A large model and knowledge graph enabled intent-driven network design method
By combining large models and knowledge graphs, an intent knowledge graph and a network state knowledge graph are constructed, which solves the problems of limited application scenarios and difficulty in representing user intent in existing intent-driven networks, and realizes flexible expression of diverse user intents and efficient management of network situation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2024-03-29
- Publication Date
- 2026-06-23
Smart Images
Figure CN118228815B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intent-driven network technology, specifically relating to a design method for intent-driven networks enabled by large models and knowledge graphs. Background Technology
[0002] With the rapid development of network technology, traditional network management faces increasing challenges. These challenges mainly include the complexity of network configuration, the diversity of user needs, and the constantly changing network environment. Traditional network management often relies on human experience and static rules, which are inadequate when dealing with complex network environments and dynamically changing user needs. Therefore, providing flexible and diverse customized services to meet the personalized needs of different users has become a research hotspot for next-generation networks. To meet the needs of on-demand services and differentiated network requirements, the new network paradigm of Intent-Driven Network (IDN) has been proposed. IDN possesses the capabilities of deep intent mining, global network state awareness, and real-time network configuration optimization, characteristics that will help it be widely applied in future networks.
[0003] Current research on intent-driven networks largely focuses on specific intent types or simple scenarios, with relatively limited in-depth research on diverse intent types. Existing intent analysis methods based on named entity recognition are mainly limited to extracting simple phrases, lacking the ability to analyze the deeper meaning of intents. Furthermore, due to the diversity of intent sources and types, the fusion and integration of intents in different scenarios faces challenges. At the same time, the large amount of unstructured data generated by the underlying network is difficult to effectively transform into useful knowledge.
[0004] Existing technology 1: Users input intents through an interactive interface, and these intents must conform to a specific template. The user-input intents are processed using a Bidirectional Long Short-Term Memory Network (BiLSTM) combined with a Conditional Random Field (CRF) to extract relevant network performance parameters or configuration information. Furthermore, an intermediate representation is created using Nile language. After user confirmation, this intermediate representation is transformed into a specific policy through an end-to-end sequence model constructed using LSTM. However, this method is limited by its relatively narrow application scenarios, lack of in-depth analysis of user intents, and the requirement for user input to follow a specific template, which restricts its application scope and flexibility.
[0005] Existing technology two: Using a specified intent language to express network requirements. In this method, users need to use a specific intent language, conforming to a predetermined framework, to express their network needs. This approach allows users to directly configure and adjust the network through specific language commands. However, the main drawback of this method is that users must master multiple intent languages to operate different types of network devices. When faced with diverse and complex network devices, this not only increases the learning cost but may also lead to a decrease in management efficiency.
[0006] Existing technology 3: Relying on traditional artificial intelligence algorithms to convert intentions into policies, thereby enabling network configuration and operation. Typical examples include methods based on ABAC policy rules, K-means clustering algorithm, and minimum spanning tree algorithm. However, the above methods have low utilization rate for big data and low efficiency for large-scale network management.
[0007] In summary, the existing technology has the following problems:
[0008] (1) The application scenarios are limited, the user learning and usage costs are high, there is a lack of in-depth exploration of user intentions, and user input must follow a specific template, which limits its application scope and flexibility.
[0009] (2) There are many types of user intentions in the network, making it difficult to accurately represent heterogeneous and diverse intentions. In addition, a large amount of unstructured data in the network is difficult to effectively transform into useful knowledge. There is a lack of unified representation means for user intentions and network status, resulting in a semantic gap. Summary of the Invention
[0010] To overcome the shortcomings of the existing technologies, the present invention aims to provide an intent-driven network design method enabled by large models and knowledge graphs. This method integrates the application of large models and knowledge graphs in intent-driven networks, constructs an intent knowledge graph and a network state knowledge graph, and utilizes the deep intent understanding and reasoning capabilities of the large model to call relevant underlying APIs to achieve intent semantic mining and the adaptation of intent requirements with underlying resource capabilities. This bridges the semantic gap between user intent, policy management, and the underlying network, thereby improving network operating efficiency.
