A communication network architecture endogenous to generative AI and a hierarchical collaborative scheduling method

Abstract

The present application discloses a communication network architecture and a hierarchical cooperative scheduling method endogenous to generative AI, which relates to the technical field of communication networks. The communication network architecture includes an intention extraction layer, a mapping adaptation layer, and a network execution layer. The intention extraction layer uses a generative AI model to process multimodal data to obtain a semantic requirement representation of the user. The mapping adaptation layer uses a generative AI model to map the semantic requirement representation to a quality of service indicator, maps the real-time network state of heterogeneous networks in the network execution layer to a network performance indicator, and generates a transmission scheduling policy based on the quality of service indicator and the network performance indicator. The network execution layer operates based on the transmission scheduling policy to achieve end-to-end deterministic transmission scheduling of service data packets. The present application can achieve efficient deterministic transmission scheduling of intelligent computing services in 6G networks.

Description

Technical Field

[0001] The present application relates to the technical field of communication networks, and particularly to a communication network architecture with generative AI (Artificial Intelligence) embedded and a hierarchical collaborative scheduling method. Background Art

[0002] With the continuous improvement of society's requirements for networks, 6G communication systems have become the focus of research by international industrial, academic, and standardization groups. Compared with 5G networks, 6G networks not only have significant improvements in performance indicators such as bandwidth and latency, but more importantly, they achieve ubiquitous intelligence and deep integration of computing and networks. Artificial intelligence technology will be deeply embedded in the 6G network architecture as a core driving force. Especially in large-scale intelligent access scenarios such as the Internet of Vehicles, 6G networks need to process and transmit massive amounts of high-dimensional multi-modal information for intelligent computing services, which poses a huge challenge to traditional optimization algorithms.

[0003] The deterministic transmission scheduling of intelligent computing services in 6G networks faces multiple technical problems. First, at the information processing level, it is necessary to solve the problem of unified feature extraction of multi-modal data, which directly affects the perception efficiency of network status and also requires accurate understanding and expression of users' semantic requirements. Second, the frequent interaction between terminal nodes and computing centers often spans multiple heterogeneous networks, which makes the deterministic transmission scheduling of multi-modal data more complex. Facing these challenges, generative artificial intelligence provides a new solution idea for solving the problem of efficient deterministic transmission scheduling in 6G scenarios due to its unique generation ability and generalization ability.

[0004] In response to the above challenges, there is an urgent need to design a new communication network architecture with generative AI embedded and a hierarchical collaborative scheduling method to support emerging intelligent computing services in 6G networks. Summary of the Invention

[0005] The purpose of the present application is to provide a communication network architecture with generative AI embedded and a hierarchical collaborative scheduling method, which can achieve efficient deterministic transmission scheduling of intelligent computing services in 6G networks and support emerging intelligent computing services in 6G networks.

[0006] To achieve the above purpose, the present application provides the following solutions:

[0007] In a first aspect, the present application provides a communication network architecture with generative AI embedded, and the communication network architecture with generative AI embedded includes:

[0008] An intention extraction layer, configured to process multi-modal data by using a generative AI model to obtain a semantic requirement representation of a user; the multi-modal data is requirement data of a user during the service execution of an intelligent computing service in a 6G network;

[0009] A mapping adaptation layer, which is respectively connected to the intent extraction layer and the network execution layer, is used to map the semantic requirement representation to service quality indicators by using a generative AI model, map the real-time network status of heterogeneous networks in the network execution layer to network performance indicators, and generate a transmission scheduling strategy based on the service quality indicators and the network performance indicators;

[0010] The network execution layer is used to work based on the transmission scheduling strategy to achieve end-to-end deterministic transmission scheduling of business data packets.

[0011] Optionally, the intent extraction layer includes:

[0012] A multimodal processing module, which is used to extract features from multimodal data to obtain user data;

[0013] A generation processing module, which is used to reason about user behavior and preferences based on user data and predict and generate synthetic data;

[0014] A semantic requirement processing module, which is used to perform semantic extraction, extension and compression by combining user data and synthetic data to obtain the user's semantic requirement representation;

[0015] Among them, both the generation processing module and the semantic requirement processing module adopt generative AI models.

[0016] Optionally, the multimodal data includes text, audio, pictures and location information. At this time, the multimodal processing module includes a CNN model, a ViT model and a GNN model. The CNN model is used to extract features from text to obtain a first feature. The CNN model is used to extract features from audio to obtain a second feature. The ViT model is used to extract features from pictures to obtain a third feature. The GNN model is used to extract features from location information to obtain a fourth feature. The first feature, the second feature, the third feature and the fourth feature are combined to form user data.

