A task-driven distributed network intelligent management system and method
Through a task-driven distributed network intelligent management system, using deep reinforcement learning and graph neural networks, the automated configuration and fault detection of unmanned platform networks are realized, and the lack of differentiated task guarantee and self-healing capabilities in traditional network management mode is solved. The problem improves the efficiency and resilience of the network.
Patent Information
- Application Number
- CN202410968949.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-07-19
AI Technical Summary
The traditional communication network management model cannot meet the needs of unmanned platform networks for differentiated task performance guarantee and efficient intelligent control, and there are problems such as low manual configuration efficiency and insufficient fault self-healing ability.
Adopt a task-driven distributed network intelligent management system, and realizes automated network configuration and failure through task intention translation module, network strategy generation module, policy verification and negotiation module and network performance statistics module, combined with deep reinforcement learning and graph neural network. Detection and recovery to adapt to dynamically changing network environments.
It improves the automation, efficiency and user-friendliness of network management, reduces algorithm complexity, enhances the resilience and adaptability of the network, and ensures efficient and reliable communications of the unmanned platform network.
Smart Images

Figure CN118509325B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of distributed network management, and more specifically, to a task-driven distributed network intelligent management system and method. Background Art
[0002] In recent years, unmanned platforms have developed rapidly due to their broad application prospects and are widely used in military, mapping, navigation, communication relay and other fields. A single unmanned platform has limited load and relatively simple functions. In order to make up for the lack of capabilities of a single unmanned platform, unmanned nodes act in a cluster manner, interact to form a distributed self-organizing network, and collaborate to complete tasks. The distributed network composed of unmanned platforms is a fast-moving self-organizing network that does not rely on a centralized node and is built when needed. Compared with traditional centralized networks, distributed self-organizing networks have better task execution capabilities, better fault tolerance and robustness, and higher economic affordability.
[0003] As the requirements for unmanned platform network performance in tasks such as perception, navigation, and rescue become increasingly higher, it is necessary to ensure efficient and reliable communication of unmanned node collaboration. On the one hand, various unmanned units will generate a variety of task types through collaboration, including reconnaissance, detection, collaboration, command, situation, network management, and other tasks. Each task includes voice, text, image, video, data, and other business types. The quality of service (QoS) requirements of each task type and business attribute are also different. The communication network needs to have differentiated performance assurance capabilities to match different application scenarios. On the other hand, with the continuous emergence and improvement of advanced technologies such as information, communication, and artificial intelligence, network functions such as 5G cellular, high-speed Wi-Fi, vehicle-mounted self-organization, and aviation self-organization are becoming increasingly complex. Managing such a large communication network and meeting the needs of each participating unit is also a key part of the informatization of unmanned platforms.
[0004] Traditional communication networks use manual methods to configure network protocols and communication resources, which is slow and has a high error rate. At the same time, traditional network architecture has poor adaptability to new application scenarios and cannot meet the network informationization needs of unmanned platforms. Summary of the invention
[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a task-driven distributed network intelligent management system and method, so as to realize the automatic configuration of the network driven by user tasks in a dynamically changing network environment, monitor and accurately predict the network status in real time, and optimize and adjust the networking strategy in time to deal with possible network failures and performance degradation.
[0006] The object of the present invention is achieved through the following solutions:
[0007] A task-driven distributed network intelligent management system, comprising:
[0008] The task intent translation module is used to parse the task instructions. After obtaining the elements of the task instructions, it parses the task instructions into the requirements of network performance indicators through the template mapping mechanism and passes the results to the networking strategy generation module;
[0009] The networking strategy generation module is used to generate networking control strategies based on deep reinforcement learning according to the requirements of the received mission network performance indicators and network status information, and then generate configuration parameters based on the strategies to achieve hierarchical mission service quality assurance;
[0010] The policy verification and negotiation module is used to verify the configuration of the networking control policy, determine whether the current network can meet the network performance requirements of the user task, and adjust the policy configuration according to the judgment result. When the network resources are insufficient, multiple tasks are allowed to negotiate the service quality degradation, and the minimum transmission is achieved by reducing the network performance requirements of some tasks. After the final configuration is generated, it can be deployed on the corresponding network nodes to meet the user task requirements;
[0011] The network performance statistics module is used to monitor and count network performance in real time during network operation, discover network failures to complete resilient reconstruction, and ensure network operation.
[0012] Furthermore, the task instruction is a task instruction formed after the user describes the task instruction through an intention template, and the intention template includes five key elements: target object, business type, task action, user level and duration.
[0013] Furthermore, the task instructions are parsed into network performance indicator requirements through a template mapping mechanism, specifically including: evaluating task attributes from multiple dimensions to design translation templates suitable for different business flows to ensure that the requirements of important tasks can be met first and can be accurately mapped to represent network resource requirements.
[0014] Furthermore, the template mapping mechanism includes: using the numerical values "3", "2", "1" and "0" to represent the requirement levels of "high", "medium", "low" and "no requirement" of each network performance indicator respectively; then integrating the QoS characteristics of multiple dimensions to obtain numerical values from 0-9; and then dividing the final QoS requirement level based on the obtained results; wherein the multiple dimensions include task type, service type and user level.
[0015] Furthermore, the network control strategy generated based on deep reinforcement learning specifically adopts the Markov decision process to model the strategy selection problem: the network control node corrects the action by interacting with the network environment to obtain rewards, and the process is described as a triple<A,S,R> , where A represents the set of networking control strategy actions, S represents the state set of task intentions, and R represents the feedback of the network environment to the network control node after the network control node gives a specific networking strategy.
[0016] Furthermore, the generation of the networking control strategy based on deep reinforcement learning also includes designing a strategy selection algorithm to solve the Markov decision process, and the algorithm specifically includes:
[0017] Step (1), the memory slot inputs the current system state s and the network control selection strategy a into the main network, and Deep neural network obtains approximate action-value function predictions ;
[0018] Step (2), the memory slot stores the system state at the next moment , Network control selection strategy Input the target network and pass it with the second weight parameter Deep Neural Networks Get True Values ;
[0019] Step (3), construct the Multi-agent learning model loss function , specifically expressed as:
[0020] ;
[0021] is the size of the mth agent dataset, Represents a memory slot, represents the discount factor, represents the reward signal of the mth network control node, represents the true value of the mth network control node, represents the predicted value of the mth network control node;
[0022] Step (4), the main network is obtained by minimizing the loss function Updates; every once in a while, the main network will be updated Directly copy to the target network ; Define the optimal model parameters as the following global loss function minimization problem for network control nodes Update the agent model parameters:
[0023] ;
[0024] Among them, network control nodes The weight of the loss function ,in, is the size of all datasets, To express the meaning of any To make Take the minimum value , M represents the number of network control nodes, Represents the loss function.
