AIGC flow scheduling method based on diffusion model and deep reinforcement learning

Abstract

The invention provides an AIGC flow scheduling method based on a diffusion model and deep reinforcement learning. The method comprises the following steps: an AIoT terminal generates an AIGC service request according to a service demand, an edge gateway analyzes the received AIGC service traffic, extracts traffic feature information, fuses the traffic feature information with network link state information and obtains system state information, and an SDN controller sends the AIGC service request to the edge gateway based on the received system state information. A deep reinforcement learning scheduling mechanism based on a diffusion model is called, a flow scheduling strategy meeting current network state information and AIGC service requirements is generated through a multi-step denoising strategy generation process and is issued to an edge gateway, and the edge gateway executes differentiated forwarding and queue scheduling operations on different AIGC service flows according to the flow scheduling strategy, so that the AIGC service flows are forwarded and queued. And flow priority control and dynamic bandwidth allocation are realized. According to the method, the scheduling of the AIGC flow in the edge network environment is optimized, low time delay and high flow completion rate are ensured, and meanwhile, the challenges of resource competition and bandwidth limitation are coped with.

Description

Technical Field

[0001] This invention relates to the field of computer application technology, and in particular to an AIGC (Artificial Intelligence Generated Content) method based on diffusion models and deep reinforcement learning. Background Technology

[0002] With the rapid development of artificial intelligence technology, AIGC has gradually become one of the important technological innovations in various industries. AIGC services cover the automatic generation of various types of content, from text and images to videos, and are widely used in entertainment, advertising, industrial manufacturing, and healthcare. Modern AIGC applications (such as Generative Adversarial Networks (GANs) and diffusion models) typically require a large amount of computing resources and bandwidth, thus necessitating efficient computing resource scheduling and data transmission mechanisms.

[0003] However, in real-world applications, especially in edge network environments, AIGC traffic faces several major challenges. First, there's bandwidth limitation. Edge network devices typically have relatively limited bandwidth, while AIGC services generate massive amounts of traffic during execution, particularly for image generation, video streaming, and other computationally intensive tasks. Second, there's dynamic traffic demand. Due to the computational complexity and data transmission requirements of AIGC tasks, traffic demands in edge network environments vary depending on the task, and traffic priorities and transmission needs also change over time. Third, there's latency sensitivity. Many AIGC services, especially real-time generative content (such as online image generation and speech synthesis), are extremely sensitive to latency. Excessive latency directly impacts service quality and can even lead to task failure. Finally, there are heterogeneous network and resource constraints. Edge network environments typically contain various devices and networks with different computing capabilities, bandwidth resources, and communication protocols. This heterogeneity complicates traffic scheduling, especially given the varying priorities, data sizes, and computational resource requirements of different tasks.

[0004] One existing AIGC request scheduling method based on DRL (Deep Reinforcement Learning) involves adjusting traffic priorities through online learning and optimizing network resource allocation in real time to meet the needs of latency-sensitive AIGC services. While AC (Actor-Critic) based algorithms improve the average quality of generated content and meet deadline requirements by leveraging deep reinforcement learning to capture global scheduling features.

[0005] The drawbacks of the aforementioned existing AIGC request scheduling method based on DRL include: First, DRL is prone to getting trapped in local optima and failing to effectively explore the global optimum, thus reducing the efficiency of traffic scheduling and the full utilization of system resources. Second, the DRL model requires significant computational resources and time, especially in edge network environments, increasing the latency of real-time scheduling. Finally, the training process is unstable and easily affected by network fluctuations and resource changes, particularly in complex and dynamic edge network environments, leading to poor convergence speed and effectiveness of the scheduling strategy.

[0006] Another existing reinforcement learning scheduling framework based on a diffusion model generates diverse scheduling policies. Through multi-step generation during AIGC traffic scheduling, this framework provides more flexible and efficient strategies, improving scheduling robustness. The distributed diffusion framework reduces resource consumption by sharing intermediate inference steps through request clustering and ensures generation quality through a refined CLIP scoring metric.