[0017] Optionally, the mapping adaptation layer includes:

[0018] A service requirement knowledge base, which is used to map the semantic requirement representation to service quality indicators; the service quality indicators include: basic performance indicators, reliability indicators, resource indicators and business feature indicators. The basic performance indicators include bandwidth, latency and service priority. The reliability indicators include packet loss rate and jitter. The resource indicators include computing resources and network resources. The business feature indicators include packet size, traffic pattern and service duration;

[0019] A network resource knowledge base for mapping the real-time network status of heterogeneous networks in the network execution layer into network performance metrics; the network performance metrics include: link status metrics, cache resource metrics, timing performance metrics, and device capability metrics. The link status metrics include available transmission bandwidth and link utilization rate. The cache resource metrics include queue length and cache utilization rate. The timing performance metrics include current link delay and jitter range. The device capability metrics include port status;

[0020] An adaptation decision module for dynamically matching requirements and resources based on service quality metrics and network performance metrics to generate a transmission scheduling strategy; the transmission scheduling strategy includes: a routing decision matrix, a bandwidth allocation matrix, and the configuration of heterogeneous networks. The routing decision matrix includes the next-hop selection of each flow at each node. The bandwidth allocation matrix includes the bandwidth allocated to each flow on each link. The node is a network node in the heterogeneous network;

[0021] Among them, generative AI models are deployed in both the service demand knowledge base and the network resource knowledge base. The adaptation decision module uses an end-to-end deterministic transmission scheduling algorithm based on the diffusion model enhanced D3QN to generate a transmission scheduling strategy.

[0022] Optionally, the network execution layer includes a heterogeneous network, which includes a first edge network, a core network, and a second edge network connected in sequence. Both the first edge network and the second edge network adopt time-sensitive networks, and the core network adopts a deterministic network; among them, time-aware shapers are deployed in the first edge network and the second edge network, and a time-aware shaper and a credit-based shaper are deployed in the core network.

[0023] Optionally, the network execution layer further includes:

[0024] A terminal processing module, connected to the mapping adaptation layer, for obtaining the transmission scheduling strategy from the mapping adaptation layer in real time;

[0025] A global processing module, connected to the terminal processing module, for storing the transmission scheduling strategy obtained by the terminal processing module in the long term and sending it to the heterogeneous network in the storage order.

[0026] Optionally, the configuration of the heterogeneous network includes the gating list of each node in the first edge network and the second edge network, as well as the gating list, idle rate, and transmission rate of each node in the core network.

[0027] In a second aspect, the present application provides a hierarchical cooperative scheduling method for a generative AI endogenous communication network architecture, which is applied to the generative AI endogenous communication network architecture described in any one of the above. The hierarchical cooperative scheduling method for the generative AI endogenous communication network architecture includes:

[0028] The intent extraction layer processes multimodal data using a generative AI model to obtain a semantic requirement representation of the user;

[0029] The mapping and adaptation layer uses a generative AI model to map the semantic requirement representation to service quality indicators, maps the real-time network states of heterogeneous networks in the network execution layer to network performance indicators, and generates a transmission scheduling policy based on the service quality indicators and network performance indicators;

[0030] The network execution layer works based on the transmission scheduling policy to achieve end-to-end deterministic transmission scheduling of service data packets.

[0031] Optionally, generating a transmission scheduling policy based on service quality indicators and network performance indicators specifically includes:

[0032] Standardize the service quality indicators and network performance indicators to obtain standardized data; the standardized data includes network topology, link load, queue status, flow characteristics, and latency requirements;

[0033] Use the standardized data as the state vector, and generate an action vector based on the state vector, and use the action vector as the transmission scheduling policy.

[0034] Optionally, generating an action vector based on the state vector specifically includes: using the state vector as input and determining the action vector using a trained policy network;

[0035] Among them, the reward function used when training the trained policy network is:

[0036] R(s, a) = w1·R delay + w2·R jit + w3·R thr - w4·C vio ;

[0037] Among them, R(s, a) is the reward; w1 is the first weight; R delay is the latency; w2 is the second weight; R jit is the jitter; w3 is the third weight; R thr is the throughput; w4 is the fourth weight; C vip is the penalty for violating the service quality indicator.

[0038] According to the specific embodiments provided by this application, this application has the following technical effects:

[0039] The present application provides a communication network architecture and a hierarchical collaborative scheduling method endogenous to generative AI, including an intention extraction layer, a mapping adaptation layer, and a network execution layer. The intention extraction layer uses a generative AI model to process multimodal data to obtain a semantic requirement representation of the user. The mapping adaptation layer uses a generative AI model to map the semantic requirement representation to service quality indicators, maps the real-time network state of heterogeneous networks in the network execution layer to network performance indicators, and generates a transmission scheduling strategy based on the service quality indicators and network performance indicators. The network execution layer works based on the transmission scheduling strategy to achieve end-to-end deterministic transmission scheduling of service data packets. By integrating the generative AI model into the intention extraction layer and the mapping adaptation layer, the present application can solve the problems of unified extraction of features of multimodal data, the need to accurately understand and express the semantic requirements of users, and the more complex problem of deterministic transmission scheduling of multimodal data, so as to achieve efficient deterministic transmission scheduling of intelligent computing services in the 6G network and support emerging intelligent computing services in the 6G network. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0041] Figure 1 It is a schematic diagram of a novel communication network architecture endogenous to generative AI for the 6G network provided in Embodiment 1 of the present application.

[0042] Figure 2 It is a schematic flowchart of a method for extracting semantic requirement representation provided in Embodiment 1 of the present application.

[0043] Figure 3 It is a schematic flowchart of a method for mapping requirements and resources provided in Embodiment 1 of the present application.

