Method and system for enhancing large language model auxiliary network operation and maintenance capability

Abstract

A method and system for enhancing large language model auxiliary network operation and maintenance capability, the method comprising: collecting a training sample, the sample collection process comprising: inputting analog network data into a network simulator, generating a text attribute graph of an analog network and a corresponding natural language query, and collecting network behavior data according to the natural language query; a large language model is constructed and trained, and the working process is as follows: natural language query and equipment configuration texts are generated through a configuration encoder, text attribute graphs are obtained through a topology encoder, then the text attribute graphs are spliced into network representation, and then the network representation is mapped into a Q layer and a V layer of an attention mechanism of the large language model in a low-rank fine tuning mode; and the large language model receives natural language query input by the user and the text attribute graph of the current network for reasoning output. The invention relates to the field of information communication networks, can efficiently model multi-modal network data, enhances the understanding ability of a large model for the current network state, and effectively assists network operation and maintenance.

Description

A method and system for enhancing the network operation and maintenance capabilities of large language models. Technical Field

[0001] This invention relates to a method and system for enhancing the network operation and maintenance capabilities of large language models, and relates to the field of information and communication networks. Background Technology

[0002] Modern telecommunications networks are increasingly characterized by their distributed and complex nature. Network devices, such as routers and switches, generate multimodal data, including device configuration, network topology, and performance metrics. Network configuration data is text-based; network topology data is structured; and performance metrics are numerical. This data is the core basis for network operations and maintenance. Network administrators need to analyze this data in conjunction with specific business requirements to achieve key operational tasks such as route optimization, device management, and performance evaluation.

[0003] Network operations and maintenance currently face significant technical bottlenecks. Traditional manual analysis methods are time-consuming and prone to errors, and even minor configuration errors can cause the entire network to crash. Rule-based network tools have strict requirements for input formats, requiring operations and maintenance personnel to have in-depth professional knowledge, which leads to a decrease in operational efficiency.

[0004] Large Language Models (LLMs) offer a new technical approach to network operations and maintenance (O&M) thanks to their powerful natural language understanding and semantic reasoning capabilities. However, due to the deep semantic relationships between the multimodal data of network devices—for example, a change in the configuration of one device may affect the traffic forwarding of all devices on a certain path, resulting in an increase in end-to-end latency—understanding these cross-modal data relationships and achieving a holistic understanding of the current network state is crucial for effective and accurate network O&M.

[0005] Current technical solutions include patent application CN 202510893805.1 (application title: Intelligent Network Operation and Maintenance Method and System Based on Performance Analysis and Large Language Model, applicant: Beijing NetDeep Technology Co., Ltd., application date: 2025.06.30). This patent solution proposes a technical approach of applying a general large language model to the network operation and maintenance process. By fine-tuning the large language model to enable it to understand network-related professional terminology, it can automatically analyze and summarize text information such as system logs and performance monitoring data. This solution mainly focuses on the passive interpretation and simple question-and-answer of historical log data. Its process is usually as follows: receiving natural language queries from users about network status, and directly calling the large language model to generate text responses based on historical data statistics. However, this solution treats the large language model as an independent and closed text processing module, failing to deeply embed it into the core closed loop of network application performance analysis. It lacks the ability to perform real-time and in-depth analysis of performance indicators and struggles to effectively combine the semantic understanding of the large model with the actual indicator detection, data association, and root cause reasoning capabilities of the performance analysis platform.

[0006] Therefore, how to efficiently model multimodal network data to enhance the ability of large language models to understand the current network state, thereby effectively assisting network operation and maintenance and improving operation and maintenance efficiency, has become a key technical issue of concern for technical personnel. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method and system for enhancing the network operation and maintenance capabilities of large language models, which can efficiently model multimodal network data to enhance the understanding of the current network state of large language models, thereby effectively assisting network operation and maintenance and improving operation and maintenance efficiency.

[0008] To achieve the above objectives, this invention provides a method for enhancing the network operation and maintenance capabilities of large language models, comprising:

[0009] Step 1: Collect multiple training samples to form a training sample set. The collection process for each training sample is as follows: Randomly construct simulated network data, then input the simulated network data into the network simulator, and collect device configuration data, network topology data, and communication link status data from the simulation data of the network simulator to generate a text attribute graph of the simulated network. At the same time, generate a natural language query based on the natural language requirement template, and finally collect the corresponding network behavior data from the simulation data generated by the network simulator based on the natural language query.

