Artificial intelligence generated content task allocation method, collaborative processing method and device in industrial computing power network

Abstract

This application provides a method, collaborative processing method, and device for allocating AI-generated content tasks in an industrial computing network, relating to the field of digital data processing. The allocation method includes: edge servers performing multiple rounds of federated learning; constructing a training consortium based on the comprehensive evaluation scores of terminal devices; collaboratively training and aggregating a hybrid expert model equipped with a resource-aware gating network; and simultaneously, cyclically executing resource competition. Based on terminal device requests, available resource capacity, and their resource contribution status, a competition mechanism is used to determine the target device and resource allocation scheme, and update the resource contribution status. The updated status serves as a feedback signal, dynamically influencing the comprehensive evaluation of subsequent federated learning, forming a closed-loop collaborative optimization of model training and resource allocation. This application can solve the problems of low performance of AI-generated content service models, poor resource allocation efficiency, and weak system sustainability under conditions of heterogeneous data, privacy sensitivity, and dynamically changing resources in the Industrial Internet of Things (IIoT).

Description

Technical Field

[0001] This application relates to the field of electronic digital data processing technology, and in particular to a method for allocating tasks, a collaborative processing method, and a device for artificial intelligence-generated content in industrial computing networks. Background Technology

[0002] With the deep integration of the Industrial Internet and artificial intelligence technologies, industrial systems are rapidly evolving towards intelligence and autonomy. Against this backdrop, industrial computing networks, as a new generation of industrial infrastructure, provide crucial support for intelligent industrial applications by integrating and coordinating the scheduling of computing power, network resources, and storage resources. Especially with the rapid development of generative artificial intelligence technology, Artificial Intelligence Generated Content (AIGC) is providing powerful content generation and decision support capabilities for key applications such as industrial quality inspection, predictive maintenance of equipment, production process optimization, and digital twin construction. To meet the stringent requirements of real-time performance, reliability, and data privacy in industrial settings, AIGC models are evolving towards lightweight and distributed architectures, gradually moving from the cloud to the edge and terminal sides, forming a three-tiered collaborative computing architecture from cloud to edge to endpoint.

[0003] However, compared to traditional consumer-grade AIGC services, the quality measurement dimensions for AIGC services in industrial computing networks are more complex and stringent. Terminal device data exhibits non-independent and identically distributed characteristics and is privacy-sensitive, making it difficult to train high-performance, high-generalization AIGC models under these constraints using traditional centralized training or simple federated learning. In edge environments with limited and dynamically fluctuating computing and communication resources, traditional static resource allocation methods cannot stably provide high-quality generation services while ensuring low-latency response to AIGC tasks. In distributed collaboration, the lack of effective mechanisms to fairly compensate terminal devices for their data and computing power contributions leads to low node participation and hinders the formation of a healthy and sustainable collaborative ecosystem. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method for allocating tasks for AI-generated content in industrial computing networks, a collaborative processing method, and a device to eliminate or improve one or more defects existing in the prior art.

[0005] The first aspect of this application provides a method for allocating AI-generated content tasks in an industrial computing network, executed by an edge server in the industrial computing network. The allocation method includes: The federated learning process involves multiple iterations, including: constructing a federated learning training consortium based on the comprehensive evaluation scores of each terminal device participating in the federated training within the industrial computing power network; distributing global model parameters for AI-generated content corresponding to the current iteration to each terminal device in the selected consortium; receiving updated global model parameters returned by each terminal device in the training consortium after local training based on a hybrid expert model with a gating network; and aggregating the updated global model parameters based on the comprehensive evaluation scores of each terminal device to obtain the global model parameters corresponding to the next iteration; the comprehensive evaluation score is calculated based on the data distribution feature vector, real-time resource state vector, and historical contribution dynamically updated through contribution feedback signals of the terminal device. A resource contention step is executed cyclically at discrete time intervals. This resource contention step includes: receiving a request from the terminal device for an AI-generated content task; determining the target terminal device to be served, the resource allocation scheme, and the resource consumption quantification index based on all the requests, the currently available edge resource capacity locally, and the real-time resource contribution status of each terminal device that issued the request; sending the resource consumption quantification index to the target terminal device, and updating the resource contribution status of the target terminal device according to the resource consumption quantification index as the latest contribution feedback signal.

[0006] In some embodiments of this application, the construction of the federated learning training alliance based on the comprehensive evaluation scores of each terminal device to participate in federated training in the industrial computing power network includes: Based on the latest contribution feedback signal of each terminal device currently participating in federated training in the industrial computing power network, the historical contribution of each terminal device is dynamically updated. Based on preset weighting coefficients, the real-time resource status vector, data distribution feature vector, and historical contribution of each terminal device are weighted and summed to obtain the comprehensive evaluation score of each terminal device. Based on the comprehensive evaluation scores from high to low, a preset number of terminal devices are selected to form a federated learning training alliance, according to the constraints. The constraints include: the total number of selected terminal devices is less than or equal to the preset upper limit of the alliance size, and the dimensions representing power consumption and computing power in the real-time resource status vectors of the selected terminal devices are both higher than the preset participation threshold.

[0007] In some embodiments of this application, prior to the federated learning step of performing multiple iterations, the method further includes: The incentive strategy information for participating in federated learning is broadcast to each terminal device in the industrial computing network. The incentive strategy information is used to indicate that, in the resource contention step, the data distribution feature vector and real-time resource status vector reported by the terminal device are the basis for determining whether the terminal device can become a target terminal device waiting for local service. The system receives a registration application from a terminal device, wherein the registration application includes the data distribution feature vector and the real-time resource status vector of the terminal device; the data distribution feature vector is calculated by the terminal device based on its local private data and reported in an anonymized form; the real-time resource status vector includes at least one of the following: the remaining battery percentage of the terminal device, the ratio of current available computing power to nominal capacity, and the ratio of current available communication bandwidth to maximum bandwidth.

[0008] In some embodiments of this application, the gating network is configured to dynamically adjust the output weights of the global model (as a global expert) and the local model (as a local expert) in the hybrid expert model based on the real-time resource state vector of the terminal device. Furthermore, the training objectives of the gating network include: minimizing the loss function of the AI-generated content task, while optimizing the energy efficiency and expected latency of the terminal device executing the AI-generated content task.

[0009] In some embodiments of this application, the step of determining the target terminal device to be served, the resource allocation scheme, and the resource consumption quantification index based on all the requests, the currently available edge resource capacity locally, and the real-time resource contribution status of each terminal device that issued the request, and updating the resource contribution status of the target terminal device according to the resource consumption quantification index as the latest contribution feedback signal includes: Determine each of the requests received within the current time interval, wherein each request contains resource requirement information and corresponding original estimates for the AI-generated content task, and the resource requirement information includes: the type of computing resources requested, the quantity of computing resources, and the maximum expected task processing latency. Obtain the currently available edge resource capacity locally, and query the current resource contribution status of the terminal devices that issued each request; wherein, the resource contribution status includes budget status variables and budget consumption status variables; A resource allocation optimization problem is constructed with the goal of maximizing the social welfare of the system, wherein the social welfare of the system is the sum of the social welfare contribution values ​​of each of the terminal devices that issued the request within the current time interval; the social welfare contribution value of each terminal device is obtained by adjusting the original estimate in its request according to the current resource contribution status of the terminal device. Solve the resource allocation optimization problem, and under the constraint of the available edge resource capacity, determine the resource allocation scheme to maximize the social welfare of the system, and determine the terminal device that obtains the resources as the target terminal device according to the resource allocation scheme. Based on the VCG auction mechanism, the resource consumption quantitative indicators to be undertaken by each of the target terminal devices are calculated, and the resource consumption quantitative indicators include: the payment price determined according to the VCG auction mechanism; The payment price is deducted from the budget status variable of the target terminal device, and the budget consumption status variable of the target terminal device is updated according to a preset nonlinear update rule as the latest contribution feedback signal.

[0010] The second aspect of this application provides a collaborative processing method for AI-generated content tasks in an industrial computing network, executed by terminal devices in the industrial computing network. The collaborative processing method includes: The collaborative training process involves multiple iterations, including: when a device is selected into the federated learning training consortium based on a comprehensive evaluation score, it receives global model parameters for AI-generated content distributed by an edge server in the industrial computing network; locally, it uses these global model parameters as a global expert, together with an existing local model as a local expert and a gating network, to form a hybrid expert model; it trains the hybrid expert model using local private data; and after training, it uploads the updated global model parameters to the edge server. An AI-generated content task is generated, and based on the AI-generated content task and the local real-time resource status vector, it is determined whether the AI-generated content task should be offloaded to the edge server; if so, a request for the AI-generated content task is generated and the request is sent to the edge server. If a resource consumption quantification indicator is received from the edge server, the local resource contribution status is updated according to the resource consumption quantification indicator to synchronize with the edge server; wherein, the updated resource contribution status serves as the latest contribution feedback signal and affects the subsequent calculation of the comprehensive evaluation score.

[0011] In some embodiments of this application, prior to the collaborative training step of performing multiple iterations, the method further includes: If, based on the incentive strategy information for participating in federated learning broadcast by the edge server, it is determined to submit a registration application to the edge server, then the bulldozer distance between the local data distribution and the global reference distribution is calculated, and based on the bulldozer distance, a desensitized data distribution feature vector is generated. The terminal device obtains its own real-time resource status vector, wherein the real-time resource status vector includes at least one of the following: the remaining power percentage of the terminal device, the ratio of the current available computing power to the nominal power, and the ratio of the current available communication bandwidth to the maximum bandwidth. A registration application is generated based on the data distribution feature vector and the real-time resource status vector, and the registration application is reported to the edge server.

[0012] In some embodiments of this application, the gating network is configured to dynamically adjust the output weights of the global model (as a global expert) and the local model (as a local expert) in the hybrid expert model based on the real-time resource state vector of the terminal device. Furthermore, the training objectives of the gating network include: minimizing the loss function of the AI-generated content task, while optimizing the energy efficiency and expected latency of the terminal device executing the AI-generated content task.

[0013] A third aspect of this application provides an AI-generated content task allocation device in an industrial computing network, which is installed in an edge server of the industrial computing network. The allocation device includes: The federated learning module is used to execute federated learning steps in multiple iteration rounds. These steps include: constructing a federated learning training consortium based on the comprehensive evaluation scores of each terminal device participating in the federated training within the industrial computing power network; distributing global model parameters for AI-generated content corresponding to the current iteration round to each terminal device in the selected consortium; receiving updated global model parameters returned by each terminal device in the training consortium after local training based on a hybrid expert model with a gating network; and aggregating the updated global model parameters based on the comprehensive evaluation scores of each terminal device to obtain the global model parameters corresponding to the next iteration round; the comprehensive evaluation score is calculated based on the data distribution feature vector, real-time resource state vector, and historical contribution dynamically updated through contribution feedback signals of the terminal device. The resource contention module is used to cyclically execute a resource contention step within discrete time intervals. This resource contention step includes: receiving a request from the terminal device for an AI-generated content task; determining the target terminal device to be served, a resource allocation scheme, and a resource consumption quantification index based on all requests, the currently available edge resource capacity locally, and the real-time resource contribution status of each terminal device that issued the request; sending the resource consumption quantification index to the target terminal device, and updating the resource contribution status of the target terminal device according to the resource consumption quantification index as the latest contribution feedback signal.

[0014] A fourth aspect of this application provides a collaborative processing device for artificial intelligence-generated content tasks in an industrial computing network, which is installed in a terminal device of the industrial computing network. The collaborative processing device includes: The collaborative training module is used to execute a collaborative training step that involves multiple iterations. This collaborative training step includes: when the module is selected into the federated learning training consortium based on a comprehensive evaluation score, it receives global model parameters for AI-generated content distributed by the edge server in the industrial computing power network; locally, it uses the global model parameters as a global expert, and together with an existing local model as a local expert and a gating network, it forms a hybrid expert model; it trains the hybrid expert model using local private data, and after training, it uploads the updated global model parameters to the edge server. The task offloading decision module is used to generate AI-generated content tasks and determine whether to offload the AI-generated content tasks to the edge server based on the AI-generated content tasks and the local real-time resource status vector; if so, it generates a request for the AI-generated content tasks and sends the request to the edge server. The resource contribution synchronization module is used to update the local resource contribution status to synchronize with the edge server based on the resource consumption quantification index received from the edge server; wherein, the updated resource contribution status serves as the latest contribution feedback signal and affects the subsequent calculation of the comprehensive evaluation score.

[0015] The fifth aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the AI-generated content task allocation method in the industrial computing network, or implements the AI-generated content task collaborative processing method in the industrial computing network.

[0016] The sixth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the AI-generated content task allocation method in the industrial computing network, or implements the AI-generated content task collaborative processing method in the industrial computing network.

[0017] The seventh aspect of this application provides a computer program product comprising a computer program that, when executed by a processor, implements the AI-generated content task allocation method in the industrial computing network, or implements the AI-generated content task collaborative processing method in the industrial computing network.