[0007] Another drawback of the aforementioned existing reinforcement learning scheduling framework based on a diffusion model includes: high computational complexity, especially in large-scale traffic scheduling scenarios, increasing time and computational costs and making it difficult to meet real-time requirements. Particularly in resource-constrained edge network environments, it may lead to increased latency. Furthermore, the diffusion model has poor adaptability, failing to quickly adjust scheduling strategies according to network changes, and its lengthy generation process may delay scheduling decisions, affecting the effectiveness of real-time response. Summary of the Invention

[0008] The embodiments of the present invention provide an AIGC traffic scheduling method based on a diffusion model and deep reinforcement learning, so as to effectively improve the scheduling efficiency of AIGC traffic in edge network environments.

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] An AIGC traffic scheduling method based on diffusion model and deep reinforcement learning is provided. The method constructs an AIGC service traffic scheduling architecture based on edge network environment and utilizing artificial intelligence technology. The edge AIGC scheduling framework is trained using a deep reinforcement learning training method based on diffusion enhancement to obtain a trained edge AIGC scheduling framework. The method includes:

[0011] The AIoT terminal generates AIGC service requests based on business needs. The AIGC service request includes task type, model size, latency constraints and data volume information, and forms corresponding AIGC service traffic. The AIoT terminal sends the AIGC service traffic to the edge gateway through the access network.

[0012] The edge gateway parses the received AIGC service traffic, extracts traffic feature information, fuses the traffic feature information with the network link status information to obtain system status information, and reports the system status information to the SDN controller.

[0013] Based on the received system state information, the SDN controller invokes a deep reinforcement learning scheduling mechanism based on a diffusion model to generate a traffic scheduling strategy that meets the current network state information and AIGC service requirements through a multi-step denoising strategy generation process.

[0014] The SDN controller distributes the generated traffic scheduling policy to the edge gateway. The edge gateway performs differentiated forwarding and queue scheduling operations on different AIGC service traffic according to the traffic scheduling policy, thereby realizing traffic priority control and dynamic bandwidth allocation.

[0015] Preferably, the method further includes: the AIGC service traffic scheduled by the edge gateway is forwarded to the corresponding edge server, which completes AIGC model loading, inference calculation and result generation according to resource availability, and returns the generated result to the corresponding AIoT terminal device;

[0016] The SDN controller continuously collects feedback information on the completion time, latency satisfaction, and resource utilization of AIGC service traffic, and inputs this feedback information as a reward signal into the deep reinforcement learning scheduling mechanism based on the diffusion model to achieve adaptive optimization of the traffic scheduling strategy.

[0017] Preferably, the construction of the AIGC service traffic scheduling architecture based on edge network environment using artificial intelligence technology involves training the edge AIGC scheduling framework using a deep reinforcement learning training method based on diffusion enhancement to obtain a trained edge AIGC scheduling framework, including:

[0018] An AIGC service traffic scheduling architecture based on an edge network environment is constructed, which includes AIoT terminals, edge gateways, SDN controllers, and edge servers.

[0019] The edge AIGC scheduling framework is trained using a deep reinforcement learning training method based on diffusion enhancement. This training method includes a closed-loop interactive learning mechanism based on Markov decision process and a training mechanism based on dual evaluation network. The closed-loop interactive learning mechanism based on Markov decision process perceives the network and computing environment in each decision time slot to form system status information. The system status information includes the service characteristics of the current incoming traffic, the queue occupancy and available bandwidth resources of each gateway node, and the available computing power on the edge server side.

[0020] The scheduling agent generates joint scheduling actions based on the system state information. After the system executes the joint scheduling actions, the environment provides immediate feedback rewards and the system transitions to the next state, forming a closed-loop interaction process of "state-action-reward-state". The reward function is used to characterize the comprehensive impact of scheduling decisions on system performance, and its definition is shown in (1):

[0021]

[0022] in, Indicates the degree of service latency deviation. Indicates the degree of uneven distribution of resources. This indicates the number of transactions that violate system constraints. , These are the weighting coefficients. As a penalty factor;

[0023] The training mechanism based on the dual-evaluation network includes a dual-evaluation network, a policy network, a target network, and an Adam optimizer. The dual-evaluation network consists of two evaluation networks with identical structures but independent parameters. Each network estimates the Q-value of the same state-action pair, and the target estimate is constructed by taking the minimum of the two outputs during training. The policy network approximates the agent's decision policy function, with the current system state as input and the corresponding scheduling action decision as output, guiding the agent's behavior selection under different environmental states. The target network is a delayed update network introduced during training. The Adam optimizer, as a parameter update algorithm, combines an adaptive learning rate mechanism of first-order moment estimation and second-order moment estimation to perform gradient optimization on the parameters of the dual-evaluation network, the policy network, and their corresponding target network. After the training process is completed, a trained edge AIGC scheduling framework is obtained.