[0044] Figure 4 It is a schematic flowchart of an end-to-end deterministic transmission scheduling algorithm based on a diffusion model enhanced D3QN (Dueling Double Deep Q-network) provided in Embodiment 1 of the present application.

[0045] Figure 5 It is a schematic flowchart of a hierarchical collaborative scheduling method for a communication network architecture endogenous to generative AI provided in Embodiment 2 of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0047] Embodiment 1

[0048] This embodiment provides a communication network architecture endogenous to generative AI. As Figure 1 shown, the communication network architecture endogenous to generative AI includes: an intent extraction layer, a mapping adaptation layer, and a network execution layer.

[0049] The intent extraction layer is used to process multimodal data by using a generative AI model to obtain a semantic requirement representation of the user. The multimodal data is the requirement data of the user during the business execution of intelligent computing services in the 6G network.

[0050] The mapping adaptation layer is respectively connected to the intent extraction layer and the network execution layer, and is used to map the semantic requirement representation to service quality indicators by using a generative AI model, map the real-time network status of heterogeneous networks in the network execution layer to network performance indicators, and generate a transmission scheduling strategy based on the service quality indicators and the network performance indicators.

[0051] The network execution layer is used to work based on the transmission scheduling strategy to achieve end-to-end deterministic transmission scheduling of business data packets.

[0052] In this embodiment, the intent extraction layer is used to process various multimodal data to obtain user data, and fuse the actual user data and the predicted synthetic data to achieve the extraction of the semantic requirement representation of the user, so as to realize the processing of the user's multimodal data and the extraction of the semantic requirement representation. The intent extraction layer deploys a multimodal processing module, a generation processing module, and a semantic requirement processing module to efficiently extract the semantic requirement representation of the user, facilitating subsequent service requirement mapping.

[0053] Specifically, the intent extraction layer includes: a multimodal processing module, a generation processing module, and a semantic requirement processing module.

[0054] The multimodal processing module is used to extract features from multimodal data to obtain user data.

[0055] The generation processing module is used to infer the user's behavior and preferences based on the user data and predict and generate synthetic data.

[0056] The semantic requirement processing module is used to extract, expand, and compress semantics by combining user data and synthetic data, obtaining the semantic requirement representation of the user. By fusing actual user data (as the main input) and predicted synthetic data (as the auxiliary input), the extraction, expansion, and compression of semantics are realized, forming an accurate semantic requirement representation.

[0057] Among them, both the generation processing module and the semantic requirement processing module adopt generative AI models. The generative AI model adopted by the generation processing module can be an LLM (Large Language Model), and the generative AI model adopted by the semantic requirement processing module can be an LLM.

[0058] Among them, the multimodal processing module is used to process multimodal data generated in 6G application scenarios, such as text, audio, pictures, and location information. It uses a deep learning model to uniformly map data of different modalities into the feature space, forming a standardized user data representation to obtain user data. That is, in this embodiment, the multimodal data can include text, audio, pictures, and location information. At this time, the multimodal processing module includes a CNN (Convolutional Neural Networks) model, a ViT (Vision Transformer) model, and a GNN (Graph Neural Networks) model. The CNN model is used to extract features from text to obtain the first feature, the CNN model is used to extract features from audio to obtain the second feature, the ViT model is used to extract features from pictures to obtain the third feature, and the GNN model is used to extract features from location information to obtain the fourth feature. The first feature, the second feature, the third feature, and the fourth feature are combined to form user data.

[0059] In this embodiment, by integrating a generative AI model into the intent extraction layer, the problems of unified feature extraction of multimodal data and accurate understanding and expression of the user's semantic requirements in the information processing level are solved.

[0060] In this embodiment, the mapping and adaptation layer generates a transmission scheduling policy through the mapping and adaptation of requirements and resources. Two knowledge bases and an adaptation decision module are deployed in this mapping and adaptation layer to achieve the precise mapping and adaptation of service requirements and network resources as well as policy generation. The semantic requirement representation and the real-time network state are respectively mapped to specific QoS (Quality of Service) metrics and network performance metrics through the two knowledge bases, and then the adaptation decision module generates a transmission scheduling policy after adapting the requirements and resources. The service requirement knowledge base maps the semantic requirement representation extracted by the intention extraction layer to specific QoS metrics with the help of the LLM. The LLM defines an effective search space and specifies the mapping path from the semantic requirement representation to the QoS metrics. The network resource knowledge base maps the real-time network state uploaded by the network execution layer to specific network performance metrics, such as link utilization rate, queue length, etc., which are used to quantify network performance. The adaptation decision module performs dynamic matching of requirements and resources according to the QoS metrics and network performance metrics, that is, adapts the requirements and resources, and generates the optimal transmission scheduling policy based on the diffusion model enhanced D3QN end-to-end deterministic transmission scheduling algorithm, and issues it to the network execution layer, that is, the transmission scheduling policy will be deployed to each network node (i.e., device) of the network execution layer subsequently. This embodiment calls the diffusion model enhanced D3QN end-to-end deterministic transmission scheduling algorithm, and dynamically controls the transmission routing and configuration of data packets according to the adaptation situation of resources and requirements. In this way, the resource allocation in the 6G network can be optimized, and the deterministic transmission scheduling of intelligent computing services can be achieved.

