Methods and devices for updating semantic encoding and decoding models

Abstract

This application provides a method and apparatus for updating a semantic codec model. The method includes: obtaining the current data distribution drift intensity for evaluating the performance degradation of the semantic codec model; calculating the current model's performance degradation rate based on the correlation function between the drift intensity and the model performance degradation rate; constructing an average drift-aware model age expression based on the model performance degradation rate and the queuing model type; calculating the relative computational load based on the drift intensity; calculating the total service time for a single model update based on the relative computational load; and periodically executing the model update steps until the update termination requirements are met, at which point execution stops. The model update steps include: solving a preset objective function to obtain an optimal update frequency; and updating the current model at the optimal update frequency to obtain the updated semantic codec model. This application enables dynamic adjustment of the update frequency, improving system resource utilization efficiency while ensuring semantic reconstruction performance.

Description

Technical Field

[0001] This application relates to the field of semantic communication technology, and in particular to a method and apparatus for updating a semantic encoding and decoding model. Background Technology

[0002] With the deep integration of 6G and artificial intelligence, semantic communication, as a new paradigm that transcends traditional symbol transmission, demonstrates significant advantages in scenarios with limited bandwidth and poor channel quality by directly understanding and processing the semantics of information. The semantic encoding and decoding model, as the core component of the system, determines the quality of semantic information reconstruction. However, in actual deployments, the model training environment often differs from the real-time operating environment, especially in dynamic scenarios such as connected vehicles and industrial IoT. Data distribution continuously evolves with the environment, tasks, or user behavior, leading to a gradual decline in model performance and impacting the reliability of the semantic communication system.

[0003] To maintain model performance, periodic updates to the edge model are necessary. The Age of Information (AoI) and its variation in traditional state update systems have been widely used to optimize data freshness, but these metrics primarily quantify staleness over time. In practical applications, the same update interval can lead to drastically different performance losses under slight and severe drift. With slight drift, the model maintains good performance, and frequent updates waste resources. With severe drift, model performance deteriorates rapidly, and insufficient updates cannot promptly curb this degradation. Traditional AoI frameworks ignore this fundamental impact of data distribution drift on performance degradation rates, easily leading to a double inefficiency in resource-constrained edge environments: over-updating when time is needed and under-updating when necessary. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method and apparatus for updating a semantic encoding and decoding model to eliminate or improve one or more defects existing in the prior art, enabling adaptive optimization of the model update frequency under resource constraints, and improving system resource utilization efficiency while ensuring semantic reconstruction performance, thereby improving the model performance assurance capability.

[0005] One aspect of this application provides a method for updating a semantic codec model, the method comprising the following steps: Obtain the current data distribution drift intensity used to assess the performance degradation of the semantic codec model; The performance degradation rate of the current semantic codec model is calculated based on the current data distribution drift intensity and the preset model performance degradation rate correlation function, thus obtaining the model performance degradation rate corresponding to the current semantic codec model; Based on the model performance degradation rate and queuing model type, construct the average drift-aware model age expression corresponding to the current semantic codec model according to queuing theory; calculate the relative computational load corresponding to the current semantic codec model based on the current data distribution drift intensity; and calculate the total service time for a single model update based on the relative computational load. The model update step is executed periodically until a preset update termination requirement is met, at which point execution stops. The model update step includes: solving a preset objective function based on the model performance degradation rate, the current data distribution drift intensity, the relative computing load, the system effective computing rate, and the queuing model type to obtain an optimal update frequency; and updating the current semantic codec model with the optimal update frequency to obtain an updated semantic codec model.

[0006] In some embodiments of this application, the generation of the model performance degradation rate correlation function includes the following steps: Obtain the baseline performance of the current semantic codec model on the source domain image set, and the measured performance of the current semantic codec model on multiple target domain image sets with different data distribution drift intensities; Based on the baseline performance and the measured performance, calculate the model performance degradation rate corresponding to each of the data distribution drift intensities; A multinomial regression fitting is performed on the data pair of the data distribution drift intensity and the model performance degradation rate to obtain the correlation function of the model performance degradation rate.

[0007] In some embodiments of this application, the calculation process for the relative computing load is shown in the following formula: in, Indicates relative computational load; Indicates the current data distribution drift intensity; This represents the growth factor of the computational workload.

[0008] In some embodiments of this application, the calculation process for the total service time of a single model update is shown in the following formula: in, This indicates the total service time for a single model update; Indicates the number of model parameters; Indicates channel rate; This indicates the baseline calculation delay.