[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0012] An intent-driven network design method enabled by large models and knowledge graphs includes the following steps;
[0013] S101: The user inputs their intent through the front-end natural language interaction interface, and the user intent is preprocessed.
[0014] S102: Based on the few-shot learning method, the preprocessed user intent is classified, and the big-small model collaboration method is used to extract key information of user intent and construct an intent knowledge graph based on it;
[0015] S103: Construct a network state knowledge graph by obtaining network topology performance parameters such as underlying network bandwidth utilization, packet loss rate, latency, and link status through the controller.
[0016] S104: Based on the prompting project and the constructed intent knowledge graph and network state knowledge graph, the intent understanding and reasoning capabilities of the large model are used to call the application programming interface provided by the underlying layer to realize the mapping and distribution of user intent to the underlying network policy.
[0017] S105: The underlying network state is perceived in real time through the network state knowledge graph and fed back to the large model. The large model makes dynamic real-time adjustments based on the network state knowledge graph information.
[0018] In step S101, the natural language interaction interface consists of an intent example box, an intent input box, and a submit button. The user inputs the intent based on the intent example and clicks submit. After that, the user's intent is cleaned (noise information in the text is removed, such as HTML tags, non-printable characters, etc.), corrected, and segmented.
[0019] In step S102, the specific operation steps for obtaining the current user intent classification result by constructing correct classification examples as the historical dialogue of the large model based on the few-shot learning method need to be given. Based on the small-model collaboration method, information is first extracted through the large model, and then secondary information extraction is performed through the small model constructed by named entity recognition.
[0020] The intent classification process is as follows:
[0021] First, analyze the user's intent input through the natural language interaction interface and extract key feature information, including but not limited to semantic content, usage scenarios, and user history behavior;
[0022] Subsequently, using a large language model and a few-shot learning method, historical dialogue content is constructed through prompts and examples, and the user's intent type is inferred through the context learning capability of the large model.
[0023] In step S103, the network state knowledge graph construction first collects network parameter information, including network element devices, transmission protocols, topology, bandwidth, latency, packet loss rate, and jitter.
[0024] Then, based on the Flask service, an automated script was written in Python, and based on the script, the Py2neo library was used to perform CRUD operations on the Neo4j graph database to form network element devices, transmission protocols, topology, bandwidth utilization, latency, packet loss rate, jitter nodes, and the connection relationships between each node.
[0025] Specifically, S104 involves the process of constructing a network state knowledge graph by obtaining network topology performance parameters through the controller.
[0026] (1) Bandwidth information acquisition: At time t1, the controller sends a PortStatsReq message to the switch to obtain the number of received and transmitted bits of the switch port, which are rx1 and tx1, respectively. At time t2, the controller sends a PortStatsReq message to the switch to obtain the number of received and transmitted bits of the switch port, which are rx2 and tx2, respectively. The maximum bandwidth is bw. max The remaining bandwidth of the link is calculated using the following formula:
[0027]
[0028] (2) Packet loss rate information acquisition: At time point t1, a PortStatsReq message is sent to switches 1 and 2 to obtain the number of received and transmitted data packets on the ports of switches 1 and 2 at this time. For switch 1, the numbers are rxa1 and txa1 respectively, and for switch 2, the numbers are rxb1 and txb1 respectively. After one measurement period, at time point t2, a PortStatsReq message is sent to switches 1 and 2 to obtain the number of received and transmitted data packets on the ports of switches 1 and 2 at this time. For switch 1, the numbers are rxa2 and txa2 respectively, and for switch 2, the numbers are rxb2 and txb2 respectively. After collecting the above information, the packet loss rate is calculated using the following formula:
[0029]
[0030] (3) Latency Information Acquisition: First, the controller sends an LLDP (Link Layer Discovery Protocol) data packet containing a timestamp T1 to network device A. Then, the latency T from the controller to device A is determined using echo messages. A After receiving the LLDP packet, device A forwards it to the adjacent network device B, which then sends the packet back to the controller. At this point, the controller measures the delay T returned from device B. B And record the current time T2 of receiving the LLDP packet; finally, calculate T using the formula d =T2-T1-(T A +T B The link delay can be calculated.