[0061] Specifically, the mapping and adaptation layer includes: a service requirement knowledge base, a network resource knowledge base, and an adaptation decision module.

[0062] The service requirement knowledge base is used to map the semantic requirement representation to QoS metrics. The QoS metrics include: basic performance metrics, reliability metrics, resource metrics, and service characteristic metrics. The basic performance metrics include bandwidth, latency, and service priority. The reliability metrics include packet loss rate and jitter. The resource metrics include computing resources and network resources. The service characteristic metrics include packet size, traffic pattern, and service duration.

[0063] The network resource knowledge base is used to map the real-time network state of heterogeneous networks in the network execution layer to network performance metrics. The network performance metrics include: link state metrics, cache resource metrics, timing performance metrics, and device capability metrics. The link state metrics include available transmission bandwidth and link utilization rate. The cache resource metrics include queue length and cache utilization rate. The timing performance metrics include current link latency and jitter range. The device capability metrics include port state.

[0064] An adaptation decision module, which is used to perform dynamic matching of requirements and resources based on quality of service indicators and network performance indicators, and generate a transmission scheduling policy. The transmission scheduling policy includes: a routing decision matrix, a bandwidth allocation matrix, and the configuration of a heterogeneous network. The routing decision matrix includes the next-hop selection of each flow at each node. The bandwidth allocation matrix includes the bandwidth allocated to each flow on each link. The node is a network node in the heterogeneous network.

[0065] Among them, generative AI models are deployed in both the service demand knowledge base and the network resource knowledge base. The generative AI models deployed in the service demand knowledge base and the network resource knowledge base can both be LLMs. The adaptation decision module uses an end-to-end deterministic transmission scheduling algorithm based on the diffusion model enhanced D3QN to generate the transmission scheduling policy. The diffusion model belongs to the generative AI model.

[0066] In this embodiment, the network execution layer provides the required physical network devices. According to the transmission scheduling policy of the mapping adaptation layer and combined with the deterministic transmission mechanism, the data packets are transmitted deterministically to achieve end-to-end deterministic transmission scheduling of the data packets. The network execution layer includes: a terminal LLM, a global LLM, and a heterogeneous network. The end-user downloads the terminal LLM from the cache server deployed at the network edge and personalizes and initializes it. The terminal LLM obtains the transmission scheduling policy from the mapping adaptation layer. The global LLM stores the transmission scheduling policy aggregated by the terminal LLMs as long-term memory data and optimizes the transmission scheduling policy through text learning and model fine-tuning to obtain an optimized transmission scheduling policy. Subsequently, based on this optimized transmission scheduling policy, the heterogeneous network is controlled to work.

[0067] The heterogeneous network includes an edge network and a core network. The edge network adopts Time-Sensitive Networking (TSN), and the core network adopts Deterministic Networking (DetNet). TSN enables all devices in the network to send and receive data packets according to a preset time sequence by deploying a unified time synchronization mechanism and a scheduling list within the local area network, thereby achieving controllability of the data transmission delay within the local area network. DetNet reserves dedicated network resources, including bandwidth and queues, for critical business traffic in the wide area network, thereby ensuring its end-to-end deterministic transmission performance and reliability requirements.

[0068] Deploy deterministic transmission mechanisms in the edge network and the core network, such as Time Awareness Shaper (TAS) and Credit-based Shaping (CBS), and use a gating mechanism to meet the requirements of high-real-time services for transmission delay and jitter in smaller time granularity. TAS adopts a time-sensitive gating scheduling mechanism to control the transmission status of the data packet queue through the time dimension. According to the pre-configured Gate Control List (GCL), the on and off states of the queue are periodically switched: when the queue is in the on state, data packets are allowed to be transmitted; when the queue is in the off state, the data packets will wait in the queue until the state switches to on. After the gating list is determined, this scheduling mechanism will execute in a loop according to the set time period. CBS controls the data transmission scheduling by assigning a credit value to each queue. The credit value of each queue changes over time, and the rates of increase and decrease are subject to corresponding parameter constraints. Bandwidth is reserved for different priority traffic through a credit-based control mode to ensure the upper limit of transmission delay.

[0069] TSN realizes deterministic transmission within the local area network through the TAS mechanism. DetNet guarantees the end-to-end transmission performance of critical service traffic in the wide area network through the integrated mechanism of TAS and CBS, and meets the requirements of high-real-time services through refined gating scheduling in the time dimension.

[0070] Specifically, the network execution layer includes heterogeneous networks, which include a first edge network, a core network, and a second edge network connected in sequence. Both the first edge network and the second edge network adopt time-sensitive networks, and the core network adopts a deterministic network. Among them, time awareness shapers are deployed in the first edge network and the second edge network, and time awareness shapers and credit-based shapers are deployed in the core network.

[0071] At this time, in the transmission scheduling strategy, the configuration of the heterogeneous network includes the gating list of each node in the first edge network and the second edge network, as well as the gating list, idle rate, and sending rate of each node in the core network.