[0009] In some embodiments of this application, the step of solving a preset objective function based on the model performance degradation rate, the current data distribution drift intensity, the relative computational load, the system effective computational rate, and the queuing model type to obtain the optimal update frequency includes: Select the corresponding normalized queuing cost function based on the queuing model type; The objective function is constructed using the normalized queuing cost function, the model performance degradation rate, the current data distribution drift intensity, the relative computational load, and the system's effective computational rate. The optimal normalized frequency is obtained by minimizing the objective function. The optimal update frequency is calculated and obtained based on the optimal normalized frequency.

[0010] In some embodiments of this application, minimizing the objective function to obtain the optimal normalized frequency includes: If the queuing model type is edge layer M / M / 1, then the objective function is minimized using numerical optimization methods such as gradient descent to obtain the optimal normalized frequency; If the queuing model type is cloud layer D / M / 1, then the bisection method is used to search for congestion factors within a preset interval to minimize the objective function and obtain the optimal normalized frequency.

[0011] In some embodiments of this application, updating the current semantic codec model at the optimal update frequency to obtain the updated semantic codec model includes: The update monitoring steps are performed periodically: based on the optimal update frequency, the corresponding update interval is determined; and the update time difference between the current time and the last model update time is calculated; based on the update interval and the update time difference, it is determined whether to perform a model update; if the update time difference is less than the update interval, the update monitoring steps are continued; if the update time difference is greater than or equal to the update interval, an update operation is performed on the current semantic codec model to obtain the updated semantic codec model.

[0012] In some embodiments of this application, after performing an update operation on the current semantic codec model based on the optimal update frequency to obtain the updated semantic codec model, the method further includes: Deploy the updated semantic codec model to the edge and reset the age state of the drift-aware model.

[0013] Another aspect of this application provides an electronic device including a processor and a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for updating the semantic codec model.

[0014] A third aspect of this application provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method for updating the semantic encoding / decoding model.

[0015] The fourth aspect of this application provides a computer program product comprising a computer program that, when executed by a processor, implements the method for updating the semantic codec model.

[0016] The semantic codec model update method of this application includes the following steps: obtaining the current data distribution drift intensity for evaluating the performance degradation of the semantic codec model; calculating the performance degradation rate of the current semantic codec model based on the current data distribution drift intensity and a preset model performance degradation rate correlation function, to obtain the model performance degradation rate corresponding to the current semantic codec model; constructing an average drift-aware model age expression corresponding to the current semantic codec model based on the model performance degradation rate and the queuing model type, according to queuing theory; calculating the relative computational load corresponding to the current semantic codec model based on the current data distribution drift intensity; calculating the total service time for a single model update based on the relative computational load; periodically executing the model update step until a preset update termination requirement is met, and then stopping execution; wherein, the model update step includes: solving a preset objective function based on the model performance degradation rate, the current data distribution drift intensity, the relative computational load, the system effective computational rate, and the queuing model type to obtain an optimal update frequency; and updating the current semantic codec model at the optimal update frequency to obtain an updated semantic codec model. By establishing a correlation model between drift intensity and model performance degradation rate, and deriving a closed-form solution for the optimal update frequency based on queuing theory, the update frequency can be dynamically adjusted according to drift intensity and system resource status. This allows the semantic reconstruction performance of the model to be maintained in resource-constrained edge environments, thereby improving the reliability and transmission efficiency of the semantic communication system and ultimately enhancing system resource utilization efficiency.

[0017] Additional advantages, objectives, and features of this application will be set forth in part in the description which follows, and will in part become apparent to those skilled in the art upon review of the following description, or may be learned by practice of the application. The objectives and other advantages of this application can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0018] Those skilled in the art will understand that the purposes and advantages that can be achieved with this application are not limited to those specifically described above, and that the above and other purposes that this application can achieve will be more clearly understood from the following detailed description. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, do not constitute a limitation thereof. The components in the drawings are not drawn to scale but are merely for illustrating the principles of this application. For ease of illustration and description of certain parts of this application, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to this application. In the drawings: Figure 1 This is a schematic diagram of the first process of updating the semantic encoding / decoding model in one embodiment of this application.

[0020] Figure 2 This is a schematic diagram of a second process for updating a semantic encoding / decoding model in one embodiment of this application.

[0021] Figure 3 This is a schematic diagram of the third process of updating the semantic encoding and decoding model in one embodiment of this application.

[0022] Figure 4 This is a schematic diagram of a system framework for the semantic encoding / decoding model update method in a specific example of this application.

[0023] Figure 5 This is a flowchart illustrating a model performance degradation rate correlation module for a specific example of the semantic encoding / decoding model update method in this application.

[0024] Figure 6 This is a flowchart illustrating the DA-AoM theoretical framework and computational load module for updating the semantic encoding / decoding model in a specific example of this application.