[0025] Furthermore, the strategy-based generation of configuration parameters specifically includes: after determining that the networking control strategy is selected, the specific communication requirement parameters translated from the task intent are input into the networking strategy generation module as needed, and the networking strategy generation module generates a set of configuration parameters from the input communication requirement parameters according to the designed algorithm.
[0026] Furthermore, the determination of whether the current network can achieve the network performance requirements of the user task and adjusting the policy configuration according to the determination result specifically include: returning the configuration that does not meet the requirements to the networking policy generation module for modification, and sending the verified policy to each execution node.
[0027] Furthermore, in the process of policy verification, it is carried out specifically according to the network status information repository and the network status prediction model, and the traffic scheduling situation in the configuration is applied to the current network status information to simulate the network status after the configuration is issued. At this time, these data are passed into the prediction model to predict the network performance after the configuration is issued, and to determine whether the current configuration is reasonable; if it is unreasonable, the performance indicators that do not meet the requirements are marked, and the feedback is fed back to the policy generation module to modify the corresponding parameters and generate a new configuration; if it is reasonable, it will enter the subsequent processing.
[0028] Furthermore, the network status prediction model is established specifically according to the nodes, links and traffic in the network.
[0029] Furthermore, the network status prediction model is specifically implemented by establishing a graph neural network model using the cyclic dependency relationship between traffic, nodes and links.
[0030] Furthermore, when network resources are insufficient, multiple tasks are allowed to negotiate service quality degradation, and guaranteed transmission is achieved by reducing some network performance requirements of the tasks. Specifically, it also includes: checking the implementation status of the intentions affected in the process of generating configuration parameters, and determining whether these intentions are met. If not, marking the relevant links when these intentions are implemented, and then regenerating the configuration for the current intention. The link parameters used when generating the configuration will be modified according to the satisfaction of other intentions. If satisfied, subsequent processing will be performed.
[0031] A task-driven distributed network intelligent management method comprises the following steps:
[0032] S1, task intent translation: parse the task instruction, obtain the elements of the task instruction, parse the task instruction into the requirements of network performance indicators through template mapping, and pass the result to the networking strategy generation step;
[0033] S2, Networking strategy generation: Receive the requirements of task network performance indicators, combine network status information, generate networking control strategies based on deep reinforcement learning, and then generate configuration parameters based on the strategies;
[0034] S3, policy verification and negotiation: Verify the configuration of the networking control policy, determine whether the current network can meet the network performance requirements of the user task, and adjust the policy configuration according to the judgment result. When network resources are insufficient, multiple tasks are allowed to negotiate service quality degradation, and the minimum transmission is achieved by reducing some network performance requirements of the tasks. After the final configuration is generated, it can be deployed on the corresponding network nodes;
[0035] S4, network performance statistics: During network operation, real-time monitoring and statistics of network performance are carried out to detect network failures to complete resilience reconstruction and ensure network operation.
[0036] The beneficial effects of the present invention include:
[0037] (1) Automation and efficiency. The present invention processes the generation of networking control strategies and the configuration of protocol parameters in an intelligent and fully automatic manner, which reduces the workload and error rate of manual operations and improves the efficiency of network management.
[0038] (2) User-friendliness. In the present invention, users only need to describe network tasks without having to understand specific operation instructions, making network management more user-friendly and easy to use.
[0039] (3) Low algorithm complexity. The optimization of networking control strategies based on mathematical models is usually a non-deterministic polynomial problem, which is usually solved by greedy methods, optimization theory or heuristic algorithms. However, as the scale of adjustable parameters continues to increase, the corresponding computational complexity and space complexity will increase exponentially. The solution implemented by the present invention using reinforcement learning can significantly reduce the algorithm complexity and is suitable for highly dynamic networking scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0041] Figure 1 It is a task-driven network intelligent control workflow of an embodiment of the present invention;
[0042] Figure 2 The task-driven “perception-decision-action-feedback” workflow of the embodiment of the present invention;
[0043] Figure 3 A task-driven distributed heterogeneous network management and control architecture according to an embodiment of the present invention;
[0044] Figure 4 Translate the workflow for the task intent of the embodiment of the present invention;
[0045] Figure 5 An evaluation framework for task attributes of an embodiment of the present invention;
[0046] Figure 6 A problem model for generating a strategy for an embodiment of the present invention;
[0047] Figure 7 The operating principle of the strategy generation based on deep reinforcement learning in the embodiment of the present invention;
[0048] Figure 8 Policy verification and negotiation for embodiments of the present invention;
[0049] Fig. 9 A network performance indicator prediction model according to an embodiment of the present invention;
[0050] Fig.10 The interaction between the service flows, nodes and links of the embodiment of the present invention;
[0051] Fig.11 The data transmission difference caused by the time slot number difference in the embodiment of the present invention;
[0052] Fig.12 Bandwidth measurement for an embodiment of the present invention;
[0053] Fig.13 Delay measurement for an embodiment of the present invention;
[0054] Fig.14 The figure is an illustration of packet loss rate measurement according to an embodiment of the present invention. DETAILED DESCRIPTION
[0055] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.
[0056] In view of the current situation in the background, the inventor of the present invention further believes after creative analysis that: in order to manage complex communication networks and meet the task requirements of different application scenarios, it is necessary to use task intent-driven network intelligent control technology. The communication network based on task-driven is highly intelligent. Through task-driven technology, the communication network can complete the translation and analysis of task intent without human intervention, and automatically complete communication resource configuration. The resource configuration efficiency is high and the error rate is low, which can better meet the task requirements of each unmanned platform node. The present invention mainly solves the following problems:
[0057] (1) Improve differentiated mission performance assurance capabilities. Communication networks carry multiple mission types such as environmental perception, disaster relief, and node coordination. Each mission can be divided into text, voice, image, video, data, and other information. Each service has different requirements for QoS indicators such as link transmission bandwidth, latency, jitter, and packet loss rate, and the importance of different services varies according to application requirements. Traditional chimney-style communication networks are dedicated to specific networks and cannot provide differentiated performance assurance services. Differentiated QoS control can enable the network to better adapt to different types of application requirements and service levels. For example, in real-time video transmission, low latency, high bandwidth, and low packet loss rate are required, while when transmitting text files, some bandwidth and reliability can be appropriately sacrificed in exchange for higher transmission speeds. In this case, differentiated QoS control can adaptively adjust network resource allocation according to the needs of different applications, thereby maximizing the satisfaction of application needs.
