channel decoding model parameters and semantic decoding model parameters as the initial channel decoding model parameters and initial semantic decoding model parameters for the iteration.

[0091] Further, the edge server performs parameter aggregation based on the semantic coding model parameters and channel coding model parameters of each terminal-side device to obtain global model parameters, including:

[0092] The edge server respectively compares the transmission loss values corresponding to each terminal-side device, performs weight allocation for each terminal-side device based on the transmission loss values corresponding to each terminal-side device, and the smaller the transmission loss value, the greater the corresponding weight; based on the weight allocation result, performs parameter aggregation on the semantic coding model parameters and channel coding model parameters of each terminal-side device to obtain global model parameters, expressed as:

[0093]

[0094]

[0095] where α and β are global model parameters; K is the number of terminal-side devices; ω k is the allocated weight of the kth terminal-side device; α k is the semantic coding model parameter of the kth terminal-side device; β k is the channel coding model parameter of the kth terminal-side device; N k is the channel noise of the wireless communication between the kth terminal-side device and the edge server.

[0096] In the embodiments of the present invention, the edge server will receive the model parameters of the semantic encoders and channel encoders of multiple terminal-side devices. Each terminal device independently trains and adjusts these parameters according to its own environment and data characteristics. Since the environmental complexities and data characteristics of different terminal devices are different, their model parameters may vary. Therefore, the task of the edge server is to aggregate these model parameters to generate a set of general model parameters that can adapt to the global situation.

[0097] Furthermore, during the aggregation process, the edge server will consider the transmission loss value (loss) of each terminal device and assign corresponding weights based on this. The calculation of the weights takes into account the degree of contribution of the model performance of each terminal device. If the model of a certain terminal device performs excellently in image restoration and has a low loss value, it indicates that the model parameters of this device have high representativeness and should be given a higher weight; otherwise, its weight is reduced. The calculation process of the weights includes two main parts. The first part is to normalize the loss values of all terminal devices to ensure that the sum of all weights is 1; the second part ensures a linear relationship between the weights and the loss values. The specific calculation formula is:

[0098]

[0099] Through this way of weight assignment, the system can more accurately reflect the model contributions of each terminal device, making the aggregated global model parameters more adaptable to the characteristics of the global data.

[0100] Furthermore, after the aggregation of the global model parameters is completed, the edge server will perform the first decoding of the transmission sample set based on these parameters. After the decoding is completed, the channel decoding model parameters and the semantic decoding model parameters will be initialized and used for subsequent decoding processes. This process ensures that in subsequent decoding processes, the system can start from an optimized initial state and reduce the number of iteration rounds.

[0101] Furthermore, during each iteration process, the edge server is configured to: respectively send the corresponding semantic coding model parameters and channel coding model parameters, and update the semantic coding model parameters and channel coding model parameters of each terminal-side device based on the transmission loss value after this round of iteration, as the initial semantic coding model parameters and initial channel coding model parameters of the corresponding terminal-side device in the next iteration round; perform sequential learning on each terminal-side device based on the semantic coding model parameters and channel coding model parameters of each terminal-side device. After each terminal-side device's learning is completed, update the channel decoding model parameters and semantic decoding model parameters once until this round of iteration is completed, and use the updated channel decoding model parameters and semantic decoding model parameters after the last terminal-side device's learning as the initial channel decoding model parameters and initial semantic decoding model parameters of the edge server in the next iteration round.

[0102] In the embodiments of the present invention, at the beginning of each round of iteration, the edge server sends the semantic coding model parameters and channel coding model parameters aggregated in the previous round to each terminal device. After receiving these parameters, the terminal device uses them as initial parameters and combines its own local data and environmental characteristics for further learning and adjustment. This method ensures that each terminal device can train under more optimized initial conditions in the new round of iteration, thereby improving the convergence speed and model accuracy of the entire system. During each iteration process, the edge server updates the semantic coding model parameters and channel coding model parameters of each terminal device in real time based on the transmission loss value of the current round. These updated parameters are not only used for the optimization of this round of iteration but also serve as the initial conditions for the next round of iteration, gradually approaching the optimal solution. The edge server performs sequential learning on each terminal device, that is, processes the data of each terminal device in sequence and updates the channel decoding model parameters and semantic decoding model parameters. After each terminal device's learning is completed, the edge server will update the model parameters once. Such sequential learning ensures that the learning results of each terminal device can be gradually accumulated and optimized. Finally, the updated result after the last terminal device's learning will be used as the initial parameters for the next round of iteration, enabling the system to start from a more optimal state in each round and continuously improve the system performance.