[0025] Figure 7 This is a flowchart illustrating a DA-AoM-driven adaptive update frequency decision module, which is a specific example of the semantic encoding / decoding model update method in this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and their descriptions are used to explain this application, but are not intended to limit it.

[0027] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the structures and / or processing steps closely related to the solution according to this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0028] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0029] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0030] In the following description, embodiments of the present application will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0031] It's important to note that in real-world semantic communication deployments, data distribution drift leads to a continuous decline in the performance of semantic encoding / decoding models, necessitating timely model updates to maintain performance. However, model updates rely on data transmission from the edge to the cloud. Under conditions of limited bandwidth and computing resources, overly frequent update strategies can cause network congestion or wasted computing resources, while excessively low update frequencies fail to curb performance degradation. Traditional edge intelligence systems primarily rely on periodic retraining or fixed-interval replacement strategies for model updates. This simple mechanism can maintain performance in static environments, but it leads to a double inefficiency in dynamic scenarios. When data distribution is stable, frequent updates waste computing resources and network bandwidth; when drastic environmental changes cause severe drift, fixed-interval updates cannot promptly curb performance degradation, resulting in a prolonged period of high-error operation. Semantic communication systems have a more urgent and complex need for model updates. Unlike traditional tasks, the performance of semantic encoding / decoding models directly determines the semantic fidelity of information reconstruction. The impact of data distribution drift on semantic feature extraction is often non-linear; slight drift may be absorbed by the model's generalization ability, but performance will drastically degrade once a critical threshold is exceeded. Furthermore, semantic encoding / decoding models are typically deployed on resource-constrained edge nodes, and the update process is limited by uplink bandwidth, cloud computing queues, and the number of model parameters. Simple periodic strategies struggle to balance update timeliness with resource overhead. The model update process of a semantic communication system involves the edge device continuously collecting real-time data and detecting the intensity of data distribution drift. When an update is deemed necessary, new data or gradient information is uploaded to the edge or cloud server. The server performs model training based on the received data, generating updated model parameters. The new model is then compressed and encoded before being sent back to the edge, completing a full update loop. In this process, the generation interval of update requests, server queuing time, and computation time for model training collectively determine the model's freshness, while the intensity of data distribution drift determines the rate of performance degradation. In research on model update strategies, Age of Information (AoI) has become a fundamental indicator for measuring data freshness. AoI is defined as the elapsed time since the information source generated the last received data packet, and it is widely used to optimize update strategies for various time-sensitive systems. To more precisely characterize the freshness of different systems, researchers have proposed several extended indicators, such as Age of Processing (AoP) and Age of Task (AoT), shifting their focus from simply representing the freshness of data transmission to more comprehensively depicting the state freshness of the entire real-time update system. In AI model update scenarios, some research has proposed the concept of Age of Model (AoM), incorporating the number of training data categories into the age calculation framework to guide model updates in the edge intelligence environment of connected vehicles.The AoM (Aspect-Oriented Model) metric models the functional relationship between the rate of decline of model inference performance over time and the diversity of training data, providing a quantitative tool for balancing update frequency, computational resources, and update quality. In the above process, model update frequency is a key parameter affecting the AoM metric; here, update frequency is defined as the reciprocal of the time interval between two adjacent model update requests, denoted as . However, this study primarily focuses on the specific factor of the number of training data categories, failing to directly model the more general and critical problem of data distribution drift. In scenarios such as semantic communication, the difference in feature distribution between the source and target domains is the root cause of model performance degradation. Directly equating the number of categories with update quality may not accurately reflect the performance degradation caused by complex distribution changes. Furthermore, the intensity of distribution drift is a continuous and multidimensional measure; the same update interval can result in drastically different performance losses under slight and severe drift, and a fixed update frequency cannot adapt to such dynamic changes. Therefore, if a semantic communication system adopts a constant update frequency in the face of complex and ever-changing data distribution drift environments, it is not conducive to maintaining the high performance of the model. Based on this, the inventors of this application first conceived of dynamically adjusting the model update frequency to adapt to the changing drift environment. This is because under different data distribution drift conditions, the rate of model performance degradation and the computational load of updates vary, and a fixed update frequency may prevent the system from responding promptly to performance degradation in certain situations, potentially leading to resource waste. Therefore, the inventors of this application innovatively propose to dynamically adjust the update frequency based on different data distribution drift intensities, thereby improving the resource utilization efficiency and model performance assurance capabilities of the semantic communication system. The dynamic update method for semantic encoding and decoding models based on Drift-Aware Age of Model (DA-AoM) of this application can dynamically adjust the model update frequency according to the data distribution drift intensity and system resource status, thereby improving the resource utilization efficiency and model performance assurance capabilities of the semantic communication system.