[0024] Preferably, the edge gateway parses the received AIGC service traffic, extracts traffic feature information, fuses the traffic feature information with the network link status information to obtain system status information, and reports the system status information to the SDN controller, including:

[0025] The edge gateway parses the received AIGC service traffic and extracts traffic feature information, including traffic volume, arrival time, service priority, and corresponding quality of service requirements. The edge gateway also collects the current network link status information in real time, including link bandwidth utilization, queue length, and congestion status.

[0026] The edge gateway fuses the traffic characteristic information with the network link status information to obtain system status information, and then reports the system status information to the SDN controller.

[0027] Preferably, the SDN controller, based on the received system state information, invokes a deep reinforcement learning scheduling mechanism based on a diffusion model to generate a traffic scheduling strategy that meets the current network state information and AIGC service requirements through a multi-step denoising strategy generation process, including:

[0028] Step 1: Select an AIGC service flow to be scheduled, given the current system state. Under these conditions, an initial noise action vector is generated by sampling from a standard normal distribution with a mean of 0 and a variance of 1. This action vector is used as the initial input for the diffusion reverse generation process to characterize the random exploration starting point in the policy space;

[0029] Step 2: Convert the noise action vector of the current diffusion step System status and diffusion step index Common input diffusion policy network, for noise action vector Perform conditional denoising once to generate the noise action vector corresponding to the previous diffusion step. This process gradually guides the action vector to converge toward the feasible scheduling strategy space under system state constraints;

[0030] Step 3: Repeat Step 2, following the preset number of diffusion steps. The reverse denoising process is executed sequentially until the final noise-free motion representation is generated. This action represents the probability distribution characteristics of the scheduling strategy under the current network state and resource constraints.

[0031] Step 4: Final action representation based on the output of the diffusion model The probability density distribution of the scheduling policy is calculated using the SoftMax function, and the final scheduling action is obtained by sampling from the probability density distribution. The formula for calculating the probability density distribution of the scheduling strategy is:

[0032]

[0033] in, Let logits represent the policy corresponding to the nth candidate scheduling action;

[0034] The scheduling action 𝑎 simultaneously encodes the traffic mapping relationship, bandwidth allocation ratio, and computing resource allocation scheme. Under the premise of ensuring that high-priority and high-latency constraints of AIGC service flows are satisfied first, it comprehensively considers the gateway queue status, available bandwidth, and remaining edge computing resources to achieve joint optimization configuration of computing power and network resources.

[0035] As can be seen from the technical solutions provided by the embodiments of the present invention above, the present invention proposes a traffic scheduling method based on a diffusion model and deep reinforcement learning. This method enhances the exploratory nature of the scheduling strategy by introducing the generative capability of the diffusion model, thereby avoiding local optima, quickly adapting to dynamically changing traffic demands and network conditions, and improving scheduling efficiency and system response speed. This method aims to optimize the scheduling of AIGC traffic in edge network environments, ensuring low latency and high traffic completion rates, while addressing challenges such as resource contention and bandwidth limitations.

[0036] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 A structural diagram of an AIGC service traffic scheduling architecture based on an edge network environment provided in an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram illustrating the training of the aforementioned edge AIGC scheduling framework using a deep reinforcement learning training method based on diffusion enhancement, as provided in an embodiment of the present invention.

[0040] Figure 3 This is a flowchart illustrating an AIGC traffic scheduling method based on a diffusion model and deep reinforcement learning, provided as an embodiment of the present invention. Detailed Implementation

[0041] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0042] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0043] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0044] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0045] This invention first constructs an AIGC service traffic scheduling architecture based on an edge network environment, aiming to achieve efficient scheduling of latency-sensitive AIGC tasks. Second, it proposes an AIGC traffic scheduling method based on a diffusion model and deep reinforcement learning. This method can dynamically optimize bandwidth allocation and task scheduling through deep reinforcement learning to cope with the complex changes in edge networks. Finally, it designs a policy generation mechanism based on a generative diffusion model, which generates flexible scheduling policies through a multi-step decision-making process, thereby improving the system's scheduling capability and response speed in dynamic network environments.