[0072] The network execution layer also includes: a terminal processing module and a global processing module.

[0073] The terminal processing module (i.e., Figure 1 the terminal LLM therein), is connected to the mapping adaptation layer and is used to obtain the transmission scheduling strategy from the mapping adaptation layer in real time.

[0074] The global processing module (i.e., Figure 1The global LLM) is connected to the terminal processing module, used to store the transmission scheduling policy obtained by the terminal processing module in the long term, and send it to the heterogeneous network in the storage order.

[0075] Based on the above novel communication network architecture generated by generative AI, this embodiment further provides a hierarchical collaborative scheduling method, which is a hierarchical collaborative intelligent method, including: semantic requirement representation extraction, mapping of requirements and resources, and an end-to-end deterministic transmission scheduling algorithm based on the diffusion model enhanced D3QN.

[0076] (1) Semantic requirement representation extraction.

[0077] Semantic requirement representation extraction is the function completed by the intention extraction layer. As Figure 2 shown, it includes: (a) Using a distributed data acquisition system to collect text, audio, picture, and location information data in 6G application scenarios in real time to obtain the user's original multimodal data; (b) Establishing deep learning models, including a CNN model for processing time series data, a ViT model for processing image data, and a GNN model for processing graph structure data. Different deep learning models are used to extract features from the generated multimodal data: the CNN model extracts features of time series data (such as text and audio), the ViT model extracts features of image data (such as pictures), and the GNN model extracts features of graph structure data (such as location information). The extracted features are used as unified tokens, that is, the collected multimodal data such as text, audio, pictures, and location information are input into different deep learning models for feature extraction. The feature extraction process is specifically as follows: Input the time series data into the CNN model to extract time series pattern features through convolution and pooling operations, input the image data into the ViT model to extract visual features through patch segmentation and multi-head self-attention layers, input the graph structure data into the GNN model to extract spatial relationship features through graph attention layers, generate a unified serialized token representation, and obtain user data; (c) According to the user data, the LLM predicts some user behavior data and preference data as synthetic data. When predicting synthetic data, after unifying the multimodal data into a serialized token representation, analyze the distribution of the serialized token representation, and then generate new samples that conform to this distribution, so as to realize the prediction of user behavior data and preference data, and generate synthetic data on the basis of user data to assist in extracting the semantic requirements of users; (d) Combine the synthetic data with the user data and input it into the LLM trained by the local database for data inference, that is, directly input the user data and synthetic data into the LLM to realize semantic inference and obtain the semantic requirement representation. The LLM not only extracts the semantics of the data but also expands and compresses the semantics according to the context.

[0078] (2) Mapping of requirements and resources.

[0079] The mapping of requirements and resources is part of the functions completed by the mapping adaptation layer, including: (a) The mapping of requirements and resources is completed by a dual-knowledge base architecture consisting of a service requirement knowledge base and a network resource knowledge base. Both knowledge bases define effective search spaces through LLM and specify accurate mapping paths; (b) The service requirement knowledge base uses LLM to accurately map semantic business requirements (i.e., semantic requirement representations) to specific QoS indicators, such as the quantitative representation of bandwidth requirements, the numerical mapping of latency constraints, the hierarchical definition of service priorities, etc.; (c) The network resource knowledge base uses LLM to map the real-time network state to network performance indicators, such as key standardized resource state characterization indicators like link utilization and queue length.

[0080] As Figure 3 shown, the process of constructing the two databases in this embodiment is as follows: (a) Construction of the service requirement knowledge base: The service requirements are initially standardized into four categories of indicators: basic performance indicators (bandwidth, latency, service priority), reliability indicators (packet loss rate, jitter), resource indicators (computing resources, network resources), and service feature indicators (packet size, traffic pattern, service duration); (b) Construction of the network resource knowledge base: The resource states are initially standardized into four categories of indicators: link state indicators (available transmission bandwidth, link utilization), cache resource indicators (queue length, cache utilization), timing performance indicators (current link latency, jitter range), and device capability indicators (port state); (c) Fine-tuning and updating of the knowledge base: Based on the defined initial standardized indicators, fine-tuning is carried out in combination with different demand scenarios (such as high-real-time services, large-bandwidth services) and network types (edge network, core network). The data used for fine-tuning is mainly historical data and real-time feedback data, and the indicator content of the knowledge base is continuously updated accordingly. Fine-tuning means training the LLM in the knowledge base with some specific data on the basis of the initial standardized indicators to expand the indicators of the knowledge base. For the service requirement knowledge base, the specific data refers to tasks with different requirements. For the network resource knowledge base, the specific data refers to different networks, such as TSN network, DetNet network. The fine-tuning principle is the same under different scenarios, but only the data used is different; (d) Accurate mapping of requirements and resources: Based on the fine-tuned knowledge base, map the semantic requirement representations to quantified QoS indicators and map the real-time network state to standardized network performance indicators to achieve the accurate mapping of requirements and resources. It should be noted that after the knowledge base is established, the mapping is like looking up a dictionary. For example, if the user requests smoothness, the knowledge base directly gives the accurate requirement of a latency of 50ms.