[0032] The following examples will provide a detailed description.

[0033] This application provides a method for updating a semantic encoding / decoding model, see [link to relevant documentation]. Figure 1 The method includes the following steps: Step 100: Obtain the current data distribution drift intensity to evaluate the performance degradation of the semantic codec model; In step 100, the current data distribution drift intensity can be a normalized scalar value used to characterize the degree of distribution difference between the target domain image and the source domain image in the semantic feature space. This current data distribution drift intensity is obtained by measuring the semantic distance between the measured image and the pre-training image, specifically including: firstly, extracting deep semantic features from the pre-training image and the measured image respectively; then, calculating and fusing the multi-dimensional semantic distance based on the extracted semantic features to obtain the original data distribution drift data; and finally, normalizing the original data distribution drift data to generate a current data distribution drift intensity with a value between 0 and 1.

[0034] Step 200: Calculate the performance degradation rate of the current semantic codec model based on the current data distribution drift intensity and the preset model performance degradation rate correlation function, and obtain the model performance degradation rate corresponding to the current semantic codec model; In step 200, the model performance degradation rate can be the degree of decline in the reconstruction quality of the current semantic encoding and decoding model on the measured image relative to the baseline performance of the pre-trained image, used to quantify the impact of data distribution drift on model performance.

[0035] Step 300: Based on the model performance degradation rate and queuing model type, construct the average drift-aware model age expression corresponding to the current semantic codec model according to queuing theory; calculate the relative computational load corresponding to the current semantic codec model based on the current data distribution drift intensity; and calculate the total service time for a single model update based on the relative computational load. In step 300, the relative computational load can be the ratio of the computational load required to complete a single model update under the current data distribution drift intensity to the baseline computational load, used to quantify the impact of data distribution drift on the computational resource consumption of model updates; the total service time for a single model update refers to the sum of network transmission latency and computational processing latency experienced from the generation of the model update request to the completion of the update. When the first The model update occurs at time [time]. Completed, in the following time period. The model is currently in an unupdated state, and the current data distribution drift intensity is... The model performance degradation rate is Then, the instantaneous drift-aware model age expression (instantaneous DA-AoM) can be defined as the product of the performance degradation rate and the unupdated time, as shown in the following formula: in, Indicates the age of the instantaneous drift perception model; This represents the model performance degradation rate. Next, a geometric decomposition method is used to decompose the total area within the observation period into a series of non-overlapping geometric regions. The area of ​​each period is calculated using the update request arrival interval and the total update duration. When the system enters a steady state, the update process is smooth and ergonomic. The average DA-AoM can be further constructed using queuing theory, yielding closed-form expressions for edge layer M / M / 1 and cloud layer D / M / 1 scenarios, respectively. When the queuing model type is an edge layer M / M / 1 queue, the expression for the average drift-aware model age is: in, Indicates the update frequency; This represents the computation rate of model updates. When the queuing model type is a cloud D / M / 1 queue, the expression for the average drift-aware model age is: in, This indicates the interval at which update requests are generated in the queuing model; Represents the congestion factor. It reflects the probability of finding the service desk busy when the task arrives; it is a transcendental equation. In the interval The only real root within.

[0036] Step 400: Periodically execute the model update step until the preset update termination requirement is met, then stop execution; wherein, the model update step includes: solving a preset objective function based on the model performance degradation rate, the current data distribution drift intensity, the relative computing load, the system effective computing rate, and the queuing model type to obtain the optimal update frequency; and updating the current semantic codec model with the optimal update frequency to obtain the updated semantic codec model.

[0037] In step 400, the preset update termination requirements include determining whether the current semantic communication service has ended; if the current semantic communication service has not ended, the model update step continues; if the current semantic communication service has ended, the model update step ends, and the updated semantic encoding / decoding model is obtained. The current semantic communication service refers to the ongoing image semantic transmission task; when the task ends, the model update stops.

[0038] As described above, the semantic encoding / decoding model update method provided in this application, by establishing a nonlinear correlation model between the intensity of data distribution drift and the model performance degradation rate, can achieve efficient prediction of semantic reconstruction performance loss, enabling update decisions to be upgraded from traditional time obsolescence to semantic degradation awareness. Furthermore, by introducing the DA-AoM theoretical framework and deriving a closed-form solution for the optimal update frequency based on queuing theory, the update frequency can be dynamically adjusted according to the intensity of data distribution drift and the system resource status. When the drift is slight, update overhead is reduced to avoid resource waste; when the drift is severe, timely updates are triggered to curb performance degradation, achieving a dynamic balance between resource efficiency and model performance. Compared with traditional fixed-frequency update strategies, this application can more flexibly adapt to complex and changing data distribution environments, improving the resource utilization efficiency and model performance assurance capabilities of the semantic communication system.