[0018] The method for allocating AI-generated content tasks in an industrial computing network provided in this application involves edge servers executing multiple iterations of federated learning steps. These federated learning steps include: constructing a federated learning training consortium based on the comprehensive evaluation scores of each terminal device participating in the federated training within the industrial computing network; distributing global model parameters for AI-generated content corresponding to the current iteration to each terminal device in the selected consortium; receiving updated global model parameters returned by each terminal device in the training consortium after local training based on a hybrid expert model with a gating network; aggregating the updated global model parameters based on the comprehensive evaluation scores of each terminal device to obtain the global model parameters corresponding to the next iteration; the comprehensive evaluation score is calculated based on the data distribution feature vector, real-time resource state vector, and historical contribution dynamically updated through contribution feedback signals of the terminal device; and cyclically executing the task allocation method within discrete time intervals. The resource content competition step includes: receiving requests from terminal devices for AI-generated content tasks; determining the target terminal device to be served, the resource allocation scheme, and the resource consumption quantification index based on all requests, the currently available edge resource capacity, and the real-time resource contribution status of each terminal device that issued the request; sending the resource consumption quantification index to the target terminal device and updating the resource contribution status of the target terminal device according to the resource consumption quantification index as the latest contribution feedback signal; through the collaborative training of federated learning alliances based on dynamic contribution feedback and resource-aware hybrid expert models, the impact of non-independent and identically distributed data can be effectively mitigated while protecting data privacy, and the generation quality and accuracy of AIGC models in industrial scenarios can be significantly improved; by modeling resource allocation as an online competition process linked to model value, efficient and adaptive scheduling of computing resources can be achieved. In resource-constrained dynamic environments, it can simultaneously reduce task dropout rate and average latency, thereby ensuring a high-quality and stable service experience; by feeding back the resource allocation results to the model training stage through the resource contribution status, high-quality contributions can be incentivized. This creates a virtuous cycle of contribution, reward, performance improvement, and better service, ensuring the long-term vitality and self-optimization capabilities of industrial computing networks.

[0019] 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.

[0020] 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

[0021] 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 the task allocation method for artificial intelligence-generated content in an industrial computing network according to an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of a second process for allocating AI-generated content tasks in an industrial computing network according to an embodiment of this application.

[0023] Figure 3 This is a schematic diagram of the first process of a collaborative processing method for AI-generated content tasks in an industrial computing network according to an embodiment of this application.

[0024] Figure 4 This is a schematic diagram of a second process of a collaborative processing method for AI-generated content tasks in an industrial computing network according to an embodiment of this application.

[0025] Figure 5 This is a schematic diagram of the architecture of an AI-generated content task allocation system in an industrial computing network, which is an application example of this application.

[0026] Figure 6 This is a schematic diagram illustrating the framework of the MoE-based federated learning fine-tuning process, which is an application example of this application.

[0027] Figure 7 This is a flowchart illustrating the federated learning fine-tuning process in an application example of this application.

[0028] Figure 8(a) is a schematic diagram comparing the average latency evaluation of various algorithms under different numbers of devices; Figure 8(b) is a schematic diagram comparing the average energy consumption of various algorithms under different numbers of devices; Figure 9(a) is a schematic diagram comparing the average latency evaluation of various algorithms under different edge nodes; Figure 9(b) is a schematic diagram comparing the average energy consumption of various algorithms under different edge nodes; Figure 10(a) is a schematic diagram comparing the user satisfaction index of EOMA performance in different rounds; Figure 10(b) is a schematic diagram comparing the budget utilization rate of EOMA performance in different rounds; Figure 11(a) is a schematic diagram comparing the overall social welfare and number of users of EOMA, SPS and RAN; Figure 11(b) is a schematic diagram comparing the overall social welfare of EOMA, SPS and RAN with the number of bid sets; Figure 12(a) is a schematic diagram illustrating the effect of the total cost of the iterative optimization algorithm; Figure 12(b) is a schematic diagram illustrating the overall utility of the iterative optimization algorithm; Figure 13 This is a schematic diagram illustrating the effect of the iterative optimization algorithm on the target value. Detailed Implementation

[0029] 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.

[0030] 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.

[0031] 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.

[0032] 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.

[0033] 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.

[0034] In one or more embodiments of this application, Artificial Intelligence Generated Content (AIGC) refers to a technology that automatically generates content such as images and text using generative artificial intelligence models (such as diffusion models). Industrial Internet of Things (IIoT) refers to a network system in an industrial environment where devices, sensors, and control systems exchange data and are intelligently controlled via the internet. Federated Learning (FL) is a distributed machine learning paradigm that allows terminal devices to train models locally, uploading only model parameters rather than raw data to protect data privacy. Mixture of Experts (MoE) is a model architecture that uses multiple "expert" sub-models and a gated network to work collaboratively to adapt to different types of data and tasks. Non-Independent Identical Distribution (Non-IID) refers to inconsistent data distribution across different devices, commonly seen in industrial scenarios due to data bias caused by device heterogeneity and environmental differences. Earth Mover's Distance (EMD) measures the difference between two probability distributions and is often used to assess data quality similarity. JADE (adaptive differential evolution with optional external archive) is an evolutionary algorithm that enhances global optimization capabilities by introducing parameter adaptation and archive-based mutation strategies.

[0035] It's important to note that industrial AIGC services are determined not only by network, computing, and storage resources, but also by data quality, the precise fit between the model and the business, and the determinism of the generated results. The industrial environment is highly time-varying, complex, and strongly business-dependent: terminal devices (such as industrial cameras and sensors), due to their limited storage resources, need to constantly update locally stored industrial image data, leading to dynamic changes in data distribution. Simultaneously, the flexible addition, removal, or failure of equipment on the production line can cause drastic changes in data sources and computing nodes, resulting in instability in AIGC services (such as defect detection image generation and digital twin scenario construction). This can cause the generated content to fail to meet the personalized and high-quality requirements of specific businesses in terms of clarity, completeness of key features, or style.

[0036] Therefore, deploying data-intensive and compute-intensive AIGC services in resource-constrained, latency-sensitive industrial IoT edge and terminal environments faces three major challenges: (1) Resource constraints: Industrial Internet of Things (IIoT) terminal devices generally suffer from weak computing power, small storage space, and limited battery energy, making it difficult to run large-scale AIGC models independently. More importantly, the edge node resources in industrial sites are not static, and their availability changes dynamically with the number and priority of parallel tasks, which renders static resource allocation schemes ineffective.

[0037] (2) Data heterogeneity and privacy: Industrial data naturally exhibits non-independent and identically distributed characteristics. Data collected from different production lines, different equipment, and even different time periods show significant differences in feature distribution. This heterogeneity severely impairs the generalization ability of models trained on centralized data. Furthermore, due to strict industry privacy regulations (such as GDPR) and considerations of trade secrets, data cannot be centrally processed in the cloud outside the factory boundary. Currently, most mainstream lightweight edge models are heterogeneous, and independent fine-tuning on different Non-IID data can lead to model drift, making it difficult to aggregate into a global high-performance model.

[0038] (3) Difficulty in balancing service quality and efficiency: Industrial AIGC tasks have dual high requirements for generation quality and response latency. For example, real-time defect detection requires that the generated comparison images reach extremely high fidelity in a very short time in order to make accurate comparisons. In a dynamically changing network environment and fluctuating resource environment, there is an inherent trade-off between quality and efficiency, making it difficult to stably and predictably provide industrial-grade services that are both high-quality and low-latency.

[0039] In a cloud-edge collaborative industrial internet data analysis and processing system, IoT data processing technology is used to maintain a global model in the cloud to analyze the security of edge device data. In this system, edge devices execute data processing tasks locally through federated learning, while the cloud server includes a model aggregation module, a data storage module, a data analysis module, and a memory for storing computer programs. The model aggregation module is responsible for updating model parameters, the data storage module saves the model parameters and training records of the edge devices, and the data analysis module runs computer programs to assess and analyze data security, effectively removing malicious or harmful data, ensuring database security, and improving the efficiency and accuracy of data processing through distributed data processing and model sharing learning, while reducing data transmission and computation costs.

[0040] However, despite employing federated learning to protect data privacy, the system does not completely resolve the issues of data heterogeneity and high communication overhead. Differences in data distribution across different clients can lead to poor model convergence, impacting the performance of the final global model. While cloud-edge collaborative systems improve data processing efficiency, they rely heavily on computing resources, especially for maintaining and updating the global model in the cloud, which places high demands on hardware and increases costs. Furthermore, federated learning methods struggle to flexibly adapt to different application scenarios when dealing with highly heterogeneous industrial data, limiting the model's generalization ability.

[0041] Another big data processing and decision support system based on an industrial internet AI big data model is specifically designed to provide decision support for enterprises. This system includes modules for industrial data acquisition, preprocessing, storage, and management, as well as the ability to train and validate industrial datasets. Its unique feature is its ability to generate new decision support systems tailored to the specific needs of enterprises. Specifically, enterprise users can call upon pre-trained industrial internet AI big data models to analyze and predict their own industrial data, thereby obtaining a decision support system that meets their specific needs and completing the updating and iteration process of industrial data.

[0042] However, while the system provides powerful AI models for data analysis and decision support, in practical applications, the need for large amounts of high-quality data for training may expose enterprises to the risk of data privacy breaches. Industrial Internet AI models require a massive number of parameters and computational resources, typically necessitating training on well-resourced cloud or edge devices, which poses a significant burden for small and medium-sized enterprises (SMEs). Centralized large models struggle to cope with the highly dynamic and diverse data demands of industrial environments, and a single centralized model cannot meet the real-time decision-making needs of all scenarios.

[0043] Based on this, to address the problems of existing methods such as model performance bottlenecks under data heterogeneity and privacy protection, difficulty in balancing service quality and efficiency under dynamic resource competition, and lack of system sustainability and participation incentives, this application provides, respectively, a method for allocating AI-generated content tasks in an industrial computing network, a device for allocating AI-generated content tasks in an industrial computing network to execute the method, a method for collaborative processing AI-generated content tasks in an industrial computing network, a device for collaborative processing AI-generated content tasks in an industrial computing network to execute the method, an electronic device, a computer-readable storage medium, and a computer program product. This constructs a three-layer collaborative resource allocation system for AIGC services in an industrial internet environment, encompassing cloud, edge servers, and terminal devices. This system deeply integrates three mechanisms: model collaborative training, market incentives, and cross-layer resource scheduling, aiming to achieve global optimization of model performance, participation willingness, and system efficiency. The system forms a four-dimensional collaborative closed loop in the horizontal direction, from perception, communication, computing to intelligence, and achieves seamless connection from data perception to intelligent generation. In the vertical direction, it builds a comprehensive autonomous optimization system from underlying physical resources to top-level application requirements, from resources, models, services to value, to ensure that AIGC services can continuously, stably and efficiently meet high-quality personalized business needs in dynamic and complex industrial environments.

[0044] The following examples will provide a detailed description.

[0045] Based on this, this application provides a method for allocating AI-generated content tasks in an industrial computing network, which can be implemented by an AI-generated content task allocation device in the industrial computing network. The AI-generated content task allocation device is located in an edge server within the industrial computing network and can function as a functional module within the edge server. (See also...) Figure 1 The method for allocating AI-generated content tasks in the industrial computing network specifically includes the following: Step 100: Execute a federated learning step with multiple iterations, which includes: constructing a federated learning training alliance based on the comprehensive evaluation scores of each terminal device to be federated in the industrial computing power network, and distributing global model parameters for AI-generated content corresponding to the current iteration to each terminal device in the selected alliance; receiving updated global model parameters returned by each terminal device in the training alliance after local training based on a hybrid expert model with a gating network, and aggregating the updated global model parameters based on the comprehensive evaluation scores of each terminal device to obtain the global model parameters corresponding to the next iteration; the comprehensive evaluation score is calculated based on the data distribution feature vector, real-time resource state vector, and historical contribution dynamically updated through contribution feedback signals of the terminal device.

[0046] It should be noted that the task allocation method for AI-generated content in industrial computing networks can be centered on edge servers deployed within the industrial computing network (such as factory parks or factory areas) as the core coordinator and executor. These edge servers connect the cloud to a vast number of terminal devices (such as industrial cameras, sensors, and robotic arms). These terminal devices are the producers of the raw data and the initiators of AIGC tasks.

[0047] In step 100, each terminal device to participate in federated training can be a terminal device that has already submitted a registration application to the edge server. The global model parameters for AI-generated content can be called the model parameters of the global model corresponding to the AI-generated content service model or the pre-trained model for AIGC tasks (such as the diffusion model for defect detection).