[0031] (4) Topology acquisition: Based on the REST API provided by the controller, network topology change information is obtained. By accessing relevant interfaces within a unit of time, the connection information of the link within that period is obtained, such as the device's connection port number, identifier, and link status, to determine and analyze whether a new link or node has been established or a link has been disconnected.
[0032] By acquiring the above information within a unit of time and updating it synchronously in the network state knowledge graph, the construction and updating of the network state knowledge graph are completed.
[0033] The specific process in S104 is as follows:
[0034] By defining an external application interface that can be called by the large model, and using the application name, application function description, application input parameter type and description, and specifying the required parameters as the basis for the large model to decide whether to call the external application interface;
[0035] Then, based on its understanding of user intent and knowledge of the current network environment and configuration, the large model uses its reasoning ability acquired during training to implement different application programming interface calls.
[0036] In step S105, the specific process of dynamic real-time adjustment is as follows:
[0037] The large model combines a network state knowledge graph to perceive and analyze the state of the underlying network in real time. Combined with prompting engineering, information is fed back to the large model. The large model performs feature analysis based on the real-time feedback information and the intent knowledge graph. After training, the large model makes decisions based on the current network state. When it finds that policy adjustments are needed, it calls the relevant APIs to adjust and optimize the established network policies to ensure user needs are met.
[0038] The beneficial effects of this invention are:
[0039] This invention is adaptable to different application scenarios and user needs. It can not only process and understand diverse user intentions, but also accurately abstract and express these intentions without strict template restrictions.
[0040] To eliminate the semantic gap between user intent and network state, this invention requires the establishment of a unified representation method that can simultaneously represent user intent and network state, ensuring seamless integration and efficient collaboration between the two.
[0041] This invention integrates the application of large models and knowledge graphs in intent-driven networks, constructing an intent knowledge graph and a network state knowledge graph. Through the deep intent understanding and reasoning capabilities of the large model, it calls relevant underlying APIs to achieve intent semantic mining and the adaptation of intent requirements with underlying resource capabilities, thereby bridging the semantic gap between user intent, policy management, and the underlying network, and improving network operating efficiency. Attached Figure Description
[0042] Figure 1 This is a schematic diagram of an intent-driven network design method that enables large models and knowledge graphs, provided as an embodiment of the present invention.
[0043] Figure 2 This is a schematic diagram of an intelligent agent design based on a large model, provided for an embodiment of the present invention.
[0044] Figure 3 This is a schematic diagram of intent classification based on a large model, provided for an embodiment of the present invention.
[0045] Figure 4 This is a schematic diagram illustrating the construction of an intent knowledge graph provided in an embodiment of the present invention.
[0046] Figure 5 This is a schematic diagram illustrating the construction of a network state knowledge graph provided in an embodiment of the present invention.
[0047] Figure 6 This is a schematic diagram illustrating the dynamic adjustment of strategies by combining a large model with a knowledge base. Detailed Implementation
[0048] The present invention will now be described in further detail with reference to the accompanying drawings.
[0049] like Figure 1 As shown in the embodiment of the present invention, an intent-driven network design method enabling large models and knowledge graphs is proposed. The method includes the following steps:
[0050] S101: The user inputs their intent through the front-end natural language interaction interface, and the user's input intent is preprocessed.
[0051] S102: User intent is classified based on a few-shot learning method, and key information of user intent is extracted by a large-small model collaboration method and used to construct an intent knowledge graph.
[0052] S103: Construct a network state knowledge graph by acquiring network topology performance parameters such as underlying network bandwidth utilization, packet loss rate, latency, and link status through the controller;
[0053] S104: Based on prompting engineering and the constructed intent knowledge graph and network state knowledge graph, the intent understanding and reasoning capabilities of the large model are used to call the application programming interface provided by the underlying layer to realize the mapping and distribution of user intent to the underlying network policy.
[0054] S105: The underlying network state is perceived in real time through the network state knowledge graph and fed back to the large model. The large model is then adjusted in real time based on the network state knowledge graph information.