[0039] To further achieve efficient prediction of semantic reconstruction performance loss and upgrade update decisions from traditional temporal obsolescence to semantic decay awareness, an update method for a semantic encoding / decoding model is provided in this application embodiment, see [link to relevant documentation]. Figure 2 The generation of the model performance degradation rate correlation function includes the following steps: Step 010: Obtain the baseline performance of the current semantic codec model on the source domain image set, and the measured performance of the current semantic codec model on multiple target domain image sets with different data distribution drift intensities; In step 010, the baseline performance of the current semantic codec model can be the average reconstruction accuracy index between the original image and the reconstructed image measured on the source domain image set, used as a reference benchmark for model performance; the measured performance can be the average reconstruction accuracy index between the original image and the reconstructed image measured on the target domain image set with the current data distribution drift intensity. The original image is input into the semantic codec model, mapped to semantic features by the semantic encoder, transmitted through the channel, and then the reconstructed image is recovered by the semantic decoder. The reconstruction accuracy index (such as peak signal-to-noise ratio (PSNR)) between the original image and the reconstructed image is calculated. The baseline performance of the model is obtained by testing in the source domain, and the measured performance is obtained by testing in the target domain with the current data distribution drift intensity.

[0040] Step 020: Based on the baseline performance and the measured performance, calculate the model performance degradation rate corresponding to each of the data distribution drift intensities; In step 020, the model performance degradation rate corresponding to each of the data distribution drift intensities can be expressed by the following formula: in, Represents the source domain; Indicates the target domain; This represents the model performance degradation rate corresponding to the drift intensity of each of the aforementioned data distributions; Indicates baseline performance; This indicates the measured performance.

[0041] Step 030: Perform a multinomial regression fitting on the data pair of the data distribution drift intensity and the model performance degradation rate to obtain the correlation function of the model performance degradation rate.

[0042] In step 030, based on data obtained from numerous experiments... For the sample pairs, a correlation model was established using the multinomial regression method to obtain the correlation function for the model's decline rate, which can be expressed as follows: in, , , as well as These represent the polynomial coefficients determined by least squares regression, used to characterize the nonlinear mapping relationship between the intensity of data distribution drift and the rate of model performance degradation. (The goodness-of-fit factor is then used to...) The root mean square error (RMSE) is used to evaluate the quality of the fit, ensuring that the correlation model has good generalization ability on both the training and test sets, and enabling the prediction of model performance loss.

[0043] To further dynamically adjust the update frequency based on the data distribution drift intensity and system resource status, and to achieve a dynamic balance between resource efficiency and model performance, the calculation process of the relative computational load in the semantic encoding and decoding model update method provided in this application embodiment is shown in the following formula: in, Indicates relative computational load; Indicates the current data distribution drift intensity; This represents the growth factor of the computational workload.

[0044] In one or more embodiments of this application, a theoretical drift-aware dynamic load mechanism is used to reduce the relative computational load. Modeled as drift intensity A linearly increasing function. The theoretical drift-aware dynamic load mechanism assumes that the system has ideal adaptive capabilities, able to dynamically adjust the computational depth or complexity of the model according to the increase in the intensity of data distribution drift, thereby maintaining the accuracy of feature extraction. Under this mechanism, the relative computational load increases linearly with the drift intensity.

[0045] To further dynamically adjust the update frequency based on the data distribution drift intensity and system resource status, and to achieve a dynamic balance between resource efficiency and model performance, the calculation process of the total service time for a single model update in the semantic encoding and decoding model update method provided in this application embodiment is shown in the following formula: in, This indicates the total service time for a single model update; Indicates the number of model parameters; Indicates channel rate; This indicates the baseline calculation delay.

[0046] In one or more embodiments of this application, the total service time for a single model update is calculated based on the relative computing load, and the theoretical construction of DA-AoM is completed by combining the performance degradation rate and queuing theory model.

[0047] To further adapt more flexibly to complex and ever-changing data distribution environments and improve the resource utilization efficiency and model performance assurance capabilities of semantic communication systems, a method for updating a semantic encoding / decoding model is provided in this application embodiment, see [link to relevant documentation]. Figure 3 Step 400, which involves solving a preset objective function based on the model performance degradation rate, the current data distribution drift intensity, the relative computational load, the system's effective computational rate, and the queuing model type to obtain the optimal update frequency, includes: Step 410: Select the corresponding normalized queuing cost function according to the queuing model type; Step 420: Construct the objective function using the normalized queuing cost function, the model performance degradation rate, the current data distribution drift intensity, the relative computational load, and the system's effective computational rate to obtain the objective function; Step 430: Minimize the objective function to obtain the optimal normalized frequency; Step 440: Calculate and obtain the optimal update frequency based on the optimal normalized frequency.