[0046] The structural diagram of an AIGC service traffic scheduling architecture based on an edge network environment constructed in this embodiment of the invention is as follows: Figure 1As shown. For ease of explanation, this invention uses an AIoT scenario as a typical application scenario for generative AIGC services. The architecture employs a three-layer edge AIGC scheduling framework, including a workshop, an edge gateway, an SDN (Software-Defined Networking) controller, and an edge server. The workshop includes devices that serve as traffic sources, including AIoT (Artificial Intelligence of Things) terminals.

[0047] To address the issues of dynamic bandwidth fluctuations and multi-priority traffic contention during edge AIGC service scheduling, such as Figure 2 As shown, this embodiment of the invention employs a deep reinforcement learning training method based on diffusion enhancement to train the aforementioned edge AIGC scheduling framework. This training method includes a closed-loop interactive learning mechanism based on Markov decision processes and a training mechanism based on a dual-evaluation network.

[0048] The closed-loop interactive learning mechanism based on Markov decision processes perceives the network and computing environment in each decision time slot, forming system state information. This system state information comprehensively reflects the service characteristics of the current incoming traffic (including data scale, latency constraints, and priority), the queue occupancy and available bandwidth resources of each gateway node, and the available computing power on the edge server side. The scheduling agent generates joint scheduling actions based on this state. The actions cover the mapping relationship between traffic and gateway nodes, bandwidth resource allocation schemes, and computing resource allocation strategies, thereby achieving collaborative optimization of communication and computing resources. After the system executes the scheduling action, the environment provides an immediate reward and transitions to the next state, forming a closed-loop interactive process of "state—action—reward—state". The reward function is used to characterize the comprehensive impact of scheduling decisions on system performance, and it is defined as Equation (1):

[0049]

[0050] in, Indicates the degree of service latency deviation. Indicates the degree of uneven distribution of resources. This indicates the number of transactions that violate system constraints. , These are the weighting coefficients. This serves as a penalty factor. By maximizing long-term cumulative rewards, the agent gradually learns scheduling strategies that can improve resource utilization efficiency while satisfying latency constraints.

[0051] The training mechanism based on the dual-evaluation network mainly consists of a dual-evaluation network, a policy network, a target network, and the Adam optimizer. The dual-evaluation network comprises two structurally identical but parameter-independent evaluation networks. Each network estimates the Q-value of the same state-action pair, and the target estimate is constructed by taking the minimum of their outputs during training. This effectively suppresses Q-value overestimation and improves the stability and conservatism of the value assessment. The policy network approximates the agent's decision-making policy function. Its input is the current system state, and its output is the corresponding scheduling action decision, guiding the agent's behavioral choices under different environmental states. The target network is a delayed-update network introduced during training. Its network structure is consistent with both the evaluation and policy networks, but its parameters are updated through soft updates or periodic synchronization, providing a relatively smooth and stable reference during target value calculation and effectively reducing the risk of training oscillations. The Adam optimizer, as a parameter update algorithm, combines an adaptive learning rate mechanism of first-order moment estimation and second-order moment estimation. It is used to efficiently and stably optimize the parameters of the dual-evaluation network, policy network, and their corresponding target network, accelerating model convergence and improving overall training performance. After the training process is completed, the trained edge AIGC scheduling framework is obtained.

[0052] Based on the pre-trained edge AIGC scheduling framework described above, the processing flow of an AIGC traffic scheduling method based on a diffusion model and deep reinforcement learning provided in this embodiment of the invention is as follows: Figure 3 As shown, the processing steps include the following:

[0053] Step S10: The AIoT (Artificial Intelligence of Things) terminal in the workshop generates an AIGC service request based on business needs. The AIGC service request includes task type, model size, latency constraints and data volume information, and forms a corresponding AIGC service traffic. The AIoT terminal sends the AIGC service traffic to the edge gateway through the access network.