[0058] (2) Improve network intelligent control capabilities. The traditional network management model is that experienced engineers configure each network device to make the network achieve the intended state. This process may be carried out through various methods such as the command line interface (CLI) and automated scripts. However, in this management model, the efficiency of manual network configuration is low, and most network failures are caused by unintentional human errors during network configuration.
[0059] (3) Improve network resilience and anti-destruction capabilities. In a harsh competitive environment, compared with other types of networks, communication networks are more likely to have node failures and network anomalies, which places higher demands on the self-healing and autonomous capabilities of communication networks. Mission-intent-driven communication networks can collect network information in real time and continuously verify whether the current resource allocation strategy can meet the communication needs of unmanned platform collaboration, and continuously adjust resource allocation strategies through corresponding maintenance mechanisms to respond to network changes, and meet the network needs of unmanned platforms as much as possible through policy negotiation mechanisms. Compared with traditional communication networks, mission-driven communication networks have a higher level of self-healing autonomy.
[0060] In order to solve the above problems, the present invention aims to provide a network intelligent management method with optimization tasks as the goal and agile control as the means to achieve networking management and control capabilities such as elastic adaptation, resilience and anti-destruction, and expansion compatibility. Further develop the network intelligent control technology driven by task intent, and realize automatic configuration, optimization and management of the network according to the task intent and QoS requirements in different application scenarios in an automated and autonomous manner. It can sense changes in the network, handle the task requirements of different unmanned units, and dynamically adjust according to the network status and resource conditions. This technology can improve the efficiency and reliability of network management and ensure the smooth operation of the communication network. It is a highly elastic distributed communication network with low cost and low complexity, providing solid support for network informatization.
[0061] In a further inventive concept, task-driven network control technology is a network management technology that aims to implement network configuration and management in an automated and efficient manner to meet the user's task requirements. The present invention combines task intent-driven technology with a communication network, so that any communication device on the network can be provided with automated and reliable communication services without human intervention, which can meet the needs of high efficiency and task orientation in information network scenarios. A task-driven communication network is a network communication system that connects various heterogeneous communication devices and completes efficient processing, exchange and distribution of information based on a task intent-driven optimization closed loop. Figure 1The task-driven network intelligent agile control workflow shown in the figure is based on the user's description of the task. Through functional modules such as task intent translation, networking strategy generation, strategy verification and negotiation, and network performance statistics, it automatically implements network strategy configuration and maintains the network status required by the task.
[0062] ① Task intention translation module: The user describes the task instruction through a template and passes the instruction to the task translation module for parsing. The task translation module obtains the five main elements of the task instruction (target object, business type, task action, user level, and duration), and then parses the task instruction into network performance indicator requirements through the template mapping mechanism, and passes the result to the policy generation module.
[0063] ② Networking strategy generation module: Based on the received task network performance indicator requirements and combined with network status information, it autonomously generates networking control strategies based on deep reinforcement learning to achieve hierarchical task QoS guarantees.
[0064] ③ Policy verification and negotiation module: The policy verification module verifies the configuration of the networking control strategy, determines whether the current network can meet the network performance requirements of the user task, and adjusts the policy configuration in a timely manner. At the same time, when network resources are insufficient, it allows multiple tasks to negotiate service quality degradation, and achieves guaranteed transmission by reducing some of the network performance requirements of the task. Finally, the controller module receives the final generated configuration and deploys these configurations on the corresponding network nodes to meet the user's task requirements.
[0065] ④ Network performance statistics module: During network operation, this module monitors and counts network performance in real time, promptly detects network failures and completes resilient reconstruction to ensure network operation.
[0066] According to the above workflow, the entire network will operate closely around the mission intent, that is, the mission intent can affect the specific behavior of the network and is the driver of the operation of the entire network. Figure 2 The main workflow of task-driven network management is demonstrated. During the entire process, users only need to input the task intention to obtain the network configuration that meets their needs. The network will also provide real-time feedback on information such as network topology and network status, so that resource configuration can be adjusted in real time, realizing the closed-loop management of "perception-decision-action-feedback".
[0067] The present invention provides a task-driven distributed network management and control architecture, including four types of nodes: heterogeneous network common nodes, gateway nodes, relay nodes, and management and control nodes. Figure 3In the distributed network management and control architecture shown, the gateway node is responsible for the management, control and message forwarding of each communication network, and the gateway nodes are managed by multiple control nodes. This ensures that the network will not fail due to control node failure, and adding subnets will not bring too much burden to the control nodes, thereby improving the network's anti-destruction and scalability.
[0068] When the scale of networked nodes is small, the distributed network architecture can evolve into a centralized management and control architecture in which a single management and control node manages multiple heterogeneous networks. It should be pointed out that the intelligent control method adopted in the present invention is also applicable to the centralized management and control architecture.
[0069] (1) Ordinary nodes in heterogeneous networks. First, for member nodes in the same heterogeneous network, since their technical systems are of the same type, they can communicate directly with each other. Secondly, for ordinary nodes in different heterogeneous networks, regardless of whether the source node and the destination node belong to the same type of link system, the interconnection between the two nodes needs to go through a gateway.
[0070] (2) Gateway nodes have both Layer 2 and Layer 3 communication capabilities, can connect heterogeneous network nodes to the inter-cluster network, and achieve long-distance transmission through Layer 3 communication. Gateway nodes can perform protocol conversion, data format conversion, routing management and other operations as needed, converting communications on heterogeneous networks with different technical systems into a unified protocol and format, so that data can flow smoothly between communication networks using different physical transmission media, protocols and technologies. At the same time, gateway nodes also need to collect and measure various types of network status information and summarize them on the control node.
[0071] (3) In the inter-cluster network, the relay node mainly undertakes the network routing function, has a high transmission rate, and can realize signal relay and amplification, thereby extending the coverage of the wireless network.
[0072] (4) Control nodes. Control nodes are mainly responsible for network management functions, completing tasks such as network status perception, translating intents and generating policy configurations, so as to allocate network resources to each node and configure corresponding forwarding decisions. This project places task intent translation and policy generation on the control node for processing. In this architecture, one task intent corresponds to one data stream, which means that when a user generates an intent request, the request only needs to be sent once to the control node. Once the control node receives the task intent request, it will process it and generate the corresponding configuration. The control node analyzes and makes decisions on the intent through various policy control modules and generates appropriate configuration information (such configuration information includes routing tables, forwarding rules, resource allocation, etc.). Then, the control node sends this configuration information to the nodes related to the intent. After receiving the configuration information, the relevant nodes forward it according to the configuration. This means that when a user sends business data, the node can forward it according to the configuration information received previously, without having to query the control node every time.