[0048] In one or more embodiments of this application, the effective computing rate of the system can be the actual service rate of the current computing node when processing model update tasks, which is jointly determined by the peak computing power of the hardware, the resource allocation coefficient, and the computing load. The objective function can be expressed by the following formula: in, Represent the objective function; Indicates the normalization update frequency; This represents the normalized queuing cost function, if the queuing model type When it is an edge layer M / M / 1 queue, then It consists of a model obsolescence term and a queuing congestion term; if the queuing model type When it is a cloud layer D / M / 1 queue, then It also needs to include the congestion factor. , By solving the transcendental equations We obtain the optimal normalized update frequency by minimizing the average DA-AoM. The optimal update frequency is expressed by the following formula: in, This represents the optimal normalized frequency.

[0049] To further adapt more flexibly to complex and ever-changing data distribution environments and improve the resource utilization efficiency and model performance assurance capabilities of semantic communication systems, in a semantic encoding / decoding model update method provided in this application embodiment, step 430 includes: Step 431: If the queuing model type is edge layer M / M / 1, then use numerical optimization methods such as gradient descent to minimize the objective function and obtain the optimal normalized frequency; Step 432: If the queuing model type is cloud layer D / M / 1, then the congestion factor is searched within the preset interval using the bisection method to minimize the objective function and obtain the optimal normalized frequency.

[0050] In step 432, if the queuing model type is cloud D / M / 1, since It is an implicit function, and the optimal normalized frequency is searched within a preset interval using a bisection method. The preset interval can be... .

[0051] In one or more embodiments of this application, the average DA-AoM closed-form expression is selected based on the queuing model type, and the optimal normalized frequency under resource constraints is solved.

[0052] To further adapt more flexibly to complex and ever-changing data distribution environments and improve the resource utilization efficiency and model performance assurance capabilities of semantic communication systems, in a semantic encoding / decoding model update method provided in this application embodiment, step 400, which updates the current semantic encoding / decoding model at the optimal update frequency to obtain the updated semantic encoding / decoding model, includes: Step 450: Periodically execute the update monitoring step: Based on the optimal update frequency, determine the corresponding update interval; and calculate the update time difference between the current time and the last model update time; determine whether to perform a model update based on the update interval and the update time difference; if the update time difference is less than the update interval, continue executing the update monitoring step; if the update time difference is greater than or equal to the update interval, perform an update operation on the current semantic codec model to obtain the updated semantic codec model.

[0053] In one or more embodiments of this application, the update time difference can be expressed as follows: in, Indicates the update time difference; Indicates the current time; This represents the time of the last model update. The update interval can be expressed by the following formula: in, Indicates the update interval.

[0054] To further adapt more flexibly to complex and ever-changing data distribution environments and improve the resource utilization efficiency and model performance assurance capabilities of semantic communication systems, in a semantic codec model update method provided in this application embodiment, after performing an update operation on the current semantic codec model based on the optimal update frequency to obtain the updated semantic codec model, the method further includes: Step 460: Deploy the updated semantic codec model to the edge and reset the age state of the drift-aware model.

[0055] In one or more embodiments of this application, the system monitors the current time in real time. Compared to the last update time ,when The model is updated in time, the latest semantic codec model is deployed to the edge and the DA-AoM state is reset.

[0056] In a specific example of the semantic encoding / decoding model update method provided in this application, see [link to example]. Figure 4 The method includes the following steps: Receive current data distribution and drift intensity from external drift monitoring module. ; The performance degradation rate correlation module is based on Calculate the performance degradation rate of the model The DA-AoM theoretical framework and the evaluation of computational load modules relative to computational load ; The optimal update frequency decision module solves for the optimal normalized update frequency based on the DA-AoM closed-form expression. And convert it into the actual update frequency. ; System by The model is triggered to update, and it iterates periodically until the semantic communication service is terminated.