[0048] The comprehensive evaluation score is a quantitative indicator dynamically calculated by the edge server for each terminal device. It is used to comprehensively evaluate the device's value in participating in federated learning and is also a core criterion for subsequent consortium selection. Based on the comprehensive evaluation score, the edge server selects a subset of terminal devices from a large pool to form the current round of federated learning training consortium. This score is dynamically calculated and primarily determined based on three dimensions: (1) Data distribution feature vector: A desensitized mathematical vector used to characterize the statistical properties (such as mean, variance, and distribution shape) of the local dataset at the terminal, especially reflecting its difference (similarity or uniqueness) from the global data distribution. It can be generated by calculating metrics such as bulldozer distance (EMD) and is a key input for evaluating data quality and uniqueness.

[0049] (2) Real-time resource status vector: This is a vector reflecting the current physical resource availability of the terminal device, representing the current physical resource status of the terminal, such as remaining battery power, available computing power, network bandwidth, etc. Its typical dimensions include: remaining battery percentage, the ratio of available computing power to maximum capacity (CPU / GPU utilization), the ratio of available communication bandwidth to maximum bandwidth, etc. This vector is key to determining whether the device can stably participate in training or execute tasks.

[0050] (3) Historical Contribution: A dynamically updated scalar value used to record the long-term contribution trajectory of the terminal device to the federated learning community. Its value is not fixed, but is iteratively updated based on the contribution feedback signal from the resource competition step, thereby transforming the device's economic behavior in the resource market into its reputation capital in the model training community. The contribution feedback signal is the core closed-loop medium of this application. Essentially, it is an update event or the updated state value of the resource contribution status. This signal is generated in the resource competition step and is transmitted to the federated learning step to update the historical contribution of the terminal device, thereby realizing the value transmission from resource consumption to the influence of model training.

[0051] The edge server distributes the currently iterated, lightweight global model parameters to terminals within the consortium. Each terminal device trains locally using its private data in a hybrid expert model (MoE) architecture. The MoE consists of the distributed global model parameters (as global experts), the terminal's own local model parameters (as local experts), and a gating network. The gating network is responsible for intelligently fusing the outputs of the two experts.

[0052] After completing the training for the current iteration, each terminal only encrypts and uploads the updated global model parameters from the global expert to the edge server, while the parameters from the local expert and the gated network remain locally, thus effectively protecting data privacy and personalized knowledge. The edge server normalizes the comprehensive evaluation scores of each terminal device to obtain the aggregate weights corresponding to each terminal device. Based on the aggregate weights corresponding to each terminal device, a weighted average is calculated on the updated global model parameters received from each terminal device to generate the global model parameters for the next iteration, thereby completing one round of model evolution. If the global model has converged or reached the preset maximum number of iterations, federated training can be stopped, and the latest global model parameters for the next iteration can be used as the final global model parameters.

[0053] It should be noted that the gating network is a key control module in the hybrid expert model. In this application, the gating network is specifically designed to be resource-aware, meaning that its decision (weight allocation) depends not only on the characteristics of the input data but also integrates the real-time resource state vector of the terminal, thereby enabling it to adaptively select more energy-efficient or faster inference paths based on factors such as device power and computing power.

[0054] Step 200: Execute a resource contention step cyclically within discrete time intervals. This resource contention step includes: receiving a request from the terminal device for an AI-generated content task; determining the target terminal device to be served, the resource allocation scheme, and the resource consumption quantification index based on all the requests, the currently available edge resource capacity locally, and the real-time resource contribution status of each terminal device that issued the request; sending the resource consumption quantification index to the target terminal device, and updating the resource contribution status of the target terminal device according to the resource consumption quantification index as the latest contribution feedback signal.

[0055] In step 200, the resource contention step is triggered cyclically at discrete time intervals (e.g., every second), aiming to dynamically allocate edge computing resources for AIGC tasks arriving in real time. The edge server receives requests from terminal devices. These requests are for specific AI content generation tasks (e.g., generating a defective image for comparison). Based on all current requests, its own currently available edge resource capacity (e.g., number of idle CPU cores, memory), and the real-time resource contribution status of each requesting terminal, the edge server runs a competitive decision-making mechanism. This status is an internal system variable that records the historical resource consumption and contribution of each terminal. Through this competitive mechanism, the target terminal device that can obtain services within the current time interval, the specific resource allocation scheme (e.g., allocating 20% ​​of GPU computing power to a task), and the corresponding resource consumption quantification index (i.e., price or cost) is calculated.

[0056] The edge server then notifies the target terminal of the quantified metrics and updates the terminal's resource contribution status accordingly. This updated status becomes the latest contribution feedback signal. This signal will be fed back into the next round of federated learning to update the device terminal's historical contribution, thereby affecting its future priority and weight in training, achieving a closed-loop incentive and optimization from the resource market to model training.

[0057] In the resource competition process, the resource consumption quantification index is a quantified value representing the cost that the target terminal device must pay for the services it receives. In embodiments employing the VCG auction mechanism, this resource consumption quantification index is the payment price calculated according to the VCG rules. It serves as the direct basis for updating the resource contribution status.

[0058] The resource allocation scheme includes the types and quantities of computing resources allocated to each target terminal device. The types of computing resources can include multiple or all of the following: CPU cores / time slices, computing power of AI accelerators such as GPUs or Tensor Processing Units (TPUs), RAM capacity, I / O bandwidth of storage (such as SSDs), time usage of dedicated hardware accelerators (such as NPUs and FPGAs), and instance quotas for containers or virtual machines. The specific allocation can be set according to actual application needs. The quantity of computing resources refers to the amount or proportion of a certain type of computing resource to be allocated to a terminal device, such as: occupying 20% ​​of the GPU computing power or 50% of the CPU cores of an edge server. Alternatively, allocating 2 GPU cores or 4GB of dedicated memory. Or, allocating 5 units of computing resources, where 1 unit is defined by the system.

[0059] The resource contribution status is a system state variable that is synchronously maintained between the edge server and the terminal device. It quantifies and records the resource consumption and contribution accumulated by the terminal in the resource competition market. In the embodiments of this application, it is specifically manifested as a budget status variable (such as remaining points) and a budget consumption status variable (an auxiliary variable reflecting the consumption rate), which is the direct basis for calculating contribution feedback signals and adjusting device competitiveness.

[0060] It is understandable that the federated learning step in step 100 and the resource competition step in step 200 do not operate in isolation. The contribution feedback signal acts as a connecting link, enabling the performance (consumption and contribution) of the device terminal in the resource market to affect its status and benefits in the model collaboration community. This incentivizes the terminal to continuously provide high-quality data and compete for resources reasonably, ultimately driving the collaborative and autonomous continuous optimization of AIGC model performance and overall system resource utilization efficiency in a dynamic industrial environment.

[0061] As described above, the AI-generated content task allocation method in the industrial computing network provided in this application, through the construction of a federated learning consortium based on dynamic contribution feedback and the collaborative training of a resource-aware hybrid expert model, can effectively mitigate the impact of non-independent and identically distributed data while protecting data privacy. This significantly improves the generation quality and accuracy of AIGC models in industrial scenarios. By modeling resource allocation as an online competition process linked to model value, efficient and adaptive scheduling of computing resources can be achieved. In resource-constrained dynamic environments, it can simultaneously reduce task dropout rates and average latency, thereby ensuring a high-quality and stable service experience. Feeding resource allocation results back to the model training stage through resource contribution status incentivizes high-quality contributions. This forms a virtuous cycle of contribution, reward, performance improvement, and better service, ensuring the long-term vitality and self-optimization capabilities of the industrial computing network.

[0062] To further address the issues of traditional federated learning's random or data-volume-based client selection failing to incentivize continuous contribution and differentiate contribution quality in industrial scenarios, and to ensure training alliance quality in heterogeneous resource and data environments, this application provides a method for allocating AI-generated content tasks in an industrial computing network. The specific implementation process of constructing a federated learning training alliance based on the comprehensive evaluation scores of each terminal device participating in federated training within the federated learning step of this method is as follows: Step 110: Based on the latest contribution feedback signal of each terminal device currently participating in federated training in the industrial computing power network, dynamically update the historical contribution of each terminal device.

[0063] Step 120: Based on the preset weighting coefficients, the real-time resource status vector, data distribution feature vector, and historical contribution of each terminal device are weighted and summed to obtain the comprehensive evaluation score of each terminal device.

[0064] Step 130: Select a preset number of terminal devices to form a federated learning training alliance based on the comprehensive evaluation scores in descending order, according to the constraints. The constraints include: the total number of selected terminal devices is less than or equal to the preset upper limit of the alliance size, and the dimensions representing power consumption and computing power in the real-time resource status vectors of the selected terminal devices are both higher than the preset participation threshold.

[0065] In one example, the edge server obtains the real-time status of each terminal (such as industrial camera N1 and sensor N2): N1 has 80% battery, 60% available computing power, and a data feature vector (calculated using EMD with a similarity of 0.2 to the global distribution); N2 has 40% battery, 30% available computing power, and a similarity of 0.8. Simultaneously, the edge server queries the historical contribution of both: N1 has a high contribution due to its active bidding in the past; N2 has a low contribution. The edge server reads the latest round of resource auction results: N1 won the bid and paid the price, while N2 did not. Based on this, the server increases N1's historical contribution, while N2 remains unchanged.

[0066] The preset weighting coefficients are: data similarity weight 0.5, resource status weight 0.3, and historical contribution weight 0.2. Then, the comprehensive evaluation score is calculated. The overall assessment score for N1 is calculated as follows: 0.5 × (1 - 0.2) + 0.3 × (0.8 + 0.6) / 2 + 0.2 × 0.9 = 0.73. The overall evaluation score for N2 is calculated as follows: 0.5 × (1 - 0.8) + 0.3 × (0.4 + 0.3) / 2 + 0.2 × 0.1 = 0.22. The alliance size is set to a maximum of 5, with participation thresholds (battery power > 50%, computing power > 40%). After sorting by score, N1, which satisfies all constraints, is selected; N2 is excluded because its battery power is below the threshold.

[0067] To further address the cold start problem of proactively attracting and initiating the participation of the first batch of high-quality devices, and the problem of obtaining device evaluation information under privacy constraints, this application provides a method for allocating AI-generated content tasks in an industrial computing network, as described in the embodiments of this application. (See also...) Figure 2 The specific implementation process before step 100 in the AI-generated content task allocation method in the industrial computing network is as follows: Step 010: Broadcast incentive strategy information for participating in federated learning to each terminal device in the industrial computing network; the incentive strategy information is used to indicate that: in the resource contention step, the data distribution feature vector and real-time resource status vector reported by the device terminal are the basis for determining whether the terminal device can become a target terminal device waiting for local service.

[0068] Understandably, upon receiving the broadcast, a terminal device (such as a smart camera) intending to participate will initiate local computation: This involves statistically analyzing the locally stored defect detection image dataset, calculating its color and texture distribution histograms, and using the Bulldozer Distance (EMD) algorithm to compare it with the previously released "typical good product features" from the server, generating a "data distribution feature vector" (e.g., [similarity: 0.3, diversity: 0.8]) representing the data's uniqueness and value. This computation is performed locally on the terminal device, and the original image data is never transmitted externally. Then, the terminal device reads the current battery level (e.g., 65%), current CPU utilization (which is used to calculate the available computing power ratio = 40%), and the estimated signal strength and bandwidth of the current network connection (e.g., available bandwidth ratio = 70%) through the device operating system interface. It then encapsulates the calculated feature vector and the collected state vector, encrypting them using a lightweight encryption algorithm (e.g., AES) agreed upon with the server, forming a registration application data packet.

[0069] Step 020: Receive a registration application from a terminal device, wherein the registration application includes the data distribution feature vector and the real-time resource status vector of the terminal device; the data distribution feature vector is calculated by the terminal device based on its local private data and reported in an anonymized form; the real-time resource status vector includes at least one of the following: the remaining battery percentage of the terminal device, the ratio of current available computing power to nominal power, and the ratio of current available communication bandwidth to maximum bandwidth.

[0070] After receiving the encrypted registration request, the edge server decrypts and verifies it. The edge server parses out the data distribution feature vector and real-time resource status vector, binds them with the device's unique identifier (such as Device ID), and stores them in the candidate device resource pool database as the original basis for subsequent comprehensive evaluation and alliance selection.

[0071] Among them, incentive strategy information is a rule-based statement actively released by the edge server to attract and guide terminal devices to participate in collaborative learning; data distribution feature vector is a mathematical vector that reflects the overall distribution of the dataset after the terminal device quantifies the statistical characteristics of its local private dataset; reporting in an anonymized form means that before uploading data, the terminal device has removed content that can be directly identified or restored to the original sensitive information through specific processing, and only submits features or statistical results that cannot be inferred from the original data; registration application refers to a structured request submitted by the terminal device to the edge server in response to the incentive broadcast, indicating its willingness to participate and containing its own quantitative characteristics.

[0072] To further address the problem of static models being unable to adapt to dynamic resource environments and the contradiction between model performance and the difficulty in coordinating optimization of device energy efficiency and real-time performance, in an industrial computing network AI-generated content task allocation method provided in this application embodiment, the gating network is configured to: dynamically adjust the output weights of the global model (as a global expert) and the local model (as a local expert) in the hybrid expert model according to the real-time resource state vector of the terminal device; and the training objective of the gating network includes: minimizing the loss function of the AI-generated content task, while optimizing the energy efficiency and expected latency of the terminal device executing the AI-generated content task.