[0073] Since the scale of distributed networks is much larger than that of centralized networks, multiple gateways need to be selected as control nodes. Each control node manages its own area, and these control nodes need to be synchronized regularly to ensure that each control node has a global view of the network, so that the network can run efficiently while enhancing the network's anti-destruction capabilities.
[0074] In the task intention translation module, the user intention needs to be translated, parsed, and understood to obtain the network policy type corresponding to the intention. The workflow is as follows: Figure 4 First, the application program interface obtains keywords from the user's intent to understand the user's purpose, such as transmitting disaster relief information, querying network status, etc.; secondly, the task translation module converts the task terms that the user may use into corresponding network performance indicators, such as situation information generally refers to images or videos with high-speed requirements.
[0075] Among them, when the user inputs the intention, for the task intention instruction, the intention template contains five key elements: target object, business type, task action, user level, and duration. For the user, it is not necessary to clearly indicate the specific data of these five elements, but only to describe them through natural language, and the specific parsing is performed by the intention translation module. For example, the task instruction "Ground search and rescue personnel T requests the unmanned platform to go to the rescue area to conduct a collaborative target search from the rescue management center C". According to the template-based method, the data entered into the template are "Ground rescue personnel T to the rescue management center C", "Voice", "Request unmanned platform for collaborative target search", "Advanced", "Default". If the duration is not specified in the task instruction, the default value is used. Inputting user intentions in this way can ensure the stability of intention expression without bringing too much burden to users.
[0076] Among them, the network indicator mapping based on the template, in the task intent driven communication network, the task intent corresponds to the business flow one by one, so when designing the translation template, it is necessary to evaluate the task attributes from multiple dimensions, including task type, business type and user level. By comprehensively considering these factors, a translation template suitable for different business flows can be designed to ensure that the network can give priority to the needs of important tasks and can be accurately mapped to represent network resource requirements (i.e., requirements for latency, sensitivity to jitter and packet loss, and bandwidth requirements, etc.). Figure 5 It is an evaluation framework for task attributes.
[0077] (1) Task type
[0078] Due to the variety of tasks, there are also a variety of task types. Different task types have different requirements for data QoS. Task types are divided into three categories: environmental perception, disaster relief, and node collaboration. Figure 6 Shown are the QoS characteristics corresponding to different types of tasks.
[0079] (a) Environmental perception: This mainly involves collecting and analyzing information collected by various sensors to support decision-making and strategic planning. It has relatively low latency requirements, but also needs to provide timely data transmission and interaction, which is suitable for supporting the timely acquisition and processing of information in perception tasks.
[0080] (b) Disaster Rescue: In rescue missions, management nodes need to communicate and guide controlled nodes in real time. Such real-time interaction is sensitive to latency and jitter, and can provide high-quality audio and video transmission and strict latency requirements, making it very suitable for supporting real-time disaster rescue needs of command and control missions.
[0081] (c) Node collaboration: It is sensitive to jitter and has certain latency requirements. It can provide real-time image and sound data transmission and support participating nodes to search for disaster relief areas and identify targets. Therefore, it is suitable for supporting the real-time feedback and disaster relief needs of participating nodes.
[0082] Table 1 QoS characteristics of different types of tasks
[0083] Task Type Latency requirements Jitter requirements Packet loss rate requirement Environmental Perception middle high high Disaster Relief high middle high Node collaboration high high middle
[0084] (2) Business type
[0085] Service type is another important dimension. Different service types have different requirements for data QoS, mainly including four service types: video, image, voice and text. Their corresponding QoS characteristics are shown in Table 2.
[0086] Table 2 Qualitative characterization of QoS characteristics of different types of services
[0087]
[0088] For different service types, specific values can be given as parameters to input into different adaptive networking control strategies according to the QoS characteristics in Table 3, so as to ensure more accurate user intentions.
[0089] Table 3 Quantitative characterization of QoS characteristics of different types of services
[0090]
[0091] (3) User level.
[0092] The tasks generated by users of different levels are also a factor we need to consider. For communication networks, the service levels of different users are different. The tasks generated by users of different service levels in the present invention are divided into three levels: high-level tasks, medium-level tasks, and low-level tasks. Tasks of different levels have differences in the three QoS characteristics of priority, delay, and packet loss rate, as shown in Table 4.
[0093] Table 4 Service QoS characteristics of users at different levels
[0094]
[0095] When designing the translation template, multiple dimensions were considered, including task type, service type, and user level, to ensure that the network can give priority to meeting the needs of important services and can be accurately mapped to represent network resource requirements. Since these three dimensions have different QoS characteristics, when synthesizing the three mapping tables, we use the values "3", "2", "1", and "0" to represent the "high", "medium", "low", and "no requirements" of the required levels of each network performance indicator. Then, by combining the QoS characteristics of the three dimensions, we can get values from 0 to 9, and then divide the final QoS requirement level based on the data. Through the above template translation method, different task types (rescue, collaboration, environment) and service types (text, voice, image, video) can be parsed, and the network indicator requirement level in each case is determined according to the user level (high, medium, and low), as shown in Table 5. Subsequently, strategies will be selected and configurations will be generated based on these translation results.
[0096] Table 5 Task intent translation template standard table
[0097]
[0098] In the strategy generation module, one or more control strategies are required to ensure the realization of a task intention, and due to the different communication capabilities or task requirements of the nodes, different performance focuses are required to achieve different intentions (for example, the service delay within the subnet is sensitive, while the cross-network service needs to focus on other performance). In this way, different strategy combinations are required to achieve different task intentions. Then the problem of strategy selection must be solved first, that is, which strategies need to be selected to ensure the realization of the intention. The present invention uses deep reinforcement learning to realize the intelligent selection of networking strategies.
[0099] Among them, the strategy selection problem model, the present invention adopts the Markov decision process to model the strategy selection problem, such as Figure 6 The network control node corrects its actions by interacting with the network environment and obtaining rewards. This process can be described as a triple<A,S,R> , where A represents the set of network control strategy actions, S represents the state set of task intent, and R represents the feedback of the network environment to the control node after the network control node gives a specific networking strategy. For example, if the strategy generated by the control node meets the task intent requirements, the network environment gives a positive foul, otherwise it gives negative feedback. At time t, when the agent is in state s(t) (s(t)∈S), it takes action a(t) (a(t)∈A) according to the strategy, receives the reward r(t) (r(t)∈R) from the environment, and enters the next state s(t+1), and executes the above process repeatedly to complete the decision.