[0057] The specific implementation of the semantic encoding / decoding model update method in this embodiment mainly relies on three core modules: a model performance degradation rate correlation module, a DA-AoM theoretical framework and computational load module, and a DA-AoM-driven adaptive update frequency decision module. The main function of the model performance degradation rate correlation module is to establish the data distribution drift intensity. Performance degradation rate of semantic encoding and decoding models A quantitative correlation model between the measured and pre-trained images is used to predict model performance loss. This module first measures the data distribution drift of the current image based on the semantic distance between the measured and pre-trained images. Then calculate the model performance degradation rate under the current drift. (Image) The input semantic encoding and decoding model is mapped into semantic features by the semantic encoder, and then the semantic decoder recovers and reconstructs the image after transmission through the channel. Calculate the original image With reconstructed images The reconstruction accuracy metrics (such as peak signal-to-noise ratio, PSNR) are used to measure the model's baseline performance in the source domain. In the current data distribution drift intensity Actual performance was obtained by testing on the target domain. The model performance degradation rate is defined as Based on extensive experimental data For the sample pairs, a correlation model was established using the multinomial regression method to obtain the model's decay rate function. The cubic fitting polynomial is obtained. By goodness of fit The root mean square error (RMSE) is used to evaluate the fit quality, ensuring that the correlation model has good generalization ability on both the training and test sets. The specific algorithm for the model performance degradation rate correlation module is shown in Table 1, and the process is as follows: Figure 5 As shown.

[0058] Table 1. Detailed algorithm description of the module related to model performance degradation rate.

[0059] The main function of the DA-AoM theoretical framework and computational load module is to construct a theoretical framework for the age of drift-aware models, characterize the dual impact of data distribution drift on model freshness, and establish an experimental metric model. This module first defines instantaneous DA-AoM. Assume the... The model update occurs at time [time]. Completed, in the following time period. The model is currently in an unupdated state, and the current data distribution drift intensity is... The model performance degradation rate is Then, instantaneous DA-AoM is defined as the product of the performance degradation rate and the time without update, i.e. Further, the average DA-AoM is derived. A geometric decomposition method is used to decompose the total area within the observation period into the sum of a series of non-overlapping geometric regions. The area of ​​each period is updated to the interval via update requests. and total update duration Calculation. Assuming the system enters a steady state, the update process is smooth and ergodic, and the average DA-AoM can be further derived using queuing theory, yielding closed-form expressions for edge layer M / M / 1 and cloud layer D / M / 1 scenarios, respectively.

[0060] The closed-form expression for the average DA-AoM in the edge layer M / M / 1 scenario is: The closed-form expression for the average DA-AoM in the cloud D / M / 1 scenario is: in, To update frequency, For the computational rate of model updates, Generate intervals and congestion factors for update requests in the queuing model. It reflects the probability of finding the service desk busy when the task arrives; it is a transcendental equation. In the interval The only real root within.

[0061] This module is based on a theoretical drift-aware dynamic load mechanism, which will reduce the relative computational load. Modeled as drift intensity A linearly growing function, i.e. .based on Calculate the total service time for a single model update And combined with performance degradation rate The theoretical framework of DA-AoM is completed using queuing theory models. The specific algorithms for the DA-AoM theoretical framework and computational load module are shown in Table 2, and the process is as follows: Figure 6 As shown.

[0062] Table 2. DA-AoM Theoretical Framework and Specific Algorithm Descriptions of Computational Load Modules

[0063] The main function of the DA-AoM-driven adaptive update frequency decision module is to solve for the optimal model update frequency under resource constraints based on the DA-AoM closed-form expression. This module determines the optimal model update frequency based on the current data distribution drift intensity. Model performance degradation rate Relative computing load System effective computing speed and queuing model types (Edge layer M / M / 1 or cloud layer D / M / 1), with the objective of minimizing the average DA-AoM, solve for the optimal normalized update frequency. Thus, the actual optimal update frequency is obtained. The system monitors the current time in real time. Compared to the last update time ,when The model update is triggered periodically, deploying the latest semantic codec model to the edge and resetting the DA-AoM state. The specific algorithm for the DA-AoM-driven adaptive update frequency decision module is shown in Table 3, and the process is as follows: Figure 7 As shown.

[0064] Table 3. Detailed algorithm description of the DA-AoM-driven adaptive update frequency decision module.

[0065] This application also provides an electronic device, which may include a processor, a memory, a receiver, and a transmitter. The processor is used to execute the semantic codec model update method mentioned in the above embodiments. The processor and the memory can be connected via a bus or other means, taking a bus connection as an example. The receiver can be connected to the processor and the memory via wired or wireless means.

[0066] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0067] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the semantic codec model update method described in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the semantic codec model update method described in the above method embodiments.

[0068] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0069] The one or more modules are stored in the memory, and when executed by the processor, the semantic codec model update method described in the embodiment is executed.