[0103] Furthermore, the update rules for the edge server to update the initial coding model parameters and initial channel coding model parameters of each terminal device in the next iteration round are as follows:

[0104]

[0105]

[0106] Among them, is the initial semantic coding model parameter of the k-th terminal device in the next iteration round; is the initial channel coding model parameter of the k-th terminal device in the next iteration round; is the initial semantic coding model parameter of the k-th terminal device in this round; is the initial channel coding model parameter of the k-th terminal device in this round; η is the learning rate; is the gradient of the transmission loss value corresponding to the k-th terminal device in this round. and The calculation rules of are as follows:

[0107]

[0108]

[0109] Among them, is the channel noise of the downlink wireless link.

[0110] In the embodiments of the present invention, the edge server first calculates the transmission loss value by comparing the decoded image sample with the original image sample, and then adjusts the model parameters according to the loss value. Specifically, the loss value reflects the magnitude of the error introduced by the current model during the transmission process. Therefore, by optimizing the loss value, the model can gradually converge to a better state. After each round of training, the updated semantic coding model parameters and channel coding model parameters will be used as the initial parameters for the next round of training. This gradually updated mechanism enables the system to continuously reduce the transmission error and improve the accuracy of data transmission in multiple rounds of iteration.

[0111] Furthermore, the edge server will also perform sequential learning on each terminal device in each round of iteration. After completing the learning of one terminal device, the edge server will update its channel decoding model parameters and semantic decoding model parameters, and use these updated parameters as the initial parameters to continue training the next terminal device. This sequential learning method ensures that the training results of each terminal device can be gradually accumulated, ultimately forming a more accurate global model.

[0112] Furthermore, the edge server performs sequential learning on each terminal device based on the semantic coding model parameters and channel coding model parameters of each terminal device. After each terminal device's learning is completed, the update rule for performing one update of the channel decoding model parameters and semantic decoding model parameters is as follows:

[0113]

[0114]

[0115] Among them, is the channel decoding model parameter corresponding to the edge server when performing channel decoding on the kth terminal device; is the semantic decoding model parameter corresponding to the edge server when performing semantic decoding on the kth terminal device; is the channel decoding model parameter corresponding to the edge server when performing channel decoding on the (k - 1)th terminal device; is the semantic decoding model parameter corresponding to the edge server when performing semantic decoding on the (k - 1)th terminal device; η is the learning rate; is the gradient of the transmission loss value corresponding to the (k - 1)th terminal device.

[0116] Furthermore, the rule for the edge server to use the updated channel decoding model parameters and semantic decoding model parameters after the learning of the last terminal device as the initial channel decoding model parameters and initial semantic decoding model parameters for the next iteration round is as follows:

[0117]

[0118]

[0119] Among them, χ (t+1) is the initial channel decoding model parameter of the edge server for the next iteration round; δ (t+1) is the initial semantic decoding model parameter of the edge server for the next iteration round; is the signal decoding model parameter for the edge server to perform the channel decoding of the last terminal-side device in this round; is the semantic decoding model parameter for the edge server to perform the semantic decoding of the last terminal-side device in this round; η is the learning rate; is the gradient of the transmission loss value corresponding to the last terminal-side device in this round.

[0120] In the embodiments of the present invention, in each round of iteration, the edge server first receives and processes the data transmitted by each terminal-side device, and calculates the transmission loss value of the current round. Based on this loss value, the edge server updates the channel decoding model parameter and the semantic decoding model parameter in sequence through the Stochastic Gradient Descent (SGD) algorithm. In this sequential learning process, the adjustment of the model parameters of each terminal-side device is not limited to the local area, but affects the global model parameters of the edge server. Each time the learning of a terminal device is completed, the edge server updates the channel decoding model parameter and the semantic decoding model parameter, and uses these as the basis for the learning of the next terminal device. This sequential update strategy ensures that the system can gradually optimize the model parameters in multiple rounds of iteration, thereby continuously improving the decoding accuracy and efficiency.