[0055] The specific implementation of the present invention will be described in more detail below with reference to a specific embodiment.
[0056] Example 1:
[0057] This invention proposes an intent-driven network design method enabled by large models and knowledge graphs.
[0058] This method first captures the user's intent through a front-end natural language interactive interface. The intent is then sent to the back-end for data preprocessing, intent classification, and key information extraction to construct a structured intent knowledge graph. Next, using the SDN controller's REST API, data is obtained from the underlying network switch ports to calculate key network parameters such as bandwidth utilization, packet loss rate, latency, and link status, thereby constructing a network state knowledge graph. Finally, combining prompting engineering techniques, the user intent knowledge graph and the network state knowledge graph are integrated through a large model, and the underlying APIs are called to achieve the mapping, verification, and distribution of user intent to underlying network policies.
[0059] like Figure 2 As shown, this is a schematic diagram of an agent design based on a large model provided in an embodiment of the present invention:
[0060] Centered on a large model, which mainly consists of memory, tool / action, and planning modules, this approach expands the application potential of large models in intent-driven networks.
[0061] Memory module: mainly stores short-term memory and long-term memory. Short-term memory is the context of the conversation with the user, and long-term memory refers to the constructed intent knowledge graph and network state knowledge graph.
[0062] Tools / Actions Module: Composed of various callable APIs for subsequent calls by the large model. Each functional module can perform routing calculations, performance monitoring, and configuration management on the underlying network. The APIs can be implemented based on traditional methods or by expert models applicable to certain domains.
[0063] Leveraging few-shot learning and thought chain techniques, the large model analyzes tasks and formulates plans by providing prompts. By calling the APIs of the tool / action modules, the large model receives execution results and reports the final outcome to the user.
[0064] like Figure 3 As shown in the figure, the structure diagram of classifying user intent using a few-shot learning method provided in this embodiment of the invention is as follows:
[0065] The intent classification process can be specifically described as follows:
[0066] First, the intent input by the user through the natural language interface is analyzed to extract key feature information, including but not limited to semantic content, usage scenarios, and user history. Then, using a large language model combined with few-shot learning methods, historical dialogue content is constructed through prompts and examples. The user intent type is inferred through the context learning capabilities of the large model.
[0067] like Figure 4 As shown in the figure, this invention provides a schematic diagram of the method for extracting key user intent information based on a size model collaboration method:
[0068] Based on the size model collaboration, key user intent information is extracted and transformed into an intent knowledge graph. The specific implementation process can be described as follows:
[0069] First, based on the pre-defined target of the user intent type, key information extraction is performed. The large model guides the dialogue based on the extracted information and pre-defined prompts to ensure the comprehensiveness and accuracy of the information. Subsequently, a small entity extraction model built based on entity recognition methods is used to identify and extract key entity information from the dialogue. Entity information may include specific network performance parameters, time, location, and other elements closely related to the user intent. Through this process, the system can continuously update and refine information during the dialogue until all pre-defined information is fully collected. This process ensures that the user intent is fully understood and accurately captured, providing a data foundation for subsequent intent knowledge graph construction and network strategy generation.