[0070] In some embodiments of this application, the user equipment may include a processor, a memory, and a transceiver unit. The transceiver unit may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected via a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0071] As one implementation method, the functions of the receiver and transmitter in this application can be implemented by transceiver circuits or dedicated transceiver chips, and the processor can be implemented by dedicated processing chips, processing circuits or general-purpose chips.

[0072] As another implementation approach, the server provided in this application embodiment can be implemented using a general-purpose computer. That is, the program code implementing the processor, receiver, and transmitter functions is stored in memory, and the general-purpose processor implements the processor, receiver, and transmitter functions by executing the code in memory.

[0073] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned semantic encoding / decoding model update method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0074] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned semantic codec model update method.

[0075] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave.

[0076] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0077] In this application, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0078] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to the embodiments of this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for updating a semantic encoding / decoding model, characterized in that, The method includes: Obtain the current data distribution drift intensity used to assess the performance degradation of the semantic codec model; The performance degradation rate of the current semantic codec model is calculated based on the current data distribution drift intensity and the preset model performance degradation rate correlation function, thus obtaining the model performance degradation rate corresponding to the current semantic codec model; Based on the model performance degradation rate and queuing model type, construct the average drift-aware model age expression corresponding to the current semantic codec model according to queuing theory; calculate the relative computational load corresponding to the current semantic codec model based on the current data distribution drift intensity; and calculate the total service time for a single model update based on the relative computational load. The model update step is executed periodically until a preset update termination requirement is met, at which point execution stops. The model update step includes: solving a preset objective function based on the model performance degradation rate, the current data distribution drift intensity, the relative computing load, the system effective computing rate, and the queuing model type to obtain an optimal update frequency; and updating the current semantic codec model with the optimal update frequency to obtain an updated semantic codec model.

2. The method according to claim 1, characterized in that, The generation of the correlation function for the model performance degradation rate includes the following steps: Obtain the baseline performance of the current semantic codec model on the source domain image set, and the measured performance of the current semantic codec model on multiple target domain image sets with different data distribution drift intensities; Based on the baseline performance and the measured performance, calculate the model performance degradation rate corresponding to each of the data distribution drift intensities; A multinomial regression fitting is performed on the data pair of the data distribution drift intensity and the model performance degradation rate to obtain the correlation function of the model performance degradation rate.

3. The method according to claim 1, characterized in that, The calculation process for the relative computing load is shown in the following formula: in, Indicates relative computational load; Indicates the current data distribution drift intensity; This represents the growth factor of the computational workload.

4. The method according to claim 1, characterized in that, The calculation process for the total service time of a single model update is shown in the following formula: in, This indicates the total service time for a single model update; Indicates the number of model parameters; Indicates channel rate; This indicates the baseline calculation delay.

5. The method according to claim 1, characterized in that, The process of solving a preset objective function based on the model performance degradation rate, the current data distribution drift intensity, the relative computational load, the system's effective computational rate, and the queuing model type to obtain the optimal update frequency includes: Select the corresponding normalized queuing cost function based on the queuing model type; The objective function is constructed using the normalized queuing cost function, the model performance degradation rate, the current data distribution drift intensity, the relative computational load, and the system's effective computational rate. The optimal normalized frequency is obtained by minimizing the objective function. The optimal update frequency is calculated and obtained based on the optimal normalized frequency.

6. The method according to claim 5, characterized in that, The process of minimizing the objective function to obtain the optimal normalized frequency includes: If the queuing model type is edge layer M / M / 1, then the objective function is minimized using numerical optimization methods such as gradient descent to obtain the optimal normalized frequency; If the queuing model type is cloud layer D / M / 1, then the bisection method is used to search for congestion factors within a preset interval to minimize the objective function and obtain the optimal normalized frequency.

7. The method according to claim 1, characterized in that, The step of updating the current semantic codec model at the optimal update frequency to obtain the updated semantic codec model includes: The update monitoring steps are performed periodically: based on the optimal update frequency, the corresponding update interval is determined; and the update time difference between the current time and the last model update time is calculated; based on the update interval and the update time difference, it is determined whether to perform a model update; if the update time difference is less than the update interval, the update monitoring steps are continued; if the update time difference is greater than or equal to the update interval, an update operation is performed on the current semantic codec model to obtain the updated semantic codec model.

8. The method according to claim 7, characterized in that, After performing an update operation on the current semantic codec model based on the optimal update frequency to obtain the updated semantic codec model, the method further includes: Deploy the updated semantic codec model to the edge and reset the age state of the drift-aware model.

9. An electronic device, characterized in that, It includes a processor and a memory; when the processor executes a running program stored in the memory, it implements the method for updating the semantic codec model as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for updating the semantic codec model as described in any one of claims 1 to 8.

Images