[0121] Furthermore, at the end of one round, the edge server will finally determine the channel decoding model parameter and the semantic decoding model parameter according to the learning result of the last terminal-side device. These parameters will be used as the initial parameters in the next round to start a new iteration. This cyclic process will continue until the transmission loss value of the system reaches the minimum and stabilizes, indicating that the system has reached the optimal state.

[0122] Preferably, in each iteration round, the edge server updates the model parameters of channel decoding and semantic decoding by sequentially processing the data of each terminal device. After each update, the system calculates the transmission loss value of the current round, that is, the difference between the decoded image and the original image. Since the model parameter update in each round is based on the optimization result of the previous round, the transmission loss value usually gradually decreases as the number of rounds increases. This loop process does not continue indefinitely but stops after meeting certain conditions. For example, a threshold or convergence criterion is preset. When the transmission loss value drops to a certain level and does not change significantly in several consecutive iterations, the system considers that the model has converged, that is, the transmission loss value of the system has reached the minimum and tends to be stable. At this time, the model parameters of the system have been fully optimized and no further adjustment is required, and the system can be considered to have reached the global optimal state.

[0123] Step S50: Each terminal device periodically sends the collected actual image data to the corresponding edge server based on the established communication connection.

[0124] Specifically, after the iteration is completed, the system will output the optimal semantic encoding model parameters and channel encoding model parameters for each terminal device. These parameters are responsible for efficiently encoding the actual image data on the terminal device side to retain key information and reduce the data volume during transmission. At the same time, the edge server will also output the global optimal channel decoding model parameters and semantic decoding model parameters for decoding and restoring the semantic information received from the terminal device to ensure that the restored data is close to the accuracy of the original data. After obtaining these optimal parameters, the system will re - establish the communication connection between each terminal device and the edge server based on these parameters. This connection is not only the physical - layer link establishment, but more importantly, it is based on the optimized model parameters to ensure the efficiency and accuracy of data transmission and decoding. Specifically, the application of these optimal parameters enables the terminal device to achieve efficient image data encoding and transmission with the minimum communication resource consumption in subsequent operations. At the same time, the edge server can also process and restore the received data with higher decoding accuracy and speed.

[0125] Furthermore, in the subsequent usage process, each terminal device will periodically send the collected actual image data to the corresponding edge server based on these optimal model parameters through the established communication connection. Since these connections and parameters are obtained through iterative optimization, they can minimize information loss and errors during transmission and ensure the transmission efficiency and accuracy of data.

[0126] Based on the solution of the present invention, the optimal semantic coding model parameters and channel coding model parameters enable the terminal device to perform data transmission with a higher compression ratio and stronger anti-interference ability. This greatly reduces the bandwidth required for transmission and improves the overall transmission efficiency. The globally optimal channel decoding model parameters and semantic decoding model parameters adopted by the edge server ensure the accurate restoration of the received coded data. Even in the case of high noise and interference, the system can still maintain a high decoding accuracy, making the restored data highly consistent with the original data. The communication connection established based on the optimal parameters not only ensures the reliability of data transmission but also enables the system to maintain stable performance during long-term operation.

[0127] Figure 2 Figure 4 is the system structure diagram of the edge-end collaborative semantic communication system for the distribution network provided by an embodiment of the present invention. As Figure 2 shown, an embodiment of the present invention provides an edge-end collaborative semantic communication system for the distribution network, and the system includes: a terminal-side device for executing the above-mentioned edge-end collaborative semantic communication method for the distribution network, and an edge server for executing the above-mentioned edge-end collaborative semantic communication method for the distribution network.

[0128] A fourth aspect of the present invention provides a computer-readable storage medium. Instructions are stored on the computer-readable storage medium, and when running on a computer, the instructions cause the computer to execute the above-mentioned edge-end collaborative semantic communication method for the distribution network.

[0129] A fifth aspect of the present invention provides a computer program product, including a computer program, and the computer program realizes the above-mentioned edge-end collaborative semantic communication method when executed by a processor.

[0130] Those skilled in the art can understand that all or part of the steps in the methods of the above embodiments can be completed by instructing relevant hardware through a program. The program is stored in a storage medium, including several instructions for causing a single-chip microcomputer, a chip, or a processor to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0131] The optional embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. In addition, it should be noted that, among the various specific technical features described in the above specific embodiments, they can be combined in any appropriate manner without conflict. To avoid unnecessary repetition, the embodiments of the present invention will not separately describe various possible combination methods.