[0070] like Figure 5 As shown in the figure, this embodiment of the invention provides a schematic diagram of constructing a network state knowledge graph by obtaining network topology performance parameters through a controller:
[0071] Furthermore, the specific process of constructing a network state knowledge graph diagram by obtaining network topology performance parameters through the controller can be described as follows:
[0072] (1) Bandwidth information acquisition: At time t1, the controller sends a PortStatsReq message to the switch to obtain the number of received and transmitted bits of the switch port, which are rx1 and tx1, respectively. At time t2, the controller sends a PortStatsReq message to the switch to obtain the number of received and transmitted bits of the switch port, which are rx2 and tx2, respectively. The maximum bandwidth is bw. max The remaining bandwidth of the link is calculated using the following formula:
[0073]
[0074] (2) Packet loss rate information acquisition: At time point t1, a PortStatsReq message is sent to switches 1 and 2 to obtain the number of received and transmitted data packets on the ports of switches 1 and 2 at this time. For switch 1, the numbers are rxa1 and txa1 respectively, and for switch 2, the numbers are rxb1 and txb1 respectively. After one measurement period, at time point t2, a PortStatsReq message is sent to switches 1 and 2 to obtain the number of received and transmitted data packets on the ports of switches 1 and 2 at this time. For switch 1, the numbers are rxa2 and txa2 respectively, and for switch 2, the numbers are rxb2 and txb2 respectively. After collecting the above information, the packet loss rate is calculated using the following formula:
[0075]
[0076] (3) Latency Information Acquisition: First, the controller sends an LLDP (Link Layer Discovery Protocol) data packet containing a timestamp T1 to network device A. Then, the latency T from the controller to device A is determined using echo messages. A After receiving the LLDP packet, device A forwards it to the adjacent network device B, which then sends the packet back to the controller. At this point, the controller measures the delay T returned from device B. B And record the current time T2 when the LLDP packet is received. Finally, calculate T using the formula... d =T2-T1-(T A +T B The link delay can be obtained.
[0077] (4) Topology acquisition: Based on the REST API provided by the controller, network topology change information is obtained. By accessing relevant interfaces within a unit of time, the connection information of the link within that period is obtained, such as the connection port number, identifier number and link status of the device, to determine and analyze whether a new link or node has been established or a link has been disconnected.
[0078] By acquiring the above information within a unit of time and updating it synchronously in the network state knowledge graph, the construction and updating of the network state knowledge graph are completed.
[0079] like Figure 6 As shown, the flowchart for dynamic strategy adjustment using a large model combined with a knowledge base provided in this embodiment of the invention is as follows:
[0080] Furthermore, the specific process for dynamically adjusting the procedure is as follows:
[0081] The large model combines network state knowledge graph to perceive and analyze the state of the underlying network in real time. Combined with prompting engineering, information is fed back to the large model. The large model performs feature analysis based on the real-time feedback information and the intent knowledge graph. When it finds that policy adjustments are needed, it calls the relevant APIs to adjust and optimize the established network policies to ensure user needs are met.
[0082] This invention classifies user intent based on a few-shot learning method, extracts key information of user intent using a large-small model collaboration method, and constructs an intent knowledge graph based on this information.
[0083] A network state knowledge graph is constructed by acquiring network topology performance parameters such as bandwidth utilization, packet loss rate, latency, and link status from the underlying network through a Software Defined Network (SDN) controller.
[0084] Based on prompting engineering and the constructed intent knowledge graph and network state knowledge graph, the intent understanding and reasoning capabilities of the large model are used to call the application programming interface (API) provided by the underlying layer to realize the mapping, verification and distribution of user intent to the underlying network policy;
[0085] The underlying network state is perceived in real time by the network state knowledge graph and fed back to the large model, which then makes real-time adjustments based on the network state knowledge graph information.
Claims
1. A method for designing intent-driven networks enabled by large models and knowledge graphs, characterized in that, Includes the following steps; S101: The user inputs their intent through the front-end natural language interaction interface, and the user intent is preprocessed. S102: Based on the few-shot learning method, the preprocessed user intent is classified, and the big-small model collaboration method is used to extract key information of user intent and construct an intent knowledge graph based on it; S103: Construct a network state knowledge graph by obtaining network topology performance parameters such as underlying network bandwidth utilization, packet loss rate, latency, and link status through the controller. S104: Based on the prompt engineering development and optimization of prompt words and the constructed intent knowledge graph and network state knowledge graph, the intent understanding and reasoning capabilities of the large model are used to call the application programming interface provided by the underlying layer to realize the mapping and distribution of user intent to the underlying network policy. S105: Real-time perception of the underlying network state through the network state knowledge graph and feedback to the large model, which then dynamically adjusts itself based on the network state knowledge graph information. In S101, the natural language interaction interface consists of an intent example box, an intent input box, and a submit button. The user inputs the intent through the intent input box based on the intent example in the intent example box. After clicking the submit button to submit, the user's intent is preprocessed by text cleaning, text correction, and word segmentation. In step S102, the specific operation steps are required to obtain the current user intent classification result by constructing correct classification examples as the historical dialogue of the large model based on the few-shot learning method. Based on the small and large model collaboration method, information is first extracted through the large model, and then information is extracted again through the small model constructed by named entity recognition. The specific process for intent classification is as follows: First, analyze the user's intent input through the natural language interaction interface and extract key feature information, including semantic content, usage scenario, and user history behavior; Subsequently, using a large language model and a few-shot learning method, historical dialogue content is constructed through prompts and examples, and the user's intent type is inferred through the context learning capability of the large model.