[0100] (1) System state
[0101] Assume that the network control node is at time The business requirement of receiving a task intent i is the current system status , expressed as ,in, represents the priority requirement of task intention i, represents the delay requirement of task intention i, represents the bandwidth requirement of task intention i, represents the packet loss rate requirement of task intention i, Indicates the delay jitter requirement of task intention i. Its value corresponds to 3, 2, and 1 according to high, medium, and low.
[0102] (2) Networking strategy (Action)
[0103] Use vector Represents the impact gain of networking strategy on each communication demand, where , , , Respectively represent the influence gain of networking strategies 1, 2, n, and N of a certain task intention i on the communication demand. Similarly, 3, 2, 1, and 0 are used to represent strong, medium, weak, and no influence of the gain. It should be pointed out that the specific strategy gain can be continuously adjusted according to the specific network environment, task requirements, and performance optimization goals. For the sake of ease of understanding, the present invention takes transmission control, topology control, routing control, access control, and resource control as examples to illustrate the problem of strategy selection. It is worth noting that the algorithm of the present invention can be well extended to other networking control strategies. Assuming that transmission control ( )、Topology Control( )、Route Control( )、Access Control( )、Resource Control( ) on the communication requirements of a task intention i can be expressed as a matrix , and because the communication requirement of a task intention i can be expressed as Then the strategy selection can be expressed as a vector ,in, = 0 means that task intention i does not adopt the transmission control strategy, =1 means adopting this strategy; = 0 means that the task intention i does not adopt the topology control strategy, =1 means adopting this strategy; =0 means that the task intention does not adopt the routing control strategy, =1 means adopting this strategy; =0 means the task intention does not adopt the access control strategy, =1 means adopting this strategy; =0 means the task intention does not adopt resource control strategy, =1 means adopting this strategy.
[0104] (3) Reward
[0105] Considering the degree to which each networking strategy consumes network resources, a cost coefficient vector w can be introduced to measure the degree to which each strategy consumes network resources. The cost coefficient vector w here reflects the degree to which different strategies consume network resources, where each element w n represents the cost coefficient of the nth strategy. Our optimization goal is to select a set of strategies that can minimize the total consumption of network resources while meeting the business communication needs. The immediate reward can be expressed as (x ≥ 0). Among them, w is the cost coefficient vector, x is the strategy selection vector, A is the impact matrix of strategy on communication demand, and y is the communication demand vector.
[0106] Among them, the strategy selection algorithm, Q-learning algorithm is usually used to solve the above Markov process, but the algorithm needs to discretize the state and behavior of the system, and store the Q value by initializing the two-dimensional table of state-action to approximate the real value function. With the increase of network scale and task intent, the Q algorithm is prone to fall into the dimensionality disaster and cannot quickly learn effective network adjustment strategies. The present invention adopts deep reinforcement learning to solve the above challenges.
[0107] The operating principle framework of strategy selection based on deep reinforcement learning is as follows Figure 7 As shown. The functions of this algorithm mainly include the main network, the target network and the memory slot. Among them, the memory slot ( ) can be expressed as , s is the current system state, a is the network control selection strategy, r is the reward signal, is the system state at the next moment ( Figure 7 TS is used in the memory slot, m represents the mth agent (network control node ), is the size of the mth agent dataset, and the specific operation process is as follows:
[0108] (1) The memory slot inputs the current system state s and the network control selection strategy a into the main network through the weight parameter The deep neural network obtains the approximate predicted value of the action value function ;
[0109] (2) The memory slot stores the system state at the next moment. , Network control selection strategy Input the target network and pass it with weight parameters Deep Neural Networks Get True Values ;
[0110] (3) Construct the Multi-agent (control node ) Learning model loss function, specifically expressed as:
[0111] ;
[0112] in, Represents a memory slot, represents the reward signal of the mth network control node, represents the discount factor, represents the true value of the mth network control node, Represents the predicted value of the mth network control node.
[0113] Step (4), the main network is obtained by minimizing the loss function Updates; every once in a while, the main network will be updated Directly copy to the target network ; Define the optimal model parameters as the following global loss function minimization problem for network control nodes Update the agent model parameters:
[0114] ;
[0115] Among them, network control nodes The weight of the loss function ,in, is the size of all datasets, To express the meaning of any To make Take the minimum value , M represents the number of network control nodes, Represents the loss function.
[0116] Configuration parameters are generated. After determining which networking control strategies are selected, the specific communication requirement parameters of the mission intent translation are input into the selected control strategy module as needed. Each control strategy module will generate a set of configuration parameters from the input communication requirement parameters according to its specific functions and algorithms. These configuration parameters may include transmission paths, routing tables, resource allocation schemes, etc., depending on the implementation of each strategy module.
[0117] The configuration generation process usually needs to be calculated and optimized based on the logic and algorithms within the policy module. The module may consider factors such as network topology, resource availability, task priority, performance requirements, etc., and generate the optimal configuration parameters based on these factors. Once the configuration generation is completed, each control policy module will output the configuration parameters it generated. These configuration parameters can be passed to the corresponding network nodes to achieve specific networking control.
[0118] In the policy verification and negotiation module, the user task obtains the corresponding network configuration through the two stages of task translation and policy generation, and enters the policy verification step to verify whether the configuration meets the task requirements. The configuration that does not meet the requirements is returned to the policy generation module for modification. When the network status cannot meet the task requirements, the service negotiation degradation mechanism between multiple tasks is used to appropriately reduce the task requirements, achieve minimum connectivity, and send the verified policy to each execution node. Figure 8 As shown in the figure, the configuration is transferred and modified between policy generation and policy verification until it meets the requirements and is sent to the management and control node to modify the network.
[0119] Among them, in the policy verification, the availability verification of current resources is mainly carried out based on the network status information repository and the network status prediction model. The network status prediction model can predict performance indicators such as latency and packet loss rate based on the current network status information. Therefore, in the policy verification module, such as Fig. 9 As shown in the figure, the traffic scheduling in the configuration is applied to the current network status information to simulate the network status after the configuration is issued. At this time, these data are passed into the prediction model to predict the network performance such as latency and packet loss rate after the configuration is issued. It can be judged whether the current configuration is reasonable and whether there are enough network resources to implement the configuration. If it is unreasonable, the performance indicators that do not meet the requirements are marked, fed back to the policy generation module, and the corresponding parameters are modified to generate a new configuration.