[0132] In addition, any combination can be made between various different embodiments of the present invention, as long as it does not violate the idea of the embodiments of the present invention, and it should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A method for edge-end collaborative semantic communication for distribution networks, applied to semantic communication between an edge server and a terminal-side device, wherein the edge server is communicatively connected with a plurality of terminal-side devices, characterized in that: The method is performed by a terminal side device, and the method includes: Periodically collecting image data of a set position and periodically constructing a corresponding image test sample set; Sequentially perform semantic coding and channel coding on the image test sample set to obtain a transmission sample set, and send the transmission sample set to a corresponding edge server; The edge server sequentially performs channel decoding and semantic decoding on the received transmission sample sets of each terminal side device, obtains a corresponding decoded image sample set of each terminal side device, and calculates a transmission loss value based on a comparison relationship between each decoded image sample set and a corresponding image test sample set; Taking the minimum transmission loss value as the iteration goal, multiple rounds of image test sample set transmission are performed between the edge server, and the edge server performs semantic coding model parameter, channel coding model parameter, channel decoding model parameter and semantic decoding model parameter iteration based on the stochastic gradient descent algorithm, and establishes a communication connection with each terminal side device based on the corresponding output model parameters; wherein, When performing the first round of iteration, the edge server is configured to: perform parameter aggregation based on the semantic coding model parameters and channel coding model parameters of each terminal side device to obtain global model parameters; based on the global model parameters, complete a transmission sample set decoding, and use the corresponding channel decoding model parameters and semantic decoding model parameters as the initial channel decoding model parameters and initial semantic decoding model parameters of the iteration; The edge server performs parameter aggregation based on the semantic coding model parameters and channel coding model parameters of each terminal side device to obtain global model parameters, including: the edge server compares the transmission loss values ​​corresponding to each terminal side device respectively, and performs weight allocation on each terminal side device based on the transmission loss values ​​corresponding to each terminal side device, where the smaller the transmission loss value, the greater the corresponding weight; based on the weight allocation result, the semantic coding model parameters and channel coding model parameters of each terminal side device are aggregated to obtain global model parameters; The collected actual image data is periodically sent to the corresponding edge server based on the established communication connection.

2. The method according to claim 1, characterized in that: The encoding rule for performing semantic encoding on the image test sample set is: in, is the semantic encoder; α k is the semantic coding model parameter of the kth terminal side device; I is the image test sample set; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; E I It is the semantic feature after semantic encoding.

3. The method according to claim 2, characterized in that The coding rule for performing channel coding on the image test sample set is: in, is a channel encoder; β k is the channel coding model parameter of the k-th terminal side device; I is the image test sample set; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; X I is the semantic feature after channel coding.

4. The method according to claim 1, characterized in that: The transmission rule for sending the transmission sample set to the corresponding edge server is: Based on the additive white Gaussian noise channel, the corresponding edge server is used as the receiving end to perform the transmission of the transmission sample set, and the transmission rule is: in, is the semantic feature after interference; X I is the semantic feature after channel coding; N is the noise; R is the feature embedding dimension of the image; H and W are the height and width of each image, respectively.

5. The method according to claim 1, characterized in that When the edge server performs decoding on the transmission sample set received from each terminal side device, the edge server is configured as follows: Channel decoding and semantic decoding are sequentially performed on the transmission sample sets received from each terminal side device to obtain corresponding decoded image sample sets; wherein, The channel decoding rules are: The semantic decoding rules are: in, represents a channel encoder; χ represents the channel decoding model parameters deployed on the edge server; represents a channel encoder; δ represents the parameters of the semantic decoding model deployed on the edge server; R is the feature embedding dimension of the image; H and W are the height and width of each image respectively; Y I is the semantic feature after channel decoding; is a set of decoded image samples.

6. The method according to claim 1, characterized in that Based on the weight distribution result, the semantic coding model parameters and channel coding model parameters of each terminal side device are aggregated to obtain the global model parameters, which are expressed as: Among them, α and β are global model parameters; K is the number of devices on the terminal side; ω k is the allocation weight of the kth terminal side device; α k is the semantic coding model parameter of the kth terminal side device; β k is the channel coding model parameter of the k-th terminal side device; N k is the channel noise of wireless communication between the kth terminal-side device and the edge server.