[0073] Specifically, the gating network calculates and assigns different weights to each input sample to determine which expert model(s) will process that sample, which is the core of the model's flexible computation. Dynamically adjusting output weights means that the weights calculated by the gating network for each model inference are not fixed, but are recalculated and optimized in real-time and online based on the device's immediate status (e.g., battery power, computing power) and task requirements (e.g., latency) during the inference. The global model, acting as a global expert, is a general-purpose model distributed uniformly by the edge server and shared among all participating devices. It absorbs extensive data knowledge, has strong capabilities, but typically has a higher computational cost. The local model, acting as a local expert, refers to a personalized model held by the terminal device itself, primarily trained or fine-tuned using local private data. It is more adapted to the specific scenario of the device and is usually lighter and more efficient. The training objective is the optimization direction and criteria that guide the machine learning model (here referring to the gating network) in updating parameters, defining what goals the model needs to learn to achieve. Energy efficiency and expected latency are additional metrics introduced to evaluate the computation process itself, besides traditional task performance metrics (e.g., accuracy). Energy efficiency can refer to the energy consumed to complete a unit of calculation; expected latency can refer to the estimated time required to complete inference under given resource conditions.

[0074] In this process, when a local hybrid expert model on a terminal device (such as a smart camera) needs to perform intelligent analysis on an industrial image, the gating network in the model receives two inputs simultaneously: one is the feature information of the image itself, and the other is a real-time resource status report of the device (e.g., battery power is only 25%, available computing power is weak, and the analysis task requires extremely low latency). The gating network first calculates initial expert weights based on the image content, determining which expert should be relied upon more under normal circumstances. Then, it dynamically adjusts these initial weights based on the real-time resource status. For example, when extremely low battery power is detected, the gating network automatically reduces the weight of the global expert, which consumes more computing power, while increasing the weight of the more energy-efficient local expert; if the task latency requirement is very urgent, it reverses this process, prioritizing processing speed. After these adjustments, the gating network outputs a final set of normalized weight values. This set of weights determines the proportion of contribution from the global model (representing general knowledge) and the local model (representing personalized knowledge) in this image analysis task, thus achieving intelligent inference that adapts to hardware status.

[0075] It's important to note that when terminal devices participate in federated learning for local model training, the training of the gating network doesn't solely focus on task accuracy. Its training objective is a composite one, comprised of three loss components: the first is the core task loss ensuring the quality of generated content (e.g., image clarity, recognition accuracy); the second is an energy efficiency loss related to device power consumption, designed to encourage the gating network to make more energy-efficient expert choices; and the third is a latency loss related to processing speed, designed to encourage it to meet the real-time requirements of the task.

[0076] Through backpropagation, the parameters of the gating network are updated simultaneously in the direction of reducing the total loss. This enables it not only to learn how to select experts based on image content, but also to learn how to make trade-offs between saving power or speeding up when resources are scarce, ultimately learning a decision-making strategy that achieves the optimal balance between quality, energy efficiency, and speed in complex resource environments.

[0077] To further address the contradiction between short-term fairness and system efficiency in dynamic resource competition, and the challenge of online budget constraint management in a distributed incentive environment, this application provides a method for allocating AI-generated content tasks in an industrial computing network. (See also...) Figure 2 The resource content competition step in step 200 of the artificial intelligence-generated content task allocation method in the industrial computing network is specifically implemented as follows: Step 210: Determine each of the requests received within the current time interval, wherein each request contains resource requirement information and corresponding original estimates for the AI-generated content task, and the resource requirement information includes: the type of computing resources requested, the quantity of computing resources, and the maximum expected task processing latency.

[0078] Step 220: Obtain the currently available edge resource capacity locally, and query the current resource contribution status of the terminal devices that issued each request; wherein, the resource contribution status includes budget status variables and budget consumption status variables.

[0079] Step 230: Construct a resource allocation optimization problem with the goal of maximizing system social welfare, wherein the system social welfare is the sum of the social welfare contribution values ​​of each of the terminal devices that issued the request within the current time interval; the social welfare contribution value of each terminal device is obtained by adjusting the original valuation in its request according to the current resource contribution status of the terminal device.

[0080] Step 240: Solve the resource allocation optimization problem. Under the constraint of the available edge resource capacity, determine the resource allocation scheme to maximize the social welfare of the system, and determine the terminal device that obtains the resources as the target terminal device according to the resource allocation scheme.

[0081] Step 250: Based on the VCG auction mechanism, calculate the quantitative indicators of resource consumption to be undertaken by each of the target terminal devices, wherein the quantitative indicators of resource consumption include the payment price determined according to the VCG auction mechanism.

[0082] Step 260: Deduct the payment price from the budget status variable of the target terminal device, and update the budget consumption status variable of the target terminal device according to a preset nonlinear update rule as the latest contribution feedback signal.

[0083] Among them, resource demand information refers to the specific description of the edge resources required by the terminal device to complete its AIGC task, as declared in the request, and serves as the basis for quantifying resource occupancy; original valuation refers to the value that the terminal device, when submitting the request, will bring to itself if the task can obtain the required resources and be successfully completed. It is the basis for the device's bidding in resource competition; budget status variable refers to one of the components of resource contribution status, which can be compared to the device's electronic currency balance, representing the amount it can use to pay for resource services; budget consumption status variable refers to one of the components of resource contribution status, a variable calculated non-linearly and cumulatively based on historical payment data, reflecting the degree or rate of consumption of the device's budget, used to dynamically adjust its bidding competitiveness; system social welfare is the sum of the adjusted valuations (social welfare contribution values) of all participating terminal devices within a decision-making cycle. The goal of resource allocation is to maximize this sum, representing the optimal overall system efficiency; social welfare contribution value is the potential contribution of a single terminal device to the system's social welfare, dynamically calculated by the server based on the device's original valuation and its current resource contribution status (especially budget consumption status), and is an effective valuation used for optimization decisions; VCG auction mechanism is a specific auction rule. The price paid by the winner is not their own bid, but rather the reduction in the total welfare of other competitors caused by their winning action. This mechanism incentivizes participants to report their true valuations. Nonlinear update rules are mathematical rules used to update budget consumption state variables. Their characteristic is that the update magnitude and the payment price are not simply linearly proportional, but are carefully designed to ensure that long-term budget constraints are automatically met.

[0084] Specifically, within a discrete time window (e.g., 1 second), the edge server receives AIGC task offloading requests from multiple terminal devices (e.g., intelligent robotic arms, quality inspection cameras). The server parses each request packet, extracting key information: resource requirements (e.g., requiring GPU-type resources, 2 computing units, and requiring completion within 100 milliseconds) and initial valuation (e.g., the device declares that successful task completion is worth 10 virtual credits). The server queries its managed edge computing resource pool to confirm the current available resource capacity (e.g., the total remaining available GPU computing power is 10 units). The server queries its maintained database to obtain the resource contribution status of each requesting device. This status includes two parts: a budget status variable (similar to an account balance, indicating how many credits the device has left to pay) and a budget consumption status variable (an indicator reflecting its historical consumption level; the higher the consumption, the larger the value).

[0085] The server constructs a real-time optimization problem: the goal is to maximize the total social welfare contribution of all requesting devices within a constraint of no more than 10 GPU units of computing power. Key computation: The social welfare contribution of each device is not directly calculated using the original estimate (10 points) from its request. The server dynamically adjusts it based on the device's budget consumption state variable. For example, for a device with already high budget consumption, its original estimate will be lowered to prevent excessive resource consumption and ensure long-term fairness.

[0086] The server uses optimization algorithms (such as linear programming relaxation and rounding) to solve for an optimal resource allocation scheme (e.g., allocating 2 units to device A, 3 units to device B, etc.) and determines the set of target terminal devices (winning bidders) accordingly. For each winning target terminal device, the server executes the VCG auction rules to calculate the price it should pay. This price is not equal to its bid, but rather equal to the social welfare loss caused by excluding other devices due to its victory. This ensures the authenticity of the bids. The server performs liquidation: deducting the payment price from the budget state variable (account balance) of each winning device. Subsequently, the server updates the budget consumption state variable of the device according to a preset non-linear formula based on the payment price. This update increases the state variable value of devices with high consumption, thereby further reducing their bidding competitiveness in the future, achieving automatic budget management. This updated state variable is the latest generated contribution feedback signal.

[0087] This application provides a method for collaborative processing of AI-generated content tasks in an industrial computing network, which can be implemented by an AI-generated content task collaborative processing device in the industrial computing network. The AI-generated content task collaborative processing device is installed in a terminal device within the industrial computing network and can function as a functional module within the terminal device. See [link to relevant documentation]. Figure 3 The collaborative processing method for AI-generated content tasks in the industrial computing network specifically includes the following: Step 300: Perform a collaborative training step involving multiple iterations. This collaborative training step includes: when selected into the federated learning training consortium based on a comprehensive evaluation score, receiving global model parameters for AI-generated content distributed by the edge server in the industrial computing power network; using the global model parameters as a global expert locally, and combining them with an existing local model as a local expert and a gating network to form a hybrid expert model; training the hybrid expert model using local private data; and uploading the updated global model parameters to the edge server after training is completed.

[0088] Step 400: Generate an AI-generated content task, and determine whether to offload the AI-generated content task to the edge server based on the AI-generated content task and the local real-time resource status vector; if so, generate a request for the AI-generated content task and send the request to the edge server.

[0089] Step 500: If a resource consumption quantification index is received from the edge server, the local resource contribution status is updated according to the resource consumption quantification index to synchronize with the edge server; wherein, the updated resource contribution status serves as the latest contribution feedback signal and affects the subsequent calculation of the comprehensive evaluation score.

[0090] In other words, the terminal device constructs the MoE by receiving and fusing the global expert model distributed from the edge with its own lightweight local expert model. This allows it to leverage the powerful capabilities of the global model while maintaining local personalization and low energy consumption. Simultaneously, the terminal device directly participates in the optimization of the global model by contributing local data to train this hybrid model and uploading parameters. The terminal device is endowed with autonomous decision-making capabilities. Based on task characteristics and its own real-time resource state vector (such as current power consumption and computing load), it dynamically assesses the risks and probabilities of local execution. This intelligent offloading decision-making mechanism enables the device to save costs by relying on its own resources when resources are abundant, and to decisively seek assistance when resources are depleted or the task is urgent, ensuring the reliability and timeliness of task completion. The terminal device maintains a local resource contribution status, which is synchronously updated based on quantified indicators of resource consumption (such as payment prices) received from the server. This status is essentially a unified, quantified contribution and consumption profile. Its update signifies that the device has confirmed the market transaction, and this latest status will be fed back to the system as a contribution feedback signal, directly affecting its future comprehensive evaluation score (i.e., its reputation and value in federated learning). This solves the problem of how to rationally manage resource consumption and contribution for long-term benefits in the absence of a central coordinator for terminal devices.

[0091] To further achieve a reliable initial assessment of data value under strict privacy constraints, and to realize the standardization of device resources and reliable state awareness in a highly heterogeneous hardware environment, this application provides a collaborative processing method for AI-generated content tasks in an industrial computing network, see [link to relevant documentation]. Figure 4 The method for collaborative processing of AI-generated content tasks in the industrial computing network also includes the following steps prior to step 300: Step 030: If, based on the incentive strategy information for participating in federated learning broadcast by the edge server, it is determined that a registration application should be submitted to the edge server, then the bulldozer distance between the local data distribution and the global reference distribution is calculated, and based on the bulldozer distance, a desensitized data distribution feature vector is generated.

[0092] Step 040: Obtain the real-time resource status vector of the terminal device, wherein the real-time resource status vector includes at least one of the following: the remaining power percentage of the terminal device, the ratio of the current available computing power to the nominal power, and the ratio of the current available communication bandwidth to the maximum bandwidth.

[0093] Step 050: Generate a registration application based on the data distribution feature vector and the real-time resource status vector, and report the registration application to the edge server.

[0094] To further address the problem of static models being unable to adapt to dynamic resource environments and the contradiction between model performance and the difficulty in coordinating optimization of device energy efficiency and real-time performance, in an industrial computing network collaborative processing method for AI-generated content tasks provided in this application embodiment, the gated network is configured to: dynamically adjust the output weights of the global model (as a global expert) and the local model (as a local expert) in the hybrid expert model according to the real-time resource state vector of the terminal device; and the training objective of the gated network includes: minimizing the loss function of the AI-generated content task, while optimizing the energy efficiency and expected latency of the terminal device executing the AI-generated content task.