[0120] The network status prediction model will establish a prediction model based on the traffic, nodes, links and business traffic in the network. Starting from the source node, it passes through the intermediate nodes according to the routing configuration. Arrives at the destination node, and the link It exists between each connected node. is a set of business flows, is the total number of business flows, j is the business flow identification number, is a node set, is the total number of nodes, i is the node identification number, is the link set, is the total number of links, and k is the link identification number. The three elements of business flow, link, and node have their own characteristics, such as Fig.10 As shown, they will also interact with each other and ultimately affect business characteristics.
[0121] Business traffic has the characteristics of sending status, business rate and priority. The sending status is used to distinguish the existing business traffic and the new business traffic of the current network, because the model needs to predict the network status after the new business traffic is generated, and at this time there are other business traffic on the network, which can provide more data, such as delay, packet loss rate, etc., which can be combined with these data for more accurate prediction. The business rate is related to the available bandwidth and time slot resources of the node. The higher the business rate, the more node bandwidth, that is, node time slot resources, will be required for the business traffic. The priority is related to the priority queue in the node. High-priority services are sent according to the strict priority policy, and medium and low-priority services are sent according to the weighted polling policy. Services of different priorities will have different opportunities to obtain time slots when competing for time slots, which means that they spend different time passing through the node.
[0122] A node has available bandwidth, number of queue messages, and time slot characteristics. Available bandwidth is determined by the total bandwidth and used bandwidth of the node, where the total bandwidth is an inherent property of the node and can generally be regarded as a constant, while the used bandwidth is determined by all business flows passing through the node. These business flows share the bandwidth resources of the node and there is competition between them. When the available bandwidth is insufficient, some services cannot be sent normally, resulting in increased latency or packet loss rate. The number of queue messages reflects the degree of time slot competition for business messages. Messages passing through the same node will share the node's priority queue. Regardless of whether it is a high priority or medium or low priority queue, the messages in the same queue are arranged in chronological order. Therefore, when the number of messages in the queue increases, the queuing delay of the business message will also increase accordingly. The time slot includes the time slot number assigned to the node and the length of the complete time slot cycle. In a complete time slot cycle, the time slots are numbered incrementally starting from 0. The node assigned the time slot number sends data within the corresponding time. Therefore, if Fig.11 As shown, when a node with a larger time slot number sends data to a node with a smaller time slot number, the latter can only wait for the second round of time slot cycles to have a chance to send the data. On the contrary, the latter will have a chance to send data soon.
[0123] The link has the characteristics of time slot number difference, link traffic bandwidth, and packet loss rate. Fig.11The data transmission difference caused by the time slot number difference is shown. Therefore, the time slot number is an important node feature. However, it is not appropriate to encode the time slot number of the node directly by integer coding or one-hot coding, because the prediction model will be applied to different topological graphs, which may increase or decrease the time slot number. The use of integer numbering will cause the out-of-range number to be unable to be processed normally, and the use of one-hot coding will increase the amount of data. Therefore, the time slot number difference between the link sender and the receiver is used to reflect this feature, which can minimize the impact of unknown data. The link traffic bandwidth is determined by the service traffic. The node sends the service traffic through different links according to the route. At the same time, the links with the same receiver will also send traffic to this node, which will affect its available bandwidth status. The packet loss rate is also an important feature of the link. Due to the different locations of the link receivers, the signal interference in different directions may be different, resulting in different packet loss rates of different links. The packet loss rate of multiple links through which the service traffic passes determines the packet loss rate of the service traffic.
[0124] According to the above analysis, the graph neural network model is established using the cyclic dependency relationship between business traffic, nodes, and links, namely:
[0125] (1) The state of business traffic is affected by the nodes and links through which the traffic passes;
[0126] (2) The state of a node is affected by the traffic passing through it and the links with it as the receiver;
[0127] (3) The state of a link is affected by the traffic passing through it.
[0128] Table 6 NetPredict prediction model table
[0129]
[0130] The prediction algorithm based on graph neural network is shown in Table 6. The input of the algorithm is the data of business traffic, nodes and links, where the business traffic includes the current business traffic and new business traffic. First, the input data is processed by using a fully connected network and converted into a unified vector format, corresponding to rows 1 to 3, and then it is iterated in a loop. In each processing, for each business traffic, the GRU model is used to process the nodes and links it passes through in turn, and the state of the business traffic is aggregated and updated, corresponding to rows 5 to 11. Then each link is processed and the state of the business traffic passing through this link is aggregated and updated, corresponding to rows 12 to 15. Finally, each node is processed and the state of the current node, the state of all business traffic passing through it, and the state of all links with it as the receiver are aggregated and updated, corresponding to rows 16 to 19. The GRU model is also used to update the link and node states. After multiple iterations, a fully connected neural network is used. The business traffic status is processed and the final prediction result is obtained.
[0131] Among them, in the specific implementation of policy negotiation, when network resources are insufficient, multi-task negotiation service degradation is required to ensure the normal operation of the network. Service degradation refers to improving the availability and stability of the network through task service quality and functions when network resources are insufficient or failures occur. For example, the consumption of network resources can be reduced by lowering bandwidth requirements, reducing transmission rates, adjusting priorities, etc. This allows the operation of the network to remain unaffected while ensuring that the basic needs of users are met. This can be achieved in the following ways:
[0132] (1) Lowering bandwidth requirements: You can reduce bandwidth requirements by reducing the amount of data transmitted, lowering the resolution of video or audio, compressing file size, etc. This can reduce network load and allow network resources to be better allocated to other purposes.
[0133] (2) Reduce the sending rate: You can reduce the sending rate of data packets by increasing the sending interval and reducing the sending frequency of data packets to reduce network pressure and improve network availability.
[0134] (3) Priority adjustment: Adjust the service priorities of different tasks according to the importance and urgency of the tasks. Allocate resources to tasks with higher priorities to ensure that they can be met, while downgrading the services of tasks with lower priorities.
[0135] In addition, the implementation status of other deployed tasks in the network also needs to be considered. Since the generated configuration will process network traffic, it will cause changes in the network status. At this time, the implementation of other intentions will be affected, so the implementation status of these intentions needs to be checked. First, it is necessary to check whether these intentions are met, such as performance requirements such as latency and packet loss rate. If not, mark the relevant links when these intentions are implemented, and then regenerate the configuration for the current intention. The link parameters used when generating the configuration will be modified according to the satisfaction of other intentions.