[0081] (3) End-to-end deterministic transmission scheduling algorithm based on diffusion model enhanced D3QN.

[0082] The end-to-end deterministic transmission scheduling algorithm based on diffusion model enhanced D3QN, which is part of the functions completed by the mapping adaptation layer, is used to generate transmission scheduling strategies, including: (a) The data abstraction module receives QoS metrics and network performance metrics, and standardizes them into a specific data structure format to obtain standardized data; (b) Input the standardized data into the diffusion agent to integrate the QoS metrics and network performance metrics into a state vector; (c) The diffusion agent regards the current network resource state as noise, performs iterative denoising according to the QoS metrics, and generates an action vector (i.e., a deterministic transmission scheduling strategy); (d) The data abstraction module configures the transmission scheduling strategy to the data plane, completes the deterministic transmission scheduling of data packets, and obtains corresponding rewards according to latency, jitter, and throughput.

[0083] As Figure 4 shown, based on QoS metrics and network performance metrics, the training process of the end-to-end deterministic transmission scheduling algorithm using diffusion model enhanced D3QN specifically includes the following steps.

[0084] (1) Data structure construction: The data abstraction module receives QoS metrics and network performance metrics, and standardizes them into a specific data structure format, i.e., the nth state vector o n , to prepare for generating the state vector set S.

[0085] (2) State vector integration: The diffusion agent integrates the state vectors, and defines the state vector set as S = {o1, o2,..., o N}, where o n is the nth state vector, and N is the number of state vectors. o n = (G, L, Q, F, T), G is the network topology, L is the link load, Q is the queue state, F is the flow characteristic, T is the latency requirement, G is represented as a graph structure, L and T can be represented by a single digit respectively, Q and F are in matrix form, G,, Q belong to network performance metrics, F, T belong to service quality metrics, the queue state matrix represents the state (i.e., length and priority) of the jth queue of node i, I is the number of nodes, J is the number of queues, the queue state matrix consists of two parameters in total: length and priority. For example, if the current length of queue 2 of node 1 is 80 and the priority is 2, then the matrix element q 1，2 = (80, 2), the flow characteristic matrix f k is the flow characteristic of the kth flow, K is the number of flows, f k = (s k , d k , qos k ) represents the source node s k , the destination node d kand the QoS requirement qos of the k-th flow k 。

[0086] (3) Scheduling policy generation: The diffusion agent generates an action vector according to each state vector. Specifically, for a state vector, the current network resource state is regarded as noise, and iterative denoising is performed according to the service demand to generate a configuration action, that is, an action vector is generated. The action space is defined as A = {a1, a2,..., a M}, where a m is the m-th action vector and M is the number of action vectors. a m = (R, B, C tsn , C det ), is the routing decision matrix, represents the next-hop selection of flow k at node i, is the bandwidth allocation matrix, represents the bandwidth allocated to flow k on link c, and C is the number of links. is the configuration of the TSN domain, and GCL i represents the gating list of node i, is the configuration of the DetNet domain, ID i and SD i represent the idleslope (i.e., idle rate) and sendslope (i.e., sending rate) of node i, respectively.

[0087] The action vector is defined by the reinforcement learning algorithm. Given the input state vector, the algorithm learns the optimal policy and generates the action vector. The above matrices are the actions generated by the algorithm according to the state and are used to configure the core network and the edge network. Their role is to achieve deterministic transmission and forwarding of data packets, control the routing and forwarding order of data packets. That is, the action vector is the configuration for realizing deterministic transmission scheduling of data packets. For example, GCL is the gating list of TAS, which can control the opening and closing of each gate, and thus determine whether the data packet is transmitted. The transmission of data packets is controlled by various tables. The routing decision matrix controls the next node of the data packet, and the gating list GCL controls the transmission order of the data packet.

[0088] (4) Configuration deployment evaluation: The data abstraction module configures the scheduling decision to the data plane, completes the deterministic transmission scheduling of data packets, and obtains the corresponding reward according to delay, jitter, and throughput. The reward function is defined as R(s, a) = w i ·R delay + w2·R jit + w3·R thr - w4·C vio , w l , The weights assigned consider metrics such as latency, jitter, and throughput. At the same time, constraint violation penalties are used to ensure meeting QoS metrics, R delay is the latency, R jit is the jitter, R thr is the throughput, C vio is the penalty for violating the quality of service metrics.

[0089] For the generated state vector and action vector, calculate the corresponding rewards. Use the state vector, action vector, and rewards to update the policy network to obtain a trained policy network. Subsequently, directly input the state vector to generate the action vector, that is, use the above-mentioned end-to-end deterministic transmission scheduling algorithm based on the diffusion model to enhance D3QN to generate a transmission scheduling policy, which acts on the core network and edge network in the network execution layer to achieve the scheduling transmission routing plan of data packets in the network.