[0095] From a software perspective, this application also provides an industrial computing network AI-generated content task allocation device for executing all or part of the aforementioned industrial computing network AI-generated content task allocation method, which is installed in an edge server in the industrial computing network. The industrial computing network AI-generated content task allocation device specifically includes the following components: (1) A federated learning module is used to execute a federated learning step with multiple iterations. The federated learning step includes: constructing a federated learning training alliance based on the comprehensive evaluation scores of each terminal device to be federated in the industrial computing power network, and distributing global model parameters for AI-generated content corresponding to the current iteration to each terminal device in the selected alliance; receiving the updated global model parameters returned by each terminal device in the training alliance after local training based on a hybrid expert model with a gating network, and aggregating the updated global model parameters based on the comprehensive evaluation scores of each terminal device to obtain the global model parameters corresponding to the next iteration; the comprehensive evaluation score is calculated based on the data distribution feature vector, real-time resource state vector, and historical contribution dynamically updated through contribution feedback signals of the terminal device.

[0096] (2) A resource contention module is used to cyclically execute a resource contention step within discrete time intervals. The resource contention step includes: receiving a request from the terminal device for an AI-generated content task; determining the target terminal device to be served, the resource allocation scheme, and the resource consumption quantification index based on all the requests, the currently available edge resource capacity locally, and the real-time resource contribution status of each terminal device that issued the request; sending the resource consumption quantification index to the target terminal device, and updating the resource contribution status of the target terminal device according to the resource consumption quantification index as the latest contribution feedback signal.

[0097] The embodiments of the AI-generated content task allocation device in the industrial computing network provided in this application can be used to execute the processing flow of the embodiment of the AI-generated content task allocation method in the industrial computing network described above. Its functions will not be repeated here, but can be referred to the detailed description of the embodiment of the AI-generated content task allocation method in the industrial computing network described above.

[0098] From a software perspective, this application also provides an industrial computing network AI-generated content task collaborative processing device for executing all or part of the aforementioned industrial computing network AI-generated content task collaborative processing method, which is installed in a terminal device in an industrial computing network. The industrial computing network AI-generated content task collaborative processing device specifically includes the following components: (1) A collaborative training module is used to perform a collaborative training step of multiple iterations. The collaborative training step includes: when it is selected into the federated learning training alliance based on the comprehensive evaluation score, it receives global model parameters for artificial intelligence-generated content distributed by the edge server in the industrial computing power network; it uses the global model parameters as global experts locally, and together with the local model as local experts and a gating network, it forms a hybrid expert model; it trains the hybrid expert model using local private data, and after the training is completed, it uploads the updated global model parameters to the edge server.

[0099] (2) Task offloading decision module, used to generate AI-generated content task, and determine whether to offload the AI-generated content task to the edge server based on the AI-generated content task and the local real-time resource status vector; if so, generate a request for the AI-generated content task and send the request to the edge server.

[0100] (3) Resource contribution synchronization module, which is used to update the local resource contribution status to synchronize with the edge server according to the resource consumption quantification index received from the edge server; wherein the updated resource contribution status serves as the latest contribution feedback signal and affects the subsequent calculation of the comprehensive evaluation score.

[0101] The embodiments of the AI-generated content task collaborative processing device in the industrial computing network provided in this application can be used to execute the processing flow of the embodiments of the AI-generated content task collaborative processing method in the industrial computing network described above. Its functions will not be repeated here, but can be referred to the detailed description of the embodiments of the AI-generated content task collaborative processing method in the industrial computing network described above.

[0102] The aforementioned terminal device may have a communication module (i.e., a communication unit) that can communicate with a remote edge server (which may be simply referred to as a server) to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0103] The server and the terminal device can communicate using any suitable network protocol, including network protocols not yet developed as of the date of this application. Such network protocols may include, for example, TCP / IP, UDP / IP, HTTP, HTTPS, etc. Of course, such network protocols may also include, for example, RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer) protocols used on top of the aforementioned protocols.

[0104] To further illustrate the above embodiments, this application also provides an AI-generated content task allocation system in an industrial computing network. This system addresses the collaborative optimization problem of simultaneously optimizing AIGC service quality, efficiency, and resource consumption in an environment of limited resources, heterogeneous data, and dynamic changes in the Industrial Internet of Things (IIoT). This problem has strict multi-objective and cross-level collaborative characteristics. Using a single solution may result in the loss of key indicators necessary for the overall AIGC service, such as generation quality, real-time response, or system sustainability. Therefore, this application proposes an integrated collaborative optimization scheme driven by a triple approach of model, incentives, and resources.

[0105] This application provides a federated fine-tuning method based on MoE and dynamic data similarity clustering, enabling industrial terminals with similar data characteristics to train collaboratively. This effectively mitigates the impact of Non-IID data and improves the performance of both global and local models while protecting data privacy. Simultaneously, it integrates comprehensive information such as network latency, node computational load, and model performance contribution into the MoE-gated network, achieving dynamic and efficient expert allocation. This solves the problem of coordinating model updates and resource status in dynamic edge environments, ensuring both model performance and communication and computational efficiency during the update process. A market-based incentive mechanism based on auction theory is introduced to fairly and reasonably compensate participating devices, stimulating their participation and forming a virtuous cycle of resource contribution and usage. Given the drastic changes in edge resources, incentivizing nodes to spontaneously collaborate and plan resources based on real-time market supply and demand. Computational resources, model services, and data contributions are quantified into tradable assets. Through pricing and competition, efficient and autonomous transactions are guided between resource-scarce terminal devices and resource-rich edge nodes, ensuring the vitality and sustainability of the service ecosystem. Establish a top-level optimization framework capable of overall coordination and dynamic decision-making. This framework unifies the modeling of terminal device selection, resource allocation, AIGC task offloading, and edge resource scheduling in federated learning, aiming to maximize system social welfare and achieve adaptive and efficient resource allocation. It realizes systematic multi-objective joint optimization by collaboratively innovating and organically integrating federated learning, market mechanisms, and dynamic resource management mechanisms, ultimately forming a self-adaptive, self-optimizing, and sustainably serving industrial AIGC collaborative ecosystem.

[0106] Specifically, this application example addresses the core issues of resource, data, and quality efficiency collaboration in industrial AIGC services. It constructs a four-dimensional collaborative autonomous optimization system encompassing perception, communication, computing, and intelligence. This system, through a hierarchical collaborative architecture and a two-way feedback mechanism, connects top-level AIGC application requirements with underlying physical resources. At the model level, a MoE architecture is adopted to fuse global models (deployed at the edge) and personalized local models (deployed at the terminal) through an adaptive gating network, balancing generalization and personalization capabilities. At the resource level, the federated learning process is treated as a special task and incorporated into a unified resource allocation framework along with the AIGC inference task. Through decomposition and coordination optimization, the complex federated problem is broken down into two sub-problems: federation selection and task-resource matching, which are efficiently solved using intelligent optimization algorithms and auction mechanisms, respectively. At the incentive level, a comprehensive evaluation score based on data utility, computing cost, and communication cost is constructed, and rewards are paid to participating devices through edge nodes, forming a sustainable collaborative ecosystem.

[0107] See Figure 5The system's physical architecture is comprised of three logical layers. The cloud layer, acting as the brain, is primarily responsible for storing and maintaining large-scale pre-trained AIGC (AI Generative Computing) foundational models. Due to its near-unlimited computing and storage resources, it focuses on pre-training and periodically updating complex models that are insensitive to latency. The edge layer consists of key nodes (such as edge servers and smart gateways) deployed within industrial parks or factories, forming a distributed AIGC service resource pool. Each edge node carries a lightweight, specific AIGC model, such as a diffusion model for quality inspection. Its core value lies in its proximity to the data source, providing nearby model inference services for latency-sensitive intelligent operations. The terminal layer comprises a massive number of industrial IoT devices, such as high-definition cameras, intelligent robotic arms, and sensors. These devices are not only producers of raw data and initiators of AIGC tasks, but with the development of embedded intelligence, they are also gradually equipped with small, heterogeneous intelligent models to complete local preliminary computation and real-time control tasks. This application example requires detailed definitions of edge nodes and device nodes (i.e., terminal devices) for subsequent design explanations of the mechanism: (1) Edge nodes (i.e., edge servers): Consider the set of edge nodes in the system as represented as Each edge node g stores the AIGC model representation as follows: , The parameters of the diffusion model are represented as follows: The total diffusion steps are expressed as... Edge node resource attributes are represented as follows: ,in This represents the node's maximum computing power. This represents the number of CPU cycles required for a node to process a unit of data.

[0108] (2) Device node (i.e., terminal device): The device set is represented as Each device stores a small, heterogeneous intelligent model. , Represented as intelligent model parameters, the number of parameters in the local model is... Overall diffusion steps Each entity stores its own private dataset (containing multiple samples) containing data samples. ,in Represents the feature vector of image data. Represents image size. This represents the number of samples for device n. The resource attributes of a device node (i.e., the terminal device) are represented as follows: ,in This represents the node's maximum computing power. This represents the number of CPU cycles required for a node to process a unit of data. This indicates the node's maximum communication bandwidth. This indicates the maximum transmit power of the device node. Each device node generates an AIGC task, and each task is characterized by its data volume. and maximum tolerable latency .

[0109] Based on the above system architecture, the operation flow of the proposed framework is summarized as follows: The workflow constructed in this application example is not a simple chain of independent stages, but a collaborative optimization system with the goal of maximizing system efficiency, mediated by the flow of data value, and possessing self-regulation and continuous optimization capabilities. Its four core components are intercoupled, forming an organically unified whole, specifically manifested as follows: Phase 1, Model Validation and Participation Phase: Edge nodes pre-deploy lightweight AIGC models and attract terminal devices with abundant data resources and good computing performance to participate in the subsequent federated learning ecosystem through a contribution-based incentive mechanism. The resource profiles and anonymized data characteristics submitted by participants during registration provide a basis for the construction of the subsequent federated learning cluster and establish a profile for their reputation assessment in the future resource market. The quality of node selection at this stage directly determines the data diversity foundation for subsequent model training and the overall potential of the system's resource pool.

[0110] Phase 2, MoE-based collaborative training phase: This stage is the core computing engine for improving the quality of AIGC services in the system. The edge server, acting as a coordinator, dynamically constructs a training consortium based on the similarity of node data distribution and real-time resource status, and implements a federated fine-tuning mechanism based on a hybrid expert model. This mechanism, through a trainable adaptive gating network, deeply integrates the global model and local personalized models, enabling terminal devices to train global experts, local experts, and the gating network in parallel locally. Only global model parameter updates are encrypted and uploaded for secure aggregation. The output of this stage (model performance) directly determines the willingness of terminal devices to pay in the subsequent task unloading stage, while the market results in the fourth stage (resource allocation) form valuable contribution feedback signals to guide the construction of subsequent training consortia and the optimization of expert networks, thereby achieving a positive guidance of market value on the direction of model evolution.

[0111] Phase 3, Task Generation and Unloading Decision Phase: This stage involves autonomous decision-making by the terminal device based on multi-dimensional state awareness. Based on the characteristics of the AIGC task generated in real time (such as data scale and latency constraints) and its local resource status (computing load and network conditions), the device rationally decides whether to execute the task locally or offload it to an edge server. This decision-making process is essentially a preliminary matching of system service capabilities with local needs. Simultaneously, its offloading decision injects real and dynamic demand signals into the resource allocation market for the next stage, becoming a key input driving the optimization of resource prices and allocation strategies.

[0112] Phase 4, Resource Allocation Phase Based on Auction Theory: This stage is the core of resource allocation and the hub for value realization in the entire system. This stage employs a designed VCG auction mechanism, treating the task-initiating device as the buyer and the edge resource pool as the seller, achieving efficient resource allocation through competitive bidding. This mechanism is deeply coupled with the preceding stages: it fulfills the promise of the first-stage incentive mechanism by providing economic compensation based on the actual contribution of nodes; it evaluates and utilizes the value of the model output from the second stage, making high-performance models highly attractive targets in the market; and it responds to and optimizes the offloading decision in the third stage, aggregating dispersed individual demands into a globally optimal allocation scheme aimed at maximizing the system's social welfare.

[0113] In summary, this solution constructs a highly collaborative closed-loop optimization system: the incentive mechanism guides high-quality nodes to participate (Phase 1) and jointly build a high-performance model (Phase 2); model performance stimulates service demand and drives offloading decisions (Phase 3); market demand achieves precise resource allocation through an auction mechanism (Phase 4); and the economic signals (i.e. contribution feedback signals) and performance feedback generated by the resource allocation results, in turn, optimize the incentive strategy and the construction of the training alliance (feedback to Phase 1 and Phase 2), ultimately forming a self-evolving and continuously optimizing industrial AIGC service ecosystem.

[0114] The core technologies are explained in detail below: (I) Federated Fine-tuning Mechanism Based on MoE The content generated by the edge model may not meet the specialized requirements of device AIGC tasks. Therefore, in order to coordinate training time, training accuracy, and training samples, we consider a privacy-preserving federated learning fine-tuning method, selecting devices to form a federated learning training consortium to fine-tune the global version of the edge model.