[0136] In the network performance statistics module, as the types of services and their requirements become more complex, network strategies are more easily affected by changes in the communication network status, so more detailed data is needed to support network decisions. In order to better formulate network strategies, it is necessary to understand the traffic characteristics of the network to ensure successful implementation. We select the following important representative network indicators for measurement to evaluate the status of links in the network, including bandwidth, latency, latency jitter, and packet loss rate.
[0137] In the specific implementation of bandwidth measurement, bandwidth is used to indicate the ability of the network's communication lines to transmit data. Therefore, network bandwidth indicates the "maximum data rate" that can be passed from one point in the network to another point in a unit of time (generally 1 second). Bandwidth, as an important indicator for measuring network characteristics, is an important basis for formulating network communication development strategies. Fig.12 As shown, the measurement of the link bandwidth between node A and node B is described:
[0138] Node A at time Count the number of bytes received by its port and the number of bytes sent ; Then in time Count the number of bytes received by its port and the number of bytes sent . Then the bandwidth of the current link can be as follows:
[0139] ;
[0140] The remaining bandwidth can be used by the port at a given maximum bandwidth Subtract the current traffic bandwidth , the formula is as follows:
[0141] .
[0142] In the specific implementation of delay measurement, delay refers to the time required for data (a message or packet, or even a bit) to be transmitted from one end of the network (or link) to the other end. Network delay includes four parts: processing delay, queuing delay, transmission delay, and propagation delay. In practice, transmission delay and propagation delay are mainly considered. Our specific measurement method is as follows Fig.13 As shown: (1) Node A first sends a delay measurement (DM) query message to node B and timestamps it: , which is the time when the DM query message is sent from node A; (2) When the message arrives at node B, the following timestamp is recorded in the DM message: , which is the time when node B receives the DM query message; (3) Node B transmits the message back to node A and records the following timestamp in it: , which is the time when node B sends the DM response message; (4) When the message returns to node A, the timestamp is recorded in the DM response message: , that is, the time when the DM response message is received at node A.
[0143] Through the above operations, we can calculate:
[0144] One-way delay from node A to node B ;
[0145] One-way delay from node B to node A ;
[0146] Bidirectional channel delay between node A and node B ;
[0147] In order to make the measurement result more accurate, multiple measurements can be taken to obtain the average value.
[0148] In the specific implementation of delay jitter measurement, delay jitter refers to the maximum delay for the same link. With minimum delay The time difference between the two networks is called jitter. Jitter can be used to evaluate the stability of the network. The smaller the jitter, the more stable the network. The calculation formula is as follows:
[0149] .
[0150] In the specific implementation of packet loss rate measurement, the packet loss rate refers to the ratio of the number of lost data packets to the number of sent data packets. For example, if 100 data packets are sent and one is lost, the packet loss rate is 1%. Fig.14 Explanation of the measurement process:
[0151] For the link between node A and node B, count the number of packets sent from A to B within a period of time and the number of packets sent by B to A .
[0152] At the same time, the number of data packets received by A and B from each other during the statistical time period is , the packet loss rate is calculated as follows:
[0153] .
[0154] In summary, the present invention has the following advantages over the prior art:
[0155] (1) The present invention proposes a task intent translation scheme, which obtains the task-network requirement intent translation template by integrating the QoS characteristics of the three dimensions of business task type, business type and user level, and performing corresponding reduction processing. For users, they only need to input through natural language, and then the intent translation module first extracts keywords and then translates the template.
[0156] (2) The present invention proposes a low-complexity intelligent networking control solution. To address the problems of long static pre-planning time and poor template matching accuracy in traditional communication networks, deep reinforcement learning is used to solve the problem of autonomous strategy generation, and low-complexity agile and rapid strategy generation is achieved, providing precise matching of task requirements to networking strategies. This method is applicable to a distributed management and control architecture composed of heterogeneous networks. At the same time, when the scale of networking nodes is small and can be evolved into a centralized network architecture, the technical principles proposed by the present invention are also applicable to this network architecture.
[0157] (3) The present invention proposes a strategy verification and negotiation method, which uses a graph neural network prediction model to simulate the network status and predict performance indicators after the configuration is issued, judge whether the configuration is reasonable, judge whether the current network can meet the network performance requirements of the user task, and adjust the policy configuration in time. At the same time, when network resources are insufficient, multiple tasks are allowed to negotiate service quality degradation, and guaranteed transmission is achieved by reducing some network performance requirements of the tasks.
[0158] The units involved in the embodiments of the present invention may be implemented by software or hardware, and the units described may also be arranged in a processor. The names of these units do not, in some cases, limit the units themselves.
[0159] According to one aspect of an embodiment of the present invention, a computer program product or a computer program is provided, the computer program product or the computer program includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. A processor of a computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method provided in the above various optional implementations.
[0160] As another aspect, an embodiment of the present invention further provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the method described in the above embodiment.