[0010] Step 2: Construct and train an augmented large language model, which uses natural language queries Q and text attribute graphs. The input is the natural language query Q, and the output is the answer corresponding to the natural language query Q. The workflow is as follows: The natural language query Q and the text attribute graph are compared. The configuration text of all devices is used to extract semantic features through a configuration encoder based on a self-attention mechanism to generate configuration-query association embeddings. At the same time, the text attribute map Global topological embedding is obtained by using a graph neural network-based topological encoder as the output. Then and splicing into a network representation Then, through low-rank fine-tuning, After being mapped to a trainable low-rank matrix, it is added to the Q and V layers of the attention mechanism of the large language model. Finally, the large language model inferences and outputs the answer content corresponding to the natural language query Q.

[0011] Step 3: Receive the natural language query input by the user. Simultaneously, collect device configuration data, network topology data, and communication link status data from the current network to generate a text attribute graph of the current network. Then, input the natural language query input by the user and the text attribute graph of the current network into the trained augmented big language model. Finally, the augmented big language model infers and outputs the answer content corresponding to the natural language query input by the user.

[0012] To achieve the above objectives, the present invention also provides a system for enhancing the network operation and maintenance capabilities of large language models, comprising:

[0013] The sample acquisition device collects multiple training samples to form a training sample set. The acquisition process for each training sample is as follows: randomly construct simulated network data, then input the simulated network data into the network simulator, and collect device configuration data, network topology data, and communication link status data from the simulation data of the network simulator to generate a text attribute graph of the simulated network. At the same time, generate a natural language query based on the natural language requirement template, and finally collect the corresponding network behavior data from the simulation data generated by the network simulator based on the natural language query.

[0014] A model building apparatus is used to build and train an augmented large language model, which uses natural language queries Q and text attribute graphs. The input is the natural language query Q, and the output is the answer corresponding to the natural language query Q. The workflow is as follows: The natural language query Q and the text attribute graph are compared. The configuration text of all devices is used to extract semantic features through a configuration encoder based on a self-attention mechanism to generate configuration-query association embeddings. At the same time, the text attribute map Global topological embedding is obtained by using a graph neural network-based topological encoder as the output. Then and splicing into a network representation Then, through low-rank fine-tuning, After being mapped to a trainable low-rank matrix, it is added to the Q and V layers of the attention mechanism of the large language model. Finally, the large language model inferences and outputs the answer content corresponding to the natural language query Q.

[0015] The model application device receives natural language queries input by the user. Simultaneously, it collects device configuration data, network topology data, and communication link status data from the current network to generate a text attribute graph of the current network. Then, it inputs the natural language query input by the user and the text attribute graph of the current network into the trained augmented big language model. Finally, the augmented big language model infers and outputs the answer content corresponding to the natural language query input by the user.

[0016] To achieve the above objectives, the present invention also provides a computing device, comprising:

[0017] Memory and processor;

[0018] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for enhancing the network operation and maintenance capabilities of large language models are implemented.

[0019] To achieve the above objectives, the present invention also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the method for enhancing large language model-assisted network operation and maintenance capabilities.

[0020] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention maps network data (device configuration files, network topology, performance metrics) of different modalities to a unified low-dimensional embedding representation by designing corresponding encoders, forming a network representation, thereby enabling direct processing of network data. Furthermore, by combining the network representation with the large model inference process through low-rank fine-tuning, it provides network background information for large model inference, enabling end-to-end participation of the large model in network operation and maintenance processes. It provides direct insights based on the current network state for data from different modalities on the network side and for the natural language input needs of network operation and maintenance personnel. This method has been experimentally verified in basic network operation and maintenance tasks, specifically routing analysis, device consistency, and performance consistency tasks. After fine-tuning training, the large language model can serve as a core part of network operation and maintenance work, understanding the current network state and providing accurate and rapid answers as an expert. This invention proposes a collaborative architecture of configuration encoder and topology encoder, respectively targeting text in network operation and maintenance. This invention employs specialized encoding of configuration data, graph structure, numerical topology, and performance data to address the challenge of handling heterogeneity in multimodal network data, enabling accurate feature extraction from different types of network data. A lightweight integration scheme based on Low-Rank Adaptation (LoRA) is designed, introducing low-rank matrix updates into the Q and V layers of the large language model to achieve efficient alignment between multimodal feature embedding and the semantic space of the large language model. This avoids the high resource consumption of full-parameter fine-tuning, balancing adaptation effectiveness and computational efficiency. Furthermore, this invention proposes a self-supervised random mask training mechanism for network data features. Through two tasks—key field mask replacement and invalid placeholder replacement—the model is forced to learn the association between text configuration and graph structure topology, significantly improving the large language model's understanding and reasoning ability regarding network domain knowledge. Finally, this invention designs a model fine-tuning method adapted to network scenarios, achieving semantic space alignment of data associations. This makes the large language model the core for understanding network state and reasoning about network operation and maintenance issues, thereby improving the efficiency of network operation and maintenance. Attached Figure Description