[0115] The global model possesses rich generalized knowledge, while the device-stored local model possesses personalized knowledge. The device-local model can be utilized for two reasons: 1) By transmitting certain feature gradients from the existing local model, local data features can be acquired while protecting privacy; 2) While enabling the global model to acquire personalized knowledge, the generalization ability of the global model can be leveraged to enhance the general performance of the local model, adapting to non-IID data across different terminal devices at the data level.

[0116] The MoE system, which constructs global and local models at the device terminal, employs a federated learning fine-tuning process as follows: Figure 6 As shown.

[0117] To achieve these capabilities, the federated learning fine-tuning process incorporates a small, shareable model, much smaller than a locally local heterogeneous model. For example... Figure 7 As shown, the steps include: Step 1: Selecting a client consortium; Step 2: Distributing the global expert model; Step 3: Building the MoE architecture; Step 4: Calculating multi-task loss; Step 5: Backpropagation to update parameters; Step 6: Local convergence judgment; Step 7: Uploading global expert parameters; Step 8: Server-weighted aggregation; Step 9: Global convergence judgment.

[0118] Specifically, in the first In round-robin communication, the workflow of the federated learning fine-tuning process includes the following steps: (1) Model Distribution: The edge server, acting as a coordinator, first dynamically selects a group of terminal devices to form the current training consortium based on the data distribution of the terminal devices (assessed using methods such as EarthMover's Distance to evaluate its similarity to the global distribution), real-time resource status (including computing power, communication bandwidth, and remaining power), and historical training contributions. Subsequently, the server distributes the globally lightweight expert model parameters updated in the previous round of aggregation to these selected terminal devices. The detailed process of model distribution is as follows: a) The server selects a set of federally fine-tuned terminal devices based on the perceived device resources and data distribution. The evaluation mainly considers three aspects: data benefits, computational model, and communication model, and finally weights them to form the overall comprehensive evaluation score.

[0119] b) The server samples the terminal devices and will do so in the [number]th [day]. Global isomorphic small feature extractor with in-round aggregation Send it to them.

[0120] c) The terminal device will receive the globally isomorphic small feature extractor As a global expert, it is used to extract generalized features across all classes and integrate local heterogeneous models. As a local expert, it is used to extract personalized features of locally visible classes.

[0121] (2) Local training: Introduce lightweight, personalized local control networks of homogeneity or heterogeneity. This approach balances generalization and personalization by dynamically assigning weights to the representations of each sample from two experts. Based on traditional MoE gated network parameters, we model network computation and communication resource information into the gate parameters; that is, we define the new weight vector as the softmax normalized result of the Hadamard product of the original weights, the bandwidth vector, and the computational capability vector. The gated network modeling formula is as follows: Let the parameters be for device nodes (i.e., terminal devices)... Input data sample The output weights of a traditional gated network are ,in Represents a gated network model. These are the parameters for the gated network. This application introduces a resource state vector in its application example. : in Indicates the remaining battery power of the device. Indicates the current available computing power of the device. This indicates the maximum tolerable latency for the task. This represents the currently available bandwidth of the device. Each element of the resource vector above is normalized to [0,1]. Therefore, the gating weight for resource status awareness is calculated as follows: in, This is the resource state weight matrix. This is the bias term. We construct a MoE architecture from these three models (gated network, globally isomorphic small feature extractor, and locally heterogeneous expert). The MoE and the three models in the head are trained simultaneously in an end-to-end manner. The local training process is as follows: a) Each local data sample Feed to global experts In addition, it was also fed back to local experts. In order to generate feature vectors; b) Each data sample It is also fed into the gating network. In this process, the weights of the two experts are determined. c) Then the representations of the two experts are mixed with the weights generated by the gating network; d) Update all models simultaneously using gradient descent; e) Iterate the above process until all locally heterogeneous complete models converge.

[0122] (3) Model Aggregation: After completing local fine-tuning, each terminal device only encrypts and uploads the updated global expert model parameters to the edge server, while the personalized local experts and gating networks remain locally, ensuring that the original data and core personalized information do not leave the domain. After receiving model updates from all participating terminal devices, the server uses a weighted aggregation strategy to fuse these parameters. The aggregation weights are not simply allocated based on the amount of data, but rather comprehensively consider the data quality of the terminal devices (measured by the similarity between their data distribution and the global distribution) and the stability of training (such as the gradient magnitude of this round of updates). Through this refined aggregation, a new generation of global expert models is generated, thus completing a full round of federated fine-tuning.

[0123] (II) Market-based incentive mechanisms based on auction theory To design a market-based incentive mechanism based on auction theory, we first need to model the comprehensive evaluation score of incentives, and then design a resource competition process based on auction theory.

[0124] (1) Incentive Model This incentive mechanism first establishes a comprehensive and quantitative evaluation score, concretizing the economic benefits of resource allocation into calculable objectives. The comprehensive evaluation score consists of two parts: benefits and costs. The benefits component reflects the quality and efficiency of AIGC task completion, specifically including service accuracy benefits (the value brought by quality indicators such as the matching degree and clarity of generated content with the expected target) and service efficiency benefits (the timeliness value brought by low-latency response to industrial processes). The cost component covers compensation for resource consumption, including the model training costs paid to the device nodes (i.e., terminal devices) participating in federated learning, and the energy consumption generated during task execution. The overall system benefit is ultimately expressed as the sum of the differences between the benefits and costs of all successfully matched tasks, with the fundamental goal of maximizing this long-term overall system benefit, expressed as: in, This represents the overall system benefit of the optimization objective (i.e., the comprehensive evaluation score). and These represent the reward and penalty terms of the system model, respectively. and These represent the revenue and rewards resulting from service accuracy and service efficiency, respectively. and This represents the cost and energy penalty paid for model training.

[0125] (2) Incentive mechanism process based on VCG auction theory Based on the comprehensive evaluation scores mentioned above, the system constructs an online auction market based on the VCG (Vickrey-Clarke-Groves) theory. This market operates cyclically at discrete time intervals to address dynamic fluctuations in terminal device requests and network resources. At the start of each auction round, devices with task offloading needs act as buyers, submitting bids to their reachable edge node resource pool. Each device is allowed to submit multiple alternative bid combinations, each specifying the desired edge servers, the requested proportion of computing resources, and the valuation they are willing to pay for this combination. This design empowers devices to flexibly bid based on the urgency of their tasks and resource preferences, constructing a flexible resource scheduling mechanism oriented towards the actual needs of offloading tasks.

[0126] To manage fluctuating demand from different active terminal devices and their requests, auction-driven matching systems operate within discrete time intervals of equal duration, denoted as... In each round In this process, service providers (i.e., edge servers) make matching decisions after receiving bids from end devices. They allow each end device to... Round submission A set of optional bids, in which Each bid includes the desired node and the valuation submitted for that desired node. We provide the following for the first... Each bid defines a binary tuple. ,in Indicates the first The bid was for terminal equipment. Specified server The proportion of computing power That's the valuation in their report. In each round... At the end, the service provider (i.e., the edge server) calculates the pricing decision and matches the results. Variables Indicates terminal device Whether the requested resource was successfully obtained. This means that when a match is successful, terminal device n will pay the service provider (i.e., the edge server). To obtain the corresponding services. Assuming all bids are valid, the problem can be represented by the following binary integer programming model: In Ultimately, the incentive mechanism and the federated learning terminal devices determine the outcome. Among these constraints, the first ensures that each controller can obtain at most one edge node match in each bidding round. The second outlines the budget constraints for each device. Indicates device The third constraint considers limited resources, stipulating that the total resources allocated to the winning group cannot exceed the budget. ,in Indicates the first Controller in each bidding group The CPU resource requirements. The final constraint determines the set of winners.

[0127] The optimal social welfare problem requires complete system knowledge across all rounds, which is impractical due to the stochasticity of resources and unpredictable end-device demands and bids. Furthermore, the interdependence of cross-round decisions driven by the controller's budget constraints complicates the matching process. To address this issue, we must transform the complex offline matching problem into a series of online single-round matching problems. The specific method is as follows: a) Initialization Before the first auction round begins, the system initializes a key auxiliary variable for each terminal device participating in the federated learning ecosystem. This variable is defined as a continuous mapping function of the current budget consumption status of the equipment, with its initial value set to 0, representing the complete state of the equipment budget. Simultaneously, the system calculates a global adjustment factor, which is determined by the maximum ratio of the estimated value of any bid to its total budget across all equipment sets.

[0128] b) Modeling the single-matching problem At the start of each auction round, instead of directly solving a complex long-term optimization problem, the system decomposes it into independent single-round matching. The system first uses auxiliary variables representing the budget state of each device at the end of the previous round. The system dynamically represents the original valuation of the equipment in this round of bidding. For equipment whose budget is about to run out, its effective valuation will be significantly reduced, thereby reducing its probability of winning the bid and automatically meeting long-term budget constraints. Based on the effective valuation, the system constructs a single-round matching optimization model with the objective of maximizing current social welfare. The decision variable of this model is a binary winning bid identifier, and the constraints focus on the resource capacity limit of the current round and the rule that each equipment can win at most one bid. The problem of a single matching is represented as follows: c) Modeling the dual problem Since the single-round matching problem is still an NP-hard integer programming problem, to obtain a high-quality solution in polynomial time, we employ linear programming relaxation techniques to relax the binary variable constraints into continuous variable constraints. Furthermore, using Lagrange duality theory, we transform the primal problem into a dual problem. In the dual form, we introduce a dual variable for each device, reflecting its marginal cost in resource competition in the form of a shadow price. Simultaneously, we introduce a global dual variable for the entire resource pool, characterizing the scarcity of system resources. Solving this dual problem yields a lower bound for the primal problem and provides guidance for constructing subsequent integer solutions, expressed as: in and It is a dual variable.

[0129] d) Construction of auxiliary variables To subtly integrate the long-term budget constraints of the equipment into each round of online decision-making, we designed a set of update rules for auxiliary variables based on an initial global adjustment factor. This auxiliary variable is a non-linear function of the equipment budget consumption rate, its value monotonically increasing with the cumulative expenditure on the equipment, approaching its theoretical maximum when the budget is completely exhausted. This mechanism ensures that the satisfaction of the budget constraint does not rely on the prediction of future information, achieving decoupling of cross-period decision-making. We defined two additional variables: in This represents the maximum ratio between the valuation of any set and the budget for the equipment. This adjustment promotes balanced and efficient use of the budget and improves the stability of one-off matching.

[0130] e) Budget Update After each round of auction matching decisions, the system needs to update the budget status of all bidding equipment in real time. For equipment that did not win the bid, its remaining budget and auxiliary variables remain unchanged. For winning equipment, the system calculates its actual payable price according to the VCG payment rules and deducts it from its remaining budget. Subsequently, based on the latest payment records, the system strictly follows a preset non-linear update formula to recalculate the updated auxiliary variable values ​​for each piece of equipment. This update process is the core element ensuring the long-term and fair operation of the incentive mechanism.

[0131] In the online matching framework, for any given time... The update to its budget is expressed as follows: This represents the bid. After the budget is updated, the single VCG auction algorithm (given later) is executed to determine the set of winning bidders. For other terminal devices outside the winning set, their bids for the items that have already been bid on are set to 0.

[0132] Finally, update the budget-based auxiliary variables: f) Implementation of the VCG auction mechanism After determining the set of winning bidders, the system executes strict VCG payment rules to calculate the actual cost for each winner. Specifically, the amount a winning bidder must pay equals the social welfare loss incurred by excluding other competitors through their victory. This payment price is completely independent of their own bid valuation and depends solely on the valuation information of all other bidders. This pricing mechanism is proven true from a game theory perspective, demonstrating that for every rational device, truthfully reporting its actual valuation of the resource is the optimal strategy under any circumstances, thus fundamentally preventing strategic bidding behavior and ensuring market efficiency and fairness. (Terminal equipment) The payment is represented as: .

[0133] g) Solving for integer solutions based on LP relaxation and dual decomposition After obtaining the optimal solution to the dual problem, we face the challenge of transforming it into a feasible solution to the original integer programming problem. We first obtain the optimal fractional solution of the relaxed linear programming problem, and then, using a random rounding or deterministic decomposition algorithm based on dual decomposition, represent the fractional solution as a convex combination of a series of feasible integer solutions. By solving a matching covering linear programming problem, we can determine the coefficients of these integer solutions and ultimately select the integer solution that maximizes a certain end-device satisfaction index (such as the approximation ratio between overall utility and the optimal fractional solution) as the final matching scheme for this round of auction. This series of operations ensures that the final solution satisfies all integer constraints and approximates the global optimum with a high probability.

[0134] Based on the above approach, binary constraints are relaxed into continuous variable constraints. . A feasible solution is represented as Meanwhile, the utility of each terminal device is typically expressed as the difference between its payment and its estimated value, and is obtained through... Please provide a solution.