[0054] Step S20: The edge gateway parses the received AIGC service traffic and extracts traffic feature information, including traffic size, arrival time, service priority, and corresponding quality of service requirements; the edge gateway collects the current network link status information in real time, including link bandwidth utilization, queue length, and congestion status.

[0055] Step S30: The edge gateway fuses the traffic characteristic information with the network link status information to obtain system status information, and reports the system status information to the SDN controller for subsequent scheduling decisions.

[0056] Step S40: Based on the received system status information, the SDN controller calls the deep reinforcement learning scheduling module based on the diffusion model. Through a multi-step denoising strategy generation process, a traffic scheduling strategy that meets the current network status information and AIGC service requirements is generated. The traffic scheduling strategy includes a traffic priority allocation scheme and a bandwidth resource allocation scheme.

[0057] Step S50: The SDN controller sends the generated traffic scheduling policy to the edge gateway. The edge gateway performs differentiated forwarding and queue scheduling operations on different AIGC service traffic according to the traffic scheduling policy, so as to realize traffic priority control and dynamic bandwidth allocation.

[0058] Step S60: The scheduled AIGC service traffic is forwarded to the corresponding edge server. The edge server completes AIGC model loading, inference calculation and result generation according to resource availability, and returns the generated result to the corresponding AIoT terminal device.

[0059] Step S70: The SDN controller continuously collects feedback information such as the completion time, latency satisfaction, and resource utilization of AIGC service traffic, and inputs this feedback as a reward signal into the deep reinforcement learning model to update the diffusion scheduling strategy, thereby achieving adaptive optimization of scheduling performance.

[0060] In step S40 above, the SDN controller introduces a policy generation mechanism based on a generative diffusion model during the scheduling decision process. This mechanism generates a probability distribution of scheduling policies step-by-step under system state constraints through multi-step random denoising, achieving efficient modeling of the complex, multimodal scheduling policy space. Based on this mechanism, the generation of traffic scheduling policies maintains both random exploration capabilities and network and computational resource constraints, thereby improving the stability and adaptability of AIGC service flow scheduling. The traffic scheduling policy generation steps of this mechanism are as follows:

[0061] Step 1: Initial noise action generation. Select an AIGC service flow to be scheduled, given the current system state. Under these conditions, an initial noise action vector is generated by sampling from a standard normal distribution with a mean of 0 and a variance of 1. This action vector is used as the initial input to the diffusion reverse generation process to characterize the random exploration starting point in the policy space.

[0062] Step 2: Conditional diffusion reverse generation, generating the action vector of the current diffusion step. System status and diffusion step index The common input diffusion policy network performs conditional denoising on noisy actions, generating the action vector corresponding to the previous diffusion step. This process gradually guides the action vector to converge toward the feasible scheduling strategy space under system state constraints.

[0063] Step 3: Repeat Step 2, following the preset number of diffusion steps. The reverse denoising process is executed sequentially until the final noise-free motion representation is generated. This action represents the probability distribution characteristics of the scheduling strategy under the current network state and resource constraints.

[0064] Step 4: Calculate the policy probability density and sample actions, based on the final action representation output by the diffusion model. The probability density distribution of the scheduling policy is calculated using the SoftMax function, and the final scheduling action is obtained by sampling from it. Its policy probability density function is:

[0065]

[0066] in, This represents the strategy logits corresponding to the nth candidate scheduling action.

[0067] Through the aforementioned strategy probability density calculation and action sampling process, the scheduling decision that best matches the current network state information and AIGC service requirements can be adaptively selected from multiple candidate scheduling actions. Specifically, the sampled scheduling action 𝑎 simultaneously encodes the traffic mapping relationship, bandwidth allocation ratio, and computing resource allocation scheme. This action, while ensuring that high-priority and high-latency-constraint AIGC service flows are prioritized, comprehensively considers the gateway queue status, available bandwidth, and remaining edge computing resources to achieve joint optimization of computing power and network resources. Finally, this scheduling action is issued and executed as the traffic scheduling strategy for the current moment, thereby achieving efficient, stable, and adaptive scheduling of AIGC service flows in a dynamic network environment.