7. The method according to claim 1, characterized in that In each round of iteration, the edge server is configured as follows: Send the corresponding semantic coding model parameters and channel coding model parameters respectively, and update the semantic coding model parameters and channel coding model parameters of each terminal side device based on the transmission loss value after the current iteration, as the initial semantic coding model parameters and initial channel coding model parameters of the terminal side device corresponding to the next iteration round; Sequential learning is performed on each terminal side device based on the semantic coding model parameters and channel coding model parameters of each terminal side device. After each learning of a terminal side device is completed, the channel decoding model parameters and the semantic decoding model parameters are updated once until this round of iteration is completed. The channel decoding model parameters and the semantic decoding model parameters updated after the last terminal side device learning are used as the initial channel decoding model parameters and initial semantic decoding model parameters of the edge server in the next iteration round.

8. The method according to claim 7, characterized in that The update rule for the edge server to update the initial coding model parameters and initial channel coding model parameters of each terminal side device in the next iteration round is: in, is the initial semantic coding model parameter of the next iteration round of the k-th terminal side device; is the initial channel coding model parameter of the next iteration round of the k-th terminal side device; are the initial semantic coding model parameters of the kth terminal-side device in this round; are the initial channel coding model parameters of the kth terminal side device in this round; η is the learning rate; is the gradient of the transmission loss value corresponding to the kth terminal-side device in this round.

9. The method according to claim 7, characterized in that: The edge server performs sequential learning on each terminal side device based on the semantic coding model parameters and channel coding model parameters of each terminal side device. After each learning of a terminal side device is completed, the update rule for performing a channel decoding model parameter and a semantic decoding model parameter update is as follows: in, The channel decoding model parameters corresponding to the channel decoding of the k-th terminal side device performed by the edge server; The semantic decoding model parameters corresponding to the semantic decoding of the k-th terminal-side device performed by the edge server; The channel decoding model parameters corresponding to the channel decoding of the k-1th terminal side device performed by the edge server; The semantic decoding model parameters corresponding to the semantic decoding of the k-1th terminal-side device performed by the edge server; η is the learning rate; is the gradient of the transmission loss value corresponding to the k-1th terminal-side device.

10. The method according to claim 7, characterized in that The rule for the edge server to use the channel decoding model parameters and semantic decoding model parameters updated after the last terminal-side device learning as the initial channel decoding model parameters and initial semantic decoding model parameters of the edge server in the next iteration round is: Among them, χ (t+1) The initial channel decoding model parameters of the edge server in the next iteration round; δ (t+1) Initial semantic decoding model parameters for the edge server in the next iteration; Signal decoding model parameters for the last terminal-side device channel decoding performed by the edge server in this round; The semantic decoding model parameters for the last terminal-side device semantic decoding performed by the edge server in this round; η is the learning rate; It is the gradient of the transmission loss value corresponding to the last terminal-side device in this round.

11. A method for edge-end collaborative semantic communication for distribution networks, applied to semantic communication between an edge server and a terminal-side device, wherein the edge server is communicatively connected with a plurality of terminal-side devices, characterized in that: The method is performed by an edge server, and the method includes: Periodically receiving the transmission sample set uploaded by each terminal side device; wherein, The transmission sample set is obtained based on each terminal side device through the following rules: Periodically collecting image data of a set position and periodically constructing a corresponding image test sample set; performing semantic coding and channel coding on the image test sample set in sequence to obtain a transmission sample set; Perform channel decoding and semantic decoding on the received transmission sample sets of each terminal side device in sequence, obtain the decoded image sample sets of each terminal side device accordingly, and calculate the transmission loss value based on the comparison relationship between each decoded image sample set and the corresponding image test sample set; Taking the minimum transmission loss value as the iteration goal, perform multiple rounds of image test sample set transmission between each terminal side device, perform semantic coding model parameter, channel coding model parameter, channel decoding model parameter and semantic decoding model parameter iteration based on the stochastic gradient descent algorithm, and establish a communication connection with each edge side device based on the corresponding output model parameters; When performing the first round of iteration, the edge server is configured to: perform parameter aggregation based on the semantic coding model parameters and channel coding model parameters of each terminal side device to obtain global model parameters; based on the global model parameters, complete a transmission sample set decoding, and use the corresponding channel decoding model parameters and semantic decoding model parameters as the initial channel decoding model parameters and initial semantic decoding model parameters of the iteration; The edge server performs parameter aggregation based on the semantic coding model parameters and channel coding model parameters of each terminal side device to obtain global model parameters, including: the edge server compares the transmission loss values ​​corresponding to each terminal side device respectively, and performs weight allocation on each terminal side device based on the transmission loss values ​​corresponding to each terminal side device, where the smaller the transmission loss value, the greater the corresponding weight; based on the weight allocation result, the semantic coding model parameters and channel coding model parameters of each terminal side device are aggregated to obtain global model parameters; The actual image data collected by each edge-side device is periodically recovered based on the established communication connection.