[0090] This embodiment discloses a novel communication network architecture (abbreviated as GenNet) endogenous to generative AI and a hierarchical collaborative scheduling method. The novel communication network architecture endogenous to generative AI includes: an intent extraction layer, a mapping and adaptation layer, and a network execution layer. The intent extraction layer processes various multimodal data to achieve the extraction of user semantic demand representation. The mapping and adaptation layer generates a transmission scheduling policy through the mapping and adaptation of demands and resources. The network execution layer provides the required physical network devices and, according to the transmission scheduling policy of the mapping and adaptation layer, combines with the deterministic transmission mechanism to achieve the end-to-end deterministic transmission scheduling of data packets. The hierarchical collaborative scheduling method includes: semantic demand representation extraction, mapping of demands and resources, and an end-to-end deterministic transmission scheduling algorithm based on the diffusion model to enhance D3QN. The novel communication network architecture and hierarchical collaborative scheduling method provided in this embodiment improve the computing efficiency, ensure the transmission latency, effectively promote intelligent resource management, and support emerging intelligent computing services in 6G networks.

[0091] In the first aspect of this embodiment, a novel communication network architecture endogenous to generative AI for 6G networks is proposed using the idea of ubiquitous intelligence to support the low-latency requirements of 6G network intelligent computing services. This architecture includes three layers: an intent extraction layer, a mapping and adaptation layer, and a network execution layer. The layers cooperate with each other to promote intelligent resource management and support emerging intelligent computing services in 6G networks. In the second aspect, a hierarchical collaborative scheduling method is provided, including semantic demand representation extraction, mapping of demands and resources, and an end-to-end deterministic transmission scheduling algorithm based on the diffusion model to enhance D3QN, which supports more efficient intelligent computing service scheduling in 6G networks and has the following effects.

[0092] (1) The novel communication network architecture for 6G network generative AI endogeny provided in this embodiment enables each layer and function to have independent or collaborative deterministic guarantee capabilities. Through the demand transfer and information sharing among the three levels of the intention extraction layer, mapping adaptation layer, and network execution layer, the generative and generalization capabilities of generative AI are fully utilized, improving computational efficiency, ensuring transmission latency, promoting intelligent resource management, and supporting emerging intelligent computing services in 6G networks.

[0093] (2) The hierarchical collaborative scheduling method provided in this embodiment serves the three-layer network architecture mentioned in the novel communication network architecture. The mapping of requirements and resources solves the problem of precise matching between user semantic requirements and network resources. Through the dual knowledge base architecture and the deep semantic understanding ability of the LLM, dynamic adaptation from service requirements to network resources can be achieved, ensuring the efficiency of resource allocation and the reliability of service quality. The end-to-end deterministic transmission scheduling algorithm based on the diffusion model enhanced D3QN realizes intelligent scheduling decisions in a cross-domain network environment, ensuring the end-to-end deterministic transmission performance of traffic in heterogeneous networks.

[0094] Embodiment 2

[0095] This embodiment provides a hierarchical collaborative scheduling method for a communication network architecture with generative AI endogeny, which is applied to the communication network architecture with generative AI endogeny described in Embodiment 1, as Figure 5 shown. The hierarchical collaborative scheduling method for the communication network architecture with generative AI endogeny includes:

[0096] S1: The intention extraction layer uses a generative AI model to process multimodal data to obtain a semantic requirement representation of the user.

[0097] S2: The mapping adaptation layer uses a generative AI model to map the semantic requirement representation to service quality indicators, maps the real-time network status of heterogeneous networks in the network execution layer to network performance indicators, and generates a transmission scheduling policy based on the service quality indicators and network performance indicators.

[0098] S3: The network execution layer works based on the transmission scheduling policy to achieve end-to-end deterministic transmission scheduling of service data packets.

[0099] Among them, generating a transmission scheduling policy based on service quality indicators and network performance indicators specifically includes: standardizing the service quality indicators and network performance indicators to obtain standardized data, where the standardized data includes network topology, link load, queue status, flow characteristics, and latency requirements; using the standardized data as a state vector, generating an action vector based on the state vector, and using the action vector as the transmission scheduling policy.

[0100] Among them, generating an action vector based on the state vector specifically includes: using the state vector as input and determining the action vector by means of the trained policy network. The reward function used when training the trained policy network is:

[0101] R(s, a) = w1·R delay + w2·R jit + w3·R thr - w4·C vio ;

[0102] Among them, R(s, a) is the reward; w1 is the first weight; R delay is the latency; w2 is the second weight; R jit is the jitter; w3 is the third weight; R thr is the throughput; w4 is the fourth weight; C vip is the penalty for violating the quality of service index.