[0135] By decomposing the optimal fractional solution into a combination of feasible integer solutions based on LP duality, we can find... and a set of integer solutions Make Where L represents the index set, and The optimal fractional solution is reduced by a certain factor, and then... The scaling factor can be used to determine a feasible solution. Dual variables are introduced based on constraints 1 and 2. and The dual problem is obtained: By applying the ellipsoidal method to solve this pair of primordial-dual problems, the optimal integer solution that approximates the fractional solution can be obtained.

[0136] h) Execute in a loop The complete process from step ② to step ⑦ constitutes an independent online auction cycle. The system executes this process synchronously and automatically at each discrete time interval. As the cycle progresses, device demand, network conditions, and resource availability dynamically change. This incentive mechanism, through continuous cyclical execution, continuously collects new bids, solves new problems, and updates the system status, thereby achieving continuous, adaptive, and highly efficient optimization management of the entire federated learning resource ecosystem, forming a robust and dynamic market-oriented operational closed loop.

[0137] (III) Joint optimization mechanism for alliance selection and resource allocation (1) Core issues In dynamic industrial computing networks, the quality of AIGC services (such as image generation clarity and feature integrity) is affected by a combination of factors: dynamic device addition or removal leading to time-varying data distribution, resource contention causing computational and communication bottlenecks, and privacy constraints limiting data sharing. These factors work together to cause drastic fluctuations in AIGC service quality, making it impossible for traditional single technologies (such as independently optimized federated learning or resource allocation) to stably guarantee service quality while protecting privacy. For example, with non-independent identically distributed (Non-IID) data, federated learning model updates may be interrupted due to resource contention, thus affecting the accuracy of generated content; while static resource allocation strategies cannot adapt to dynamic changes in network conditions, leading to an increased task dropout rate.

[0138] This application example is the first to jointly model the "model training-resource allocation-economic incentives," defining a new problem that can only be solved through deep integration: achieving smoothing and long-term stable optimization of the AIGC service quality curve while protecting privacy. This problem requires simultaneously optimizing terminal device selection, resource allocation, and task offloading decisions in federated learning, with the goal of maximizing system social welfare, ensuring a synergistic improvement in model performance, resource efficiency, and economic sustainability.

[0139] Specifically, the optimization objective consists of model performance gains and resource allocation costs. However, unlike traditional methods, this application example introduces economic incentive signals and resource status awareness to enable mutual feedback between model training and task inference. in, These represent the node selection of node n for each edge node g during the federated learning process. Allocate computing power Transmission bandwidth Transmission power , The notation represents the matched node and the proportion of computing resources obtained.

[0140] Constraint 1 indicates whether a node participates in federated learning on edge server g, and constrains the total number of nodes participating in federated learning to remain at S. Constraint 2 sets an upper limit constraint on the computing power of device nodes (i.e., terminal devices), and ensures that the overall computing power remains within the maximum value range. Constraint 3 constrains the communication bandwidth between device nodes (i.e., terminal devices) and edge nodes, and ensures that the overall bandwidth utilization of device nodes (i.e., terminal devices) does not exceed the maximum uplink and downlink bandwidth. Constraint 4 ensures the upper limit of the transmit power of device nodes (i.e., terminal devices) and the overall maximum value. Constraint 5 indicates that the offloading strategy is carried out on local and edge nodes. Constraint 6 ensures the proportion of computing resources provided.

[0141] Problem (P1) contains two sets of variables to be solved, namely and resource allocation The variables are not only of different types but also strongly coupled, making P1 an NP-hard MINLP problem, difficult to solve in polynomial time. To address this issue, this paper divides it into two sub-problems: model training node selection and task offloading decision.

[0142] The node selection problem needs to consider the benefits and costs that the edge node model offloading devices can bring to the task. When the offloading decision is a given fixed value, the selection of alliance nodes is related to the system configuration of the device nodes (i.e., terminal devices), such as data resources, computing resources, and network resources. Better data resources will obtain more suitable data training samples, but this may lead to an increase in computing costs. Therefore, balancing these factors is crucial for achieving effective and reliable control. When the offloading decision is fixed, i.e. The goal of the node selection problem is to determine the alliance node selection, calculate resource allocation, communication resource allocation, and power allocation, so that the model can meet the task requirements to the greatest extent possible after the task offloading decision is given, while saving processing costs. Since it only relates to constraints C1-C4, problem P1 can be transformed into an auction-driven resource matching problem P3. (2) Solution process The joint optimization mechanism proposed in this application example decomposes a complex problem into two coupled sub-problems through a hierarchical iterative framework and introduces a non-obvious interaction mechanism to form an organic whole among the components: a) Interaction Mechanism 1: Economic signals (i.e., contribution feedback signals) guide model training in reverse. In each round of optimization, economic reward signals (such as actual node revenue and resource scarcity indicators) are generated based on the resource allocation results of the VCG auction. These signals are fed back to the federated learning consortium selection phase. Specifically, devices with high historical revenue receive higher weights in subsequent consortium formation because their data contributions and resource stability are more valuable. At the same time, the resource scarcity indicator is used to adjust the training objective of the MoE gating network, so that the model prioritizes the generation quality of high-value tasks when resources are scarce.

[0143] b) Interaction Mechanism Two: Resource State Aware MoE Gated Network This application example innovatively extends the MoE gating network, enabling it to not only rely on input data features when deciding whether to use a "global expert" or a "local expert," but also to perceive the dynamic resource status of the device in real time. When updating the gating network parameters, these resource indicators are incorporated, combined with the original weights through a Hadamard product, and then normalized using softmax. This allows the model's inference strategy to adapt to changes in the resource environment; for example, prioritizing low-complexity local experts to save energy when device power is low, while selecting high-performance global experts to ensure quality when task latency is critical.

[0144] c) Detailed steps of the solution process: c1) Initialization: Set the iteration index, convergence threshold and maximum number of iterations, and randomly initialize the resource contention strategy.

[0145] C2) Hierarchical iterative optimization: Step A: Fixed task unloading strategy, optimizing federated learning consortium and resource allocation. An improved Adaptive Differential Evolutionary Algorithm (JADE) is used to solve the node selection and resource allocation subproblems. The JADE algorithm encodes discrete selection and continuous allocation as composite individuals, driving population evolution through mutation, crossover, and selection operations. Key innovation: The fitness function not only includes model accuracy gains but also incorporates economic signals from the previous auction round (i.e., contribution feedback signals) (such as node reward weights), aligning coalition selection with market value.

[0146] Step B: Based on the updated alliance structure, optimize task unloading and resource matching. The AIGC task offloading is modeled as an online VCG auction problem, considering constraints such as equipment budget and resource capacity. Key innovation: In the auction mechanism, the bidding valuation of a node is dynamically affected by its resource status (such as remaining power and computing load), and is adjusted through the model performance expectations fed back by the MoE gating network to ensure that resource allocation matches model capabilities.

[0147] C3) Convergence judgment: Calculate the joint control cost. If the change in node selection strategy is less than the threshold or the maximum number of iterations is reached, stop; otherwise, return to step A.

[0148] The calculation process of hierarchical iterative optimization is shown in Table 1.

[0149] Table 1 The core of this application example lies in providing an industrial computing network AIGC resource allocation scheme that deeply integrates model training and task inference. Its key and intended protection point is the construction of a three-layer collaborative system consisting of cloud, edge nodes, and terminal devices, and the integration of the following technical solutions: (1) Federated fine-tuning method based on resource state awareness MoE: The core is to fuse the global model deployed at the edge and the local model on the terminal through a trainable adaptive gating network. The decision of this gating network not only depends on the features of the input data, but also integrates the dynamic resource state information of the device in real time (such as remaining power, available computing power, and task latency constraints). This method enables the model to adapt to the resource environment during inference and achieve a dynamic balance between service quality and resource consumption.

[0150] (2) Market-based incentive and resource allocation mechanism linked to model value: The core is the design of an online resource market based on VCG auction theory, in which terminal devices compete for marginal resources as buyers. The innovation of this mechanism is that: a) its bid valuation or final social welfare calculation deeply integrates the model performance benefits generated by MoE federated learning; b) the economic returns (payment price, node revenue) generated by the auction are used as key contribution feedback signals for the calculation of node selection weights in the subsequent formation of the federated learning alliance, thereby forming a closed loop of economic behavior and technological contribution.

[0151] (3) A hierarchical iterative joint optimization method for maximizing system social welfare: The core of this method is to decompose the complex mixed-integer nonlinear programming problem into two sub-problems: federated learning node selection and task unloading resource matching. An improved JADE algorithm and the aforementioned online auction mechanism are then used for hierarchical iterative solutions. The key to this method is that the two sub-problems are not solved independently, but rather through the iterative process, they pass decision context (such as a fixed unloading strategy or alliance structure) to each other, ultimately converging to a globally satisfactory solution through collaborative optimization under the guidance of the unified goal of system social welfare.

[0152] To further illustrate the beneficial effects of the above method, this application also provides specific experiments, which mainly include the following: 1. Experimental parameter settings The experimental environment was configured according to the IIoT standard to evaluate the effectiveness of the proposed adaptive collaborative design framework. The simulation setup included parameters and baseline selection in four parts: 1) Device Node Parameters: Each device node is equipped with computing power distributed in the range of [1, 5] GHz. The CPU cycles required for each data unit are in the range of [10³, 10⁴], the available bandwidth is in the range of [10, 100] MHz, and the transmit power is in the range of [0.1, 10] W. The input data size ranges from 1 to 100 MB, the local sample size is in the range of [100, 1000], and the diffusion step size is in the range of [10, 100].

[0153] 2) Edge node parameters: The computing power of each edge node is uniformly extracted from [5, 20] GHz, the model complexity is in the range of [10⁷, 10⁸], and the CPU cycles required for each data unit are in the range of [5, 20]. Within the range, the number of diffusion steps varies in [50, 200].

[0154] 3) Communication parameters: The wireless link is affected by Rayleigh fading, and the channel gain is uniformly distributed in the range of [0.5, 1.5]. The background noise power is fixed at 10⁻⁹ W, and the effective capacitance coefficient is set to 10⁻²⁷. The delay constraint for each task is within the range of [0.1, 1.0] seconds.

[0155] 4) To evaluate the robustness and performance of the proposed scheme, we compared it with several baseline algorithms: a) SA: Simulated Annealing explores the solution space by probabilistically accepting worse solutions, helping to avoid local optima during joint optimization.

[0156] b) DE: Differential Evolution (DE) evolves the population of solutions through mutation and crossover to optimize model selection and resource allocation.

[0157] c) WOA: Whale Optimization Algorithm mimics the feeding behavior of humpback whales, iteratively updating the solution through spiral and circling strategies.

[0158] d) RAN: Random Sample Concensus (RANS) is a simple baseline for allocating resources without optimization.

[0159] e) SPS: Second-Price Sealed-Bid Matching allocates resources based on bids, with the winner paying the second-highest bid.

[0160] 2. Implementation Effect Analysis To verify the effectiveness of the proposed JADE-based node selection algorithm, comparative experiments were conducted against three established optimization methods: SA, DE, and WOA. In all performance metrics, the proposed JADE algorithm consistently demonstrated superior results. Referring to Figure 8(a), compared to SA, JADE effectively reduced the average task latency by approximately 37.2%, and the performance gap further widened with the increase in the number of devices. This improvement stems from JADE's adaptive mutation strategy, which allows for a dynamic balance between exploration and exploitation throughout the optimization process. Also observed in Figure 8(b) is that JADE consumes 20.5% less energy than DE, particularly in high-density device scenarios.

[0161] Figures 9(a) and 9(b) further investigate the system performance as the number of edge nodes increases from 1 to 20. JADE demonstrates excellent performance in reducing system latency, achieving a latency reduction of approximately 55.0% compared to SA. Simultaneously, JADE consistently improves energy efficiency, consuming approximately 20%-35% less energy than DE.

[0162] Figures 10(a) and 10(b) present the performance of the proposed EOMA algorithm from a user-centric perspective: In Figure 10(a), due to the enhanced task matching algorithm, excessive rounds and the number of users lead to fragmented resource allocation, resulting in decreased satisfaction. Figure 10(b) shows that budget utilization increases with the number of rounds, indicating more efficient currency deployment in a competitive environment. Here, T represents the number of rounds.

[0163] To further validate the effectiveness of EOMA, Figures 11(a) and 11(b) compare it with RAN and SPS under different configurations. Figure 11(a) shows that the overall social welfare improves as users enhance their capabilities, illustrating the scalability of the algorithm. In Figure 11(b), the increase in the number of bid sets leads to bid diversity. EOMA achieves welfare improvements of 33.2% and 46.7% higher than the baseline at K=25. These results confirm the robustness and adaptability of EOMA across different system sizes.

[0164] Finally, Figure 12(a), Figure 12(b) and Figure 13 Evaluations of total utility, total cost, and coupling objectives are presented. Results show that control cost gradually decreases while utility gradually increases with the number of iterations. After approximately 20 iterations, both metrics tend to converge. This result reflects the algorithm's progress from local optima to global optima, validating its effectiveness. Furthermore, it is shown that increasing the number of nodes leads to higher system complexity, intensifies resource competition, and results in higher total cost.