Claims
1. A task-driven distributed network intelligent management system, characterized in that: It includes common nodes, gateway nodes, relay nodes and control nodes in heterogeneous networks. Common nodes in the same heterogeneous network communicate directly with each other, and common nodes in different heterogeneous networks are interconnected through gateways. The gateway node has both Layer 2 and Layer 3 communication capabilities, and is used to connect heterogeneous network nodes to the inter-cluster network and achieve long-distance transmission through Layer 3 communication. The gateway node is also used to perform protocol conversion, data format conversion, and routing management operations as needed, converting communications on heterogeneous networks with different technical systems into a unified protocol and format, so that data can flow smoothly between communication networks using different physical transmission media, protocols, and technologies. The gateway node is also used to collect and measure various types of network status information and summarize it on the control node. Relay nodes are used in inter-cluster networks to perform network routing functions and realize signal relay and amplification. The control node is used in the inter-cluster network to undertake network management functions, complete network status perception, translate intent and generate policy configuration tasks, so as to allocate network resources to each node and configure corresponding forwarding decisions; it also includes: The task intent translation module is used to parse the task instructions. After obtaining the elements of the task instructions, it parses the task instructions into the requirements of network performance indicators through the template mapping mechanism and passes the results to the networking strategy generation module; The networking strategy generation module is used to generate networking control strategies based on deep reinforcement learning according to the requirements of the received mission network performance indicators and network status information, and then generate configuration parameters based on the strategies to achieve hierarchical mission service quality assurance; The policy verification and negotiation module is used to verify the configuration of the networking control policy, determine whether the current network can meet the network performance requirements of the user task, and adjust the policy configuration according to the judgment result. When the network resources are insufficient, multiple tasks are allowed to negotiate the service quality degradation, and the minimum transmission is achieved by reducing the network performance requirements of some tasks. After the final configuration is generated, it can be deployed on the corresponding network nodes to meet the user task requirements; In the process of policy verification, the network status information repository and the network status prediction model are used to apply the traffic scheduling in the configuration to the current network status information to simulate the network status after the configuration is issued. At this time, these data are passed into the prediction model to predict the network performance after the configuration is issued and determine whether the current configuration is reasonable. If it is unreasonable, the performance indicators that do not meet the requirements are marked and fed back to the policy generation module to modify the corresponding parameters and generate a new configuration. If it is reasonable, it enters the subsequent processing; The network performance statistics module is used to monitor and count network performance in real time during network operation, detect network failures to complete resilience reconstruction, and ensure network operation; The network control strategy is generated based on deep reinforcement learning, and the Markov decision process is used to model the strategy selection problem: the network control node corrects the action by interacting with the network environment to obtain rewards, and the process is described as a triple<A,S,R> , where A represents the set of network control strategy actions, S represents the state set of task intentions, and R represents the feedback of the network environment to the control node after the network control node gives a specific networking strategy; The generation of the networking control strategy based on deep reinforcement learning also includes designing a strategy selection algorithm to solve the Markov decision process, and the algorithm specifically includes: Step (1), the memory slot inputs the current system state s and the network control selection strategy a into the main network, and Deep neural network obtains approximate action-value function predictions ; Step (2), the memory slot stores the system state at the next moment , Network control selection strategy Input the target network and pass it with the second weight parameter Deep Neural Networks Get True Values ; Step (3), construct the Multi-agent learning model loss function , specifically expressed as: ; is the size of the mth agent dataset, Represents a memory slot, represents the discount factor, represents the reward signal of the mth network control node, represents the true value of the mth network control node, represents the predicted value of the mth network control node; Step (4), the main network is obtained by minimizing the loss function Updates; every once in a while, the main network will be updated Directly copy to the target network ; Define the optimal model parameters as the following global loss function minimization problem for network control nodes Update the agent model parameters: ; Among them, network control nodes The weight of the loss function ,in, is the size of all datasets, To express the meaning of any To make Take the minimum value , M represents the number of network control nodes, Represents the loss function.
2. The task-driven distributed network intelligent management system according to claim 1, characterized in that: The task instruction is a task instruction formed after the user describes the task instruction through an intention template. The intention template includes five key elements: target object, business type, task action, user level and duration.
3. The task-driven distributed network intelligent management system according to claim 2, characterized in that: The process of parsing task instructions into network performance indicator requirements through a template mapping mechanism specifically includes: evaluating task attributes from multiple dimensions to design translation templates suitable for different business flows to ensure that the requirements of important tasks can be met first and can be accurately mapped to represent network resource requirements.
4. The task-driven distributed network intelligent management system according to any one of claims 1 or 3, characterized in that: The template mapping mechanism includes: using the values "3", "2", "1", and "0" to represent the requirement levels of "high", "medium", "low", and "no requirement" for each network performance indicator respectively; then integrating the QoS characteristics of multiple dimensions to obtain values from 0 to 9; and then dividing the final QoS requirement level based on the obtained results; wherein the multiple dimensions include task type, service type, and user level.
5. The task-driven distributed network intelligent management system according to claim 1, characterized in that: The strategy-based generation of configuration parameters specifically includes: after determining that a networking control strategy has been selected, the specific communication requirement parameters translated from the mission intent are input into a networking strategy generation module as needed, and the networking strategy generation module generates a set of configuration parameters from the input communication requirement parameters according to a designed algorithm.
6. The task-driven distributed network intelligent management system according to claim 1, characterized in that: The determining whether the current network can achieve the network performance requirements of the user task and adjusting the policy configuration according to the determination result specifically includes: returning the configuration that does not meet the requirements to the networking policy generation module for modification, and sending the verified policy to each execution node.
7. The task-driven distributed network intelligent management system according to claim 1, characterized in that: The network status prediction model is specifically established based on nodes, links and traffic in the network.
8. The task-driven distributed network intelligent management system according to claim 7, characterized in that: The network status prediction model is specifically implemented by establishing a graph neural network model using the cyclic dependency relationship between traffic, nodes and links.
9. The task-driven distributed network intelligent management system according to claim 1, characterized in that: When network resources are insufficient, multiple tasks are allowed to negotiate service quality degradation, and guaranteed transmission is achieved by reducing some network performance requirements of the tasks. Specifically, it also includes: checking the implementation status of the intentions affected in the process of generating configuration parameters, and determining whether these intentions are met. If not, marking the relevant links when these intentions are implemented, and then regenerating the configuration for the current intention. The link parameters used when generating the configuration will be modified according to the satisfaction of other intentions. If they are met, subsequent processing will be performed.
10. A task-driven distributed network intelligent management method, characterized in that: The task-driven distributed network intelligent management system according to claim 1 further comprises the following steps: S1, task intent translation: parse the task instruction, obtain the elements of the task instruction, parse the task instruction into the requirements of network performance indicators through template mapping, and pass the result to the networking strategy generation step; S2, Networking strategy generation: Receive the requirements of task network performance indicators, combine network status information, generate networking control strategies based on deep reinforcement learning, and then generate configuration parameters based on the strategies; S3, policy verification and negotiation: Verify the configuration of the networking control policy, determine whether the current network can meet the network performance requirements of the user task, and adjust the policy configuration according to the judgment result. When network resources are insufficient, multiple tasks are allowed to negotiate service quality degradation, and the minimum transmission is achieved by reducing some network performance requirements of the tasks. After the final configuration is generated, it can be deployed on the corresponding network nodes; In the process of policy verification, the network status information repository and the network status prediction model are used to apply the traffic scheduling in the configuration to the current network status information to simulate the network status after the configuration is issued. At this time, these data are passed into the prediction model to predict the network performance after the configuration is issued and determine whether the current configuration is reasonable. If it is unreasonable, the performance indicators that do not meet the requirements are marked and fed back to the policy generation module to modify the corresponding parameters and generate a new configuration. If it is reasonable, it enters the subsequent processing; S4, network performance statistics: During network operation, real-time monitoring and statistics of network performance are carried out to detect network failures to complete resilience reconstruction and ensure network operation.