[0103] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0104] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0105] Specific examples are used in this article to elaborate on the principle and implementation manner of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, according to the idea of this application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A communication network architecture endogenous to generative AI, characterized in that, The communication network architecture endogenous to the generative AI includes: An intent extraction layer for processing multimodal data using a generative AI model to obtain a semantic requirement representation of the user; the multimodal data is the requirement data of the user during the business execution of intelligent computing services in the 6G network; A mapping adaptation layer connected to the intent extraction layer and the network execution layer respectively, for mapping the semantic requirement representation to service quality indicators using a generative AI model, mapping the real-time network state of heterogeneous networks in the network execution layer to network performance indicators, and generating a transmission scheduling strategy based on the service quality indicators and network performance indicators; A network execution layer for working based on the transmission scheduling strategy to achieve end-to-end deterministic transmission scheduling of business data packets; The intent extraction layer includes: A multimodal processing module for extracting features from multimodal data to obtain user data; A generation processing module for reasoning about user behavior and preferences based on user data and predicting and generating synthetic data; A semantic requirement processing module for extracting, expanding, and compressing semantics by combining user data and synthetic data to obtain a semantic requirement representation of the user; Among them, both the generation processing module and the semantic requirement processing module use generative AI models; The mapping adaptation layer includes: A service requirement knowledge base for mapping the semantic requirement representation to service quality indicators; A network resource knowledge base for mapping the real-time network state of heterogeneous networks in the network execution layer to network performance indicators; An adaptation decision module for dynamically matching requirements and resources based on service quality indicators and network performance indicators to generate a transmission scheduling strategy; Among them, generative AI models are deployed in both the service requirement knowledge base and the network resource knowledge base, and the adaptation decision module uses an end-to-end deterministic transmission scheduling algorithm based on the diffusion model enhanced D3QN to generate a transmission scheduling strategy.

2. The communication network architecture endogenous to generative AI according to claim 1, characterized in that, The multimodal data includes text, audio, picture, and location information. At this time, the multimodal processing module includes a CNN model, a ViT model, and a GNN model. The CNN model is used to extract features from text to obtain a first feature, the CNN model is used to extract features from audio to obtain a second feature, the ViT model is used to extract features from pictures to obtain a third feature, and the GNN model is used to extract features from location information to obtain a fourth feature. The first feature, the second feature, the third feature, and the fourth feature are combined to form user data.

3. The communication network architecture endogenous to generative AI according to claim 1, wherein The service quality indicators include: basic performance indicators, reliability indicators, resource indicators, and service feature indicators. The basic performance indicators include bandwidth, latency, and service priority. The reliability indicators include packet loss rate and jitter. The resource indicators include computing resources and network resources. The service feature indicators include packet size, traffic pattern, and service duration; The network performance indicators include: link state indicators, cache resource indicators, timing performance indicators, and device capability indicators. The link state indicators include available transmission bandwidth and link utilization rate. The cache resource indicators include queue length and cache utilization rate. The timing performance indicators include current link latency and jitter range. The device capability indicators include port status; The transmission scheduling strategy includes: a routing decision matrix, a bandwidth allocation matrix, and the configuration of a heterogeneous network. The routing decision matrix includes the next-hop selection of each flow at each node. The bandwidth allocation matrix includes the bandwidth allocated to each flow on each link. The nodes are network nodes in the heterogeneous network.

4. The communication network architecture endogenous to generative AI according to claim 3, wherein The network execution layer includes a heterogeneous network, which includes a first edge network, a core network, and a second edge network connected in sequence. Both the first edge network and the second edge network adopt time-sensitive networks, and the core network adopts a deterministic network. Among them, time-aware shapers are deployed in the first edge network and the second edge network, and a time-aware shaper and a credit-based shaper are deployed in the core network.

5. The generative AI endogenous communication network architecture according to claim 4, wherein The network execution layer also includes: A terminal processing module, connected to the mapping adaptation layer, for obtaining the transmission scheduling strategy from the mapping adaptation layer in real time. A global processing module, connected to the terminal processing module, for storing the transmission scheduling strategy obtained by the terminal processing module in the long term and sending it to the heterogeneous network in the storage order.

6. The generative AI endogenous communication network architecture according to claim 4, wherein The configuration of the heterogeneous network includes the gating list of each node in the first edge network and the second edge network, as well as the gating list, idle rate, and transmission rate of each node in the core network.

7. A hierarchical cooperative scheduling method for a generative AI-native communication network architecture, applied to the generative AI-native communication network architecture described in any one of claims 1-6, characterized in that, The hierarchical cooperative scheduling method of the communication network architecture endogenous to the generative AI includes: The intent extraction layer uses the generative AI model to process multi-modal data to obtain the semantic requirement representation of the user. The mapping adaptation layer uses the generative AI model to map the semantic requirement representation to service quality indicators, maps the real-time network state of the heterogeneous network in the network execution layer to network performance indicators, and generates a transmission scheduling strategy based on the service quality indicators and network performance indicators. The network execution layer works based on the transmission scheduling strategy to achieve end-to-end deterministic transmission scheduling of service data packets.

8. The hierarchical collaborative scheduling method for a communication network architecture endogenous to generative AI according to claim 7, characterized in that, Generating a transmission scheduling strategy based on service quality indicators and network performance indicators specifically includes: Normalizing the service quality indicators and network performance indicators to obtain normalized data; the normalized data includes network topology, link load, queue status, flow characteristics, and delay requirements. Using the normalized data as the state vector, generating an action vector based on the state vector, and using the action vector as the transmission scheduling strategy.

9. The hierarchical cooperative scheduling method for a communication network architecture endogenous to generative AI according to claim 8, characterized in that, Generating an action vector based on the state vector specifically includes: using the state vector as the input and determining the action vector using the trained policy network. Among them, the reward function used when training the trained policy network is: ； wherein, is the reward; is the first weight; is the delay; is the second weight; is the jitter; is the third weight; is the throughput; is the fourth weight; is the penalty for violating the quality of service metrics.

Images