[0165] In summary, simulation results validate the effectiveness of the proposed hierarchical collaborative design framework in enhancing AIGC services for the Industrial Internet. The JADE-based optimization algorithm consistently outperforms SA, DE, and WOA in key metrics (task latency, energy consumption), demonstrating strong scalability with increasing device and edge node density. Furthermore, compared to RAN and SPS algorithms, the EOMA algorithm achieves higher controller satisfaction and more efficient budget utilization. These results confirm the framework's superiority in adaptive resource coordination and realistic, utility-driven task matching under dynamic, resource-constrained environments.

[0166] In other words, the core advantage of the entire joint optimization mechanism lies in its successful transformation of an NP-hard complex resource allocation problem into two sub-problems that can be solved by efficient algorithms and can co-evolve. This provides a practical and high-performance resource collaborative management solution for large-scale industrial computing network systems, achieving global optimization of service quality and resource efficiency while protecting data privacy. In highly non-independent and identically distributed (Non-IID) industrial data environments, the application example in this application effectively improves the performance of the AIGC model through a JADE-based optimization algorithm and a hybrid expert (MoE) collaborative training architecture. Compared to traditional federated learning methods, the accuracy of the model-generated content (such as defect detection images) is improved by approximately 10%. Simultaneously, the generated content shows significant improvements in key feature completeness, image clarity, and style matching with specific business needs, stably meeting the stringent requirements of industrial scenarios for AIGC service quality. Addressing the problem of low task completion rates in dynamic resource competition environments, this invention introduces a resource-state-aware gating network and an economic signal-guided alliance formation mechanism to achieve efficient resource adaptation and scheduling. Compared to traditional methods where task dropout rates generally exceed 20%, this invention successfully controls the proportion of tasks dropped due to resource shortages to below 5%. Even in large-scale scenarios (e.g., 50 devices, 10 edge nodes), the system maintains a low dropout rate, demonstrating excellent scalability and robustness. The EOMA (Enhanced Online VCG Matching Algorithm) online matching mechanism proposed in this application consistently maintains user satisfaction above 88% throughout 300 rounds of auctions, reaching a maximum of 90%, significantly outperforming baseline methods such as random matching and second-price sealed auctions. At the system level, the proposed hierarchical iterative joint optimization algorithm converges after approximately 20 iterations, improving the overall system utility by 25% to 50%, while reducing system energy consumption by 20.5% and average task latency by 37.2%, achieving synergistic optimization of multi-objective performance. This application successfully breaks the inherent trade-off between privacy protection and model performance in traditional solutions. By ensuring that all raw data remains within the domain, and through deep coupling of federated learning and market-based incentive mechanisms, the framework has achieved, for the first time, simultaneous optimization of model accuracy, response efficiency, and resource consumption. This framework not only stimulates the continuous participation of nodes and fosters a healthy data contribution ecosystem, but also ultimately promotes the long-term and stable improvement of AIGC service quality.

[0167] In other words, the application examples of this application have the following beneficial effects: (1) Significantly improve the performance of the model on Non-IID data: Through the dynamic alliance formation mechanism, devices with similar data distribution are clustered and trained collaboratively. Combined with the adaptive gating network of the MoE architecture, the negative impact of data heterogeneity is effectively mitigated.

[0168] (2) Combining privacy protection and high performance: Under the premise of strictly protecting the original data from leaving the local area, the application example of this application has achieved model performance close to that of centralized training through efficient parameter exchange and model fusion strategies, and successfully solved the problem of balancing privacy and performance.

[0169] (3) Incentivize sustainable operation of global nodes: By combining a budget-aware auction mechanism, rewards are fairly linked to data contribution and quality, which not only guarantees the income of edge nodes, but also incentivizes resource-constrained devices to continue to participate, thus forming a healthy data ecosystem and resource pool.

[0170] (4) It has high adaptability and scalability: The proposed hierarchical optimization framework does not depend on a specific model structure or resource type. It can adapt to the dynamic changes in network status, resource availability and task requirements, and is easy to extend to large-scale IIoT scenarios and various AIGC applications.

[0171] 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 AI-generated content task allocation method or the AI-generated content task collaborative processing allocation method mentioned in the above embodiments. The processor and 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 memory via wired or wireless means.

[0172] 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.

[0173] 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 AI-generated content task allocation method or the AI-generated content task collaborative processing allocation method in the industrial computing power network 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 AI-generated content task allocation method or the AI-generated content task collaborative processing allocation method in the industrial computing power network in the above method embodiments.

[0174] 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.

[0175] The one or more modules are stored in the memory, and when executed by the processor, they execute the AI-generated content task allocation method or the AI-generated content task collaborative processing allocation method in the industrial computing power network of the embodiment.

[0176] 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.

[0177] 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.

[0178] 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.

[0179] This application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of the aforementioned method for allocating AI-generated content tasks in an industrial computing network or the method for collaborative processing of AI-generated content tasks in an industrial computing network. 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.

[0180] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned method for allocating AI-generated content tasks in an industrial computing network or the method for collaborative processing of AI-generated content tasks in an industrial computing network.

[0181] 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.

[0182] 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.

[0183] 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.

[0184] 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 allocating AI-generated content tasks in an industrial computing network, characterized in that, Performed by edge servers in an industrial computing network, the method includes: The federated learning process involves multiple iterations, including: constructing a federated learning training consortium based on the comprehensive evaluation scores of each terminal device participating in the federated training within the industrial computing power network; distributing global model parameters for AI-generated content corresponding to the current iteration to each terminal device in the selected consortium; receiving updated global model parameters returned by each terminal device in the training consortium after local training based on a hybrid expert model with a gating network; and aggregating the updated global model parameters based on the comprehensive evaluation scores of each terminal device to obtain the global model parameters corresponding to the next iteration; the comprehensive evaluation score is calculated based on the data distribution feature vector, real-time resource state vector, and historical contribution dynamically updated through contribution feedback signals of the terminal device. A resource contention step is executed cyclically at discrete time intervals. This resource contention step includes: receiving a request from the terminal device for an AI-generated content task; determining the target terminal device to be served, the resource allocation scheme, and the resource consumption quantification index based on all the requests, the currently available edge resource capacity locally, and the real-time resource contribution status of each terminal device that issued the request; sending the resource consumption quantification index to the target terminal device, and updating the resource contribution status of the target terminal device according to the resource consumption quantification index as the latest contribution feedback signal.

2. The method for allocating AI-generated content tasks in an industrial computing network according to claim 1, characterized in that, The construction of the federated learning training alliance based on the comprehensive evaluation scores of each terminal device participating in federated training in the industrial computing power network includes: Based on the latest contribution feedback signal of each terminal device currently participating in federated training in the industrial computing power network, the historical contribution of each terminal device is dynamically updated. Based on preset weighting coefficients, the real-time resource status vector, data distribution feature vector, and historical contribution of each terminal device are weighted and summed to obtain the comprehensive evaluation score of each terminal device. Based on the comprehensive evaluation scores from high to low, a preset number of terminal devices are selected to form a federated learning training alliance, according to the constraints. The constraints include: the total number of selected terminal devices is less than or equal to the preset upper limit of the alliance size, and the dimensions representing power consumption and computing power in the real-time resource status vectors of the selected terminal devices are both higher than the preset participation threshold.

3. The method for allocating AI-generated content tasks in an industrial computing network according to claim 1, characterized in that, Prior to the federated learning step that performs multiple iterations, the following is also included: The incentive strategy information for participating in federated learning is broadcast to each terminal device in the industrial computing network. The incentive strategy information is used to indicate that, in the resource contention step, the data distribution feature vector and the real-time resource status vector reported by the terminal device are the basis for determining whether the terminal device can become a target terminal device to be served locally. The system receives a registration application from a terminal device, wherein the registration application includes the data distribution feature vector and the real-time resource status vector of the terminal device; the data distribution feature vector is calculated by the terminal device based on its local private data and reported in an anonymized form; the real-time resource status vector includes at least one of the following: the remaining battery percentage of the terminal device, the ratio of current available computing power to nominal capacity, and the ratio of current available communication bandwidth to maximum bandwidth.

4. The method for allocating AI-generated content tasks in an industrial computing network according to claim 1, characterized in that, The gating network is configured to dynamically adjust the output weights of the global model (as a global expert) and the local model (as a local expert) in the hybrid expert model based on the real-time resource state vector of the terminal device. Furthermore, the training objectives of the gating network include: minimizing the loss function of the AI-generated content task, while optimizing the energy efficiency and expected latency of the terminal device executing the AI-generated content task.

5. The method for allocating AI-generated content tasks in an industrial computing network according to claim 1, characterized in that, The step of determining the target terminal device to be served, the resource allocation scheme, and the resource consumption quantification index based on all the requests, the currently available edge resource capacity locally, and the real-time resource contribution status of each terminal device that issued the request, and updating the resource contribution status of the target terminal device according to the resource consumption quantification index as the latest contribution feedback signal, includes: Determine each of the requests received within the current time interval, wherein each request contains resource requirement information and corresponding original estimates for the AI-generated content task, and the resource requirement information includes: the type of computing resources requested, the quantity of computing resources, and the maximum expected task processing latency. Obtain the currently available edge resource capacity locally, and query the current resource contribution status of the terminal devices that issued each request; wherein, the resource contribution status includes budget status variables and budget consumption status variables; A resource allocation optimization problem is constructed with the goal of maximizing the social welfare of the system, wherein the social welfare of the system is the sum of the social welfare contribution values ​​of each of the terminal devices that issued the request within the current time interval; the social welfare contribution value of each terminal device is obtained by adjusting the original estimate in its request according to the current resource contribution status of the terminal device. Solve the resource allocation optimization problem, and under the constraint of the available edge resource capacity, determine the resource allocation scheme to maximize the social welfare of the system, and determine the terminal device that obtains the resources as the target terminal device according to the resource allocation scheme. Based on the VCG auction mechanism, the resource consumption quantitative indicators to be undertaken by each of the target terminal devices are calculated, and the resource consumption quantitative indicators include: the payment price determined according to the VCG auction mechanism; The payment price is deducted from the budget status variable of the target terminal device, and the budget consumption status variable of the target terminal device is updated according to a preset nonlinear update rule as the latest contribution feedback signal.

6. A collaborative processing method for AI-generated content tasks in an industrial computing network, characterized in that, Performed by terminal devices in an industrial computing network, the method includes: The collaborative training process involves multiple iterations, including: when a device is selected into the federated learning training consortium based on a comprehensive evaluation score, it receives global model parameters for AI-generated content distributed by an edge server in the industrial computing network; locally, it uses these global model parameters as a global expert, together with an existing local model as a local expert and a gating network, to form a hybrid expert model; it trains the hybrid expert model using local private data; and after training, it uploads the updated global model parameters to the edge server. An AI-generated content task is generated, and based on the AI-generated content task and the local real-time resource status vector, it is determined whether the AI-generated content task should be offloaded to the edge server; if so, a request for the AI-generated content task is generated and the request is sent to the edge server. If a resource consumption quantification indicator is received from the edge server, the local resource contribution status is updated according to the resource consumption quantification indicator to synchronize with the edge server; wherein, the updated resource contribution status serves as the latest contribution feedback signal and affects the subsequent calculation of the comprehensive evaluation score.

7. The collaborative processing method for AI-generated content tasks in an industrial computing network according to claim 6, characterized in that, Prior to the collaborative training step that performs multiple iterations, the following is also included: If, based on the incentive strategy information for participating in federated learning broadcast by the edge server, it is determined to submit a registration application to the edge server, then the bulldozer distance between the local data distribution and the global reference distribution is calculated, and based on the bulldozer distance, a desensitized data distribution feature vector is generated. The terminal device obtains its own real-time resource status vector, wherein the real-time resource status vector includes at least one of the following: the remaining power percentage of the terminal device, the ratio of the current available computing power to the nominal power, and the ratio of the current available communication bandwidth to the maximum bandwidth. A registration application is generated based on the data distribution feature vector and the real-time resource status vector, and the registration application is reported to the edge server.

8. The collaborative processing method for AI-generated content tasks in an industrial computing network according to claim 6, characterized in that, The gating network is configured to dynamically adjust the output weights of the global model (as a global expert) and the local model (as a local expert) in the hybrid expert model based on the real-time resource state vector of the terminal device. Furthermore, the training objectives of the gating network include: minimizing the loss function of the AI-generated content task, while optimizing the energy efficiency and expected latency of the terminal device executing the AI-generated content task.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the task allocation method for AI-generated content in an industrial computing network as described in any one of claims 1 to 5, or the collaborative processing method for AI-generated content in an industrial computing network as described in any one of claims 6 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 task allocation method for AI-generated content in an industrial computing network as described in any one of claims 1 to 5, or the collaborative processing method for AI-generated content in an industrial computing network as described in any one of claims 6 to 8.

Images