12. The method according to claim 11, characterized in that In each round of iteration, the method includes: Send the corresponding semantic coding model parameters and channel coding model parameters respectively, and update the semantic coding model parameters and channel coding model parameters of each terminal side device based on the transmission loss value after the current iteration, as the initial semantic coding model parameters and initial channel coding model parameters of the terminal side device corresponding to the next iteration round; Sequential learning is performed on each terminal side device based on the semantic coding model parameters and channel coding model parameters of each terminal side device. After each learning of a terminal side device is completed, the channel decoding model parameters and the semantic decoding model parameters are updated once until this round of iteration is completed. The channel decoding model parameters and the semantic decoding model parameters updated after the last terminal side device learning are used as the initial channel decoding model parameters and initial semantic decoding model parameters of the edge server in the next iteration round.

13. The method according to claim 12, characterized in that The update rule for the edge server to update the initial coding model parameters and initial channel coding model parameters of each terminal side device in the next iteration round is: in, is the initial semantic coding model parameter of the next iteration round of the k-th terminal side device; is the initial channel coding model parameter of the next iteration round of the k-th terminal side device; are the initial semantic coding model parameters of the kth terminal-side device in this round; are the initial channel coding model parameters of the kth terminal side device in this round; η is the learning rate; is the gradient of the transmission loss value corresponding to the kth terminal-side device in this round.

14. The method according to claim 12, characterized in that The edge server performs sequential learning on each terminal side device based on the semantic coding model parameters and channel coding model parameters of each terminal side device. After each learning of a terminal side device is completed, the update rule for performing a channel decoding model parameter and a semantic decoding model parameter update is as follows: in, The channel decoding model parameters corresponding to the channel decoding of the k-th terminal side device performed by the edge server; The semantic decoding model parameters corresponding to the semantic decoding of the k-th terminal-side device performed by the edge server; The channel decoding model parameters corresponding to the channel decoding of the k-1th terminal side device performed by the edge server; The semantic decoding model parameters corresponding to the semantic decoding of the k-1th terminal-side device performed by the edge server; η is the learning rate; is the gradient of the transmission loss value corresponding to the k-1th terminal-side device.

15. The method according to claim 12, characterized in that The rule for the edge server to use the channel decoding model parameters and semantic decoding model parameters updated after the last terminal-side device learning as the initial channel decoding model parameters and initial semantic decoding model parameters of the edge server in the next iteration round is: Among them, χ (t+1) The initial channel decoding model parameters of the edge server in the next iteration round; δ (t+1) Initial semantic decoding model parameters for the edge server in the next iteration; Signal decoding model parameters for the last terminal-side device channel decoding performed by the edge server in this round; The semantic decoding model parameters for the last terminal-side device semantic decoding performed by the edge server in this round; η is the learning rate; It is the gradient of the transmission loss value corresponding to the last terminal-side device in this round.

16. An edge-end collaborative semantic communication system for distribution networks, applied to semantic communication between edge servers and terminal devices, wherein the edge servers are communicatively connected with a plurality of terminal devices, characterized in that: The system includes: a terminal-side device for executing the edge-end collaborative semantic communication method for distribution networks as described in any one of claims 1-10, and an edge server for executing the edge-end collaborative semantic communication method for distribution networks as described in any one of claims 11-15.

17. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the edge collaborative semantic communication method for distribution networks as described in any one of claims 1 to 15.

18. A computer program product comprising a computer program, characterized in that When executed by a processor, the computer program implements the edge collaborative semantic communication method for distribution networks according to any one of claims 1 to 15.

Images