2. The method for designing intent-driven networks enabled by large models and knowledge graphs according to claim 1, characterized in that, In step S103, the network state knowledge graph construction first collects network topology performance parameter information, including network element devices, transmission protocols, topology, bandwidth utilization, latency, packet loss rate, and jitter. Then, based on the Flask service, an automated script was written in Python, and based on the script, the Py2neo library was used to perform CRUD operations on the Neo4j graph database to form network element devices, transmission protocols, topology, bandwidth utilization, latency, packet loss rate, jitter nodes, and the connection relationships between each node.
3. The method for designing intent-driven networks enabled by large models and knowledge graphs according to claim 2, characterized in that, Specifically, S103 involves the process of acquiring network topology performance parameters through the controller to construct a network state knowledge graph, which is as follows: (1) Bandwidth information acquisition: In The timing controller sends a PortStatsReq message to the switch to obtain the number of received and transmitted bits for the switch port, respectively. and At any moment Send a PortStatsReq message to the switch to obtain the receive and transmit bit counts for the switch port, respectively. and Maximum bandwidth is The remaining bandwidth of the link is calculated using the following formula: (2) Packet loss rate information acquisition: at a given time point The PortStatsReq message is sent to both switch 1 and switch 2 to obtain the number of received and transmitted data packets on the ports of switch 1 and switch 2 at that time. Switch 1 is respectively... and Switch 2 is respectively and After one measurement cycle, at time point The PortStatsReq message is sent to both switch 1 and switch 2 to obtain the number of received and sent data packets on the ports of switch 1 and switch 2 at that time. The number of received and sent data packets on switch 1 is as follows: and Switch 2 is respectively and After collecting the above information, the packet loss rate is calculated using the following formula: (3) Latency information acquisition: First, the controller sends a time stamp to network device A. The LLDP data packets are then used to determine the latency from the controller to device A via echo messages. After receiving the LLDP data packet, device A forwards it to the adjacent network device B. Device B then sends the LLDP data packet back to the controller. At this point, the controller measures the latency returned from device B. And record the current time of receiving LLDP data packets. Finally, through the calculation formula The link delay is calculated. (4) Topology acquisition: Based on the REST API provided by the controller, network topology change information is obtained. The connection information of the link within a unit of time is obtained by accessing the relevant interface within a unit of time. Based on the connection port number, identifier number and link status information of the device, it is determined and analyzed whether a new link or a new node is established or a link is disconnected. By acquiring the above information within a unit of time and updating it synchronously in the network state knowledge graph, the construction and updating of the network state knowledge graph are completed.
4. The method for designing intent-driven networks enabled by large models and knowledge graphs according to claim 1, characterized in that, The specific process in S104 is as follows: By defining an external application interface that can be called by the large model, and using the application name, application function description, application input parameter type and description, and specifying the required parameters as the basis for the large model to decide whether to call the external application interface; Then, based on its understanding of user intent and knowledge of the current network environment and configuration, the large model uses its reasoning ability acquired during training to implement different application programming interface calls.
5. The method for designing intent-driven networks enabled by large models and knowledge graphs according to claim 1, characterized in that, In step S105, the specific process of dynamic real-time adjustment is as follows: The large model combines a network state knowledge graph to perceive and analyze the state of the underlying network in real time. Combined with prompting engineering, information is fed back to the large model. The large model performs feature analysis based on the real-time feedback information and the intent knowledge graph. After training, the large model makes decisions based on the current network state. When it finds that policy adjustments are needed, it calls the relevant APIs to adjust and optimize the established network policies to ensure user needs are met.