[0068] Through the above steps, the generated traffic scheduling strategy can not only fully explore diverse feasible scheduling schemes during the diffusion generation process, but also be constrained by the current system state in each denoising process. Thus, it comprehensively considers dynamic factors such as network bandwidth, queue state, computing resources and flow priority, and realizes adaptive optimization of AIGC service flow bandwidth and priority allocation strategy, thereby improving overall resource utilization efficiency and scheduling robustness.

[0069] In summary, this invention provides a unified model for AIGC service flows in edge network scenarios, incorporating network bandwidth, gateway queue status, and edge network resources into the same scheduling decision process to achieve collaborative optimization of computing and network resources. Compared to existing technologies that primarily focus on inference processes and request clustering and sharing, this invention is more suitable for resource-constrained and highly dynamic edge network environments.

[0070] This invention constructs a closed-loop interactive learning mechanism based on Markov decision processes, enabling the scheduling strategy to continuously and adaptively update according to changes in real-time network load and computing resources. This maintains stable scheduling performance in high-concurrency request scenarios and effectively avoids the problem of significantly increased scheduling latency caused by intensified resource contention.

[0071] This invention introduces a generative diffusion model into the deep reinforcement learning policy generation process, and conducts structured exploration in the policy space through multi-step denoising, which improves policy diversity and robustness while ensuring policy feasibility. Unlike existing technologies that rely on request clustering and static sharing mechanisms, this invention can directly characterize the competitive relationship between multiple priority AIGC streams, making it more suitable for large-scale, real-time scheduling scenarios.

[0072] This invention achieves balanced resource utilization while satisfying latency and QoS constraints through a multi-objective joint reward design and a dual-evaluation network training mechanism, and ensures policy convergence through a stable training mechanism. Therefore, in high-concurrency and highly dynamic edge AIGC service scheduling scenarios, this invention outperforms existing technologies in terms of scheduling latency, system stability, and scalability.

[0073] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0074] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0075] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0076] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An AIGC traffic scheduling method based on diffusion model and deep reinforcement learning, characterized in that, A traffic scheduling architecture for AIGC (Artificial Intelligence Generate Content) services based on an edge network environment is constructed. The edge AIGC scheduling framework is trained using a deep reinforcement learning training method based on diffusion enhancement to obtain a trained edge AIGC scheduling framework. The method includes: The AIoT terminal generates AIGC service requests based on business needs. The AIGC service request includes task type, model size, latency constraints and data volume information, and forms corresponding AIGC service traffic. The AIoT terminal sends the AIGC service traffic to the edge gateway through the access network. The edge gateway parses the received AIGC service traffic, extracts traffic feature information, fuses the traffic feature information with the network link status information to obtain system status information, and reports the system status information to the SDN controller. Based on the received system state information, the SDN controller invokes a deep reinforcement learning scheduling mechanism based on a diffusion model to generate a traffic scheduling strategy that meets the current network state information and AIGC service requirements through a multi-step denoising strategy generation process. The SDN controller distributes the generated traffic scheduling policy to the edge gateway. The edge gateway performs differentiated forwarding and queue scheduling operations on different AIGC service traffic according to the traffic scheduling policy, thereby realizing traffic priority control and dynamic bandwidth allocation.

2. The method according to claim 1, characterized in that, The method further includes: the AIGC service traffic scheduled by the edge gateway is forwarded to the corresponding edge server, which completes AIGC model loading, inference calculation and result generation according to resource availability, and returns the generated result to the corresponding AIoT terminal device; The SDN controller continuously collects feedback information on the completion time, latency satisfaction, and resource utilization of AIGC service traffic, and inputs this feedback information as a reward signal into the deep reinforcement learning scheduling mechanism based on the diffusion model to achieve adaptive optimization of the traffic scheduling strategy.

3. The method according to claim 1 or 2, characterized in that, The aforementioned construction of an AIGC service traffic scheduling architecture based on edge network environment and utilizing artificial intelligence technology involves training the edge AIGC scheduling framework using a deep reinforcement learning training method based on diffusion enhancement to obtain a trained edge AIGC scheduling framework, including: An AIGC service traffic scheduling architecture based on an edge network environment is constructed, which includes AIoT terminals, edge gateways, SDN controllers, and edge servers. The edge AIGC scheduling framework is trained using a deep reinforcement learning training method based on diffusion enhancement. This training method includes a closed-loop interactive learning mechanism based on Markov decision process and a training mechanism based on dual evaluation network. The closed-loop interactive learning mechanism based on Markov decision process perceives the network and computing environment in each decision time slot to form system status information. The system status information includes the service characteristics of the current incoming traffic, the queue occupancy and available bandwidth resources of each gateway node, and the available computing power on the edge server side. The scheduling agent generates joint scheduling actions based on the system state information. After the system executes the joint scheduling actions, the environment provides immediate feedback rewards and the system transitions to the next state, forming a closed-loop interaction process of "state-action-reward-state". The reward function is used to characterize the comprehensive impact of scheduling decisions on system performance, and its definition is shown in (1): in, Indicates the degree of service latency deviation. Indicates the degree of uneven distribution of resources. This indicates the number of transactions that violate system constraints. , These are the weighting coefficients. As a penalty factor; The training mechanism based on the dual-evaluation network includes a dual-evaluation network, a policy network, a target network, and an Adam optimizer. The dual-evaluation network consists of two evaluation networks with identical structures but independent parameters. Each network estimates the Q-value of the same state-action pair, and the target estimate is constructed by taking the minimum of the two outputs during training. The policy network approximates the agent's decision policy function, with the current system state as input and the corresponding scheduling action decision as output, guiding the agent's behavior selection under different environmental states. The target network is a delayed update network introduced during training. The Adam optimizer, as a parameter update algorithm, combines an adaptive learning rate mechanism of first-order moment estimation and second-order moment estimation to perform gradient optimization on the parameters of the dual-evaluation network, the policy network, and their corresponding target network. After the training process is completed, a trained edge AIGC scheduling framework is obtained.

4. The method according to claim 3, characterized in that, The edge gateway parses the received AIGC service traffic, extracts traffic feature information, fuses the traffic feature information with network link status information to obtain system status information, and reports the system status information to the SDN controller, including: The edge gateway parses the received AIGC service traffic and extracts traffic feature information, which includes traffic size, arrival time, service priority, and corresponding service quality requirements. The edge gateway collects the current network link status information in real time, including link bandwidth utilization, queue length, and congestion status. The edge gateway fuses the traffic characteristic information with the network link status information to obtain system status information, and then reports the system status information to the SDN controller.

5. The method according to claim 4, characterized in that, The SDN controller, based on the received system state information, invokes a deep reinforcement learning scheduling mechanism based on a diffusion model to generate a traffic scheduling strategy that meets the current network state information and AIGC service requirements through a multi-step denoising strategy generation process, including: Step 1: Select an AIGC service flow to be scheduled, given the current system state. Under these conditions, an initial noise action vector is generated by sampling from a standard normal distribution with a mean of 0 and a variance of 1. This action vector is used as the initial input for the diffusion reverse generation process to characterize the random exploration starting point in the policy space; Step 2: Convert the noise action vector of the current diffusion step System status and diffusion step index Common input diffusion policy network, for noise action vector Perform conditional denoising once to generate the noise action vector corresponding to the previous diffusion step. This process gradually guides the action vector to converge toward the feasible scheduling strategy space under system state constraints; Step 3: Repeat Step 2, following the preset number of diffusion steps. The reverse denoising process is executed sequentially until the final noise-free motion representation is generated. This action represents the probability distribution characteristics of the scheduling strategy under the current network state and resource constraints. Step 4: Final action representation based on the output of the diffusion model The probability density distribution of the scheduling policy is calculated using the SoftMax function, and the final scheduling action is obtained by sampling from the probability density distribution. The formula for calculating the probability density distribution of the scheduling strategy is: in, Let logits represent the policy corresponding to the nth candidate scheduling action; The scheduling action 𝑎 simultaneously encodes the traffic mapping relationship, bandwidth allocation ratio, and computing resource allocation scheme. Under the premise of ensuring that high-priority and high-latency constraints of AIGC service flows are satisfied first, it comprehensively considers the gateway queue status, available bandwidth, and remaining edge computing resources to achieve joint optimization configuration of computing power and network resources.

Images