[0021] Figure 1 is a flowchart illustrating an exemplary embodiment of the present invention of a method for enhancing the network operation and maintenance capabilities assisted by a large language model.

[0022] Figure 2 is a schematic diagram illustrating an application of the present invention to a small network in an exemplary embodiment.

[0023] Figure 3 is a schematic diagram of the structure of a system for enhancing the network operation and maintenance capabilities of a large language model, as illustrated in an exemplary embodiment of the present invention.

[0024] Figure 4 is a schematic diagram of the structure of a computer device shown in an exemplary embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] As shown in Figure 1, the method for enhancing the network operation and maintenance capabilities of large language models according to the present invention includes:

[0027] Step 1: Collect multiple training samples to form a training sample set. The collection process for each training sample is as follows: Randomly construct simulated network data, which includes all network devices in the network, the configuration text of each network device, and the physical connection relationships between network devices. Then, input the simulated network data into the network simulator and collect device configuration data, network topology data, and communication link status data from the simulation data of the network simulator to generate a text attribute graph of the simulated network. At the same time, generate a natural language query based on the natural language requirement template. Finally, collect the corresponding network behavior data from the simulation data generated by the network simulator based on the natural language query. That is, the input of each training sample is a natural language query and a text attribute graph of the simulated network, and the expected output is the network behavior data collected from the simulation data generated by the network simulator based on the natural language query.

[0028] Step 2: Construct and train an augmented large language model, which uses natural language queries Q and text attribute graphs. The input is the natural language query Q, and the output is the answer corresponding to the natural language query Q. The workflow is as follows: The natural language query Q and the text attribute graph are compared. The configuration text of all devices is used to extract semantic features through a configuration encoder based on a self-attention mechanism to generate configuration-query association embeddings. At the same time, the text attribute map Global topological embedding is obtained by using a graph neural network-based topological encoder as the output. Then and splicing into a network representation Then, through low-rank (LoRA) fine-tuning, After being mapped to a trainable low-rank matrix, it is added to the Q and V layers of the attention mechanism of the large language model. Finally, the large language model inferences and outputs the answer content corresponding to the natural language query Q.

[0029] Step 3: Receive the user's natural language query. Simultaneously, collect device configuration data, network topology data, and communication link status data from the current network to generate a text attribute graph of the current network. Then, input the user's natural language query and the text attribute graph of the current network into the trained augmented big language model. Finally, the augmented big language model infers and outputs the answer content corresponding to the user's natural language query. In this way, the augmented big language model outputs accurate operation and maintenance decision support information through semantic reasoning, including specific operation suggestions, anomaly cause analysis, performance evaluation results, etc., which can directly assist operation and maintenance personnel in making decisions or automatically execute relevant network configuration operations.

[0030] In step one, due to the lack of open-source data in the networking field, this invention can use network simulators such as OMNet++ to obtain corresponding network behavior data based on network configuration and other input data. Step one, which involves randomly constructing simulated network data and then inputting it into the network simulator, can further include:

[0031] The input data required for the OMNet++ network simulator, including network devices, device configurations, and physical connections between devices, is randomly generated. This can be achieved using the INET framework supported by OMNet++, defining parameters such as the number of ports, input / output port numbers, and routing table settings. In the code, a Router class is written as the basic routing device, a Switch class as the data forwarding device within the same LAN, and a Reflector class as network devices in different autonomous systems using the BGP protocol. Simultaneously, the selection range for OSPF and BGP protocol configuration items, such as OSPF protocol weights and BGP protocol local preference values, is set, and corresponding values ​​are generated using a random algorithm. To enhance the data's relevance to real-world application scenarios, the physical connections between devices can be obtained from real-world industrial network topologies obtained from Topologyzoo data, and then preprocessed into adjacency matrix data. 500 network data examples are generated for each network topology of different sizes.

[0032] Steps one and three involve collecting device configuration data, network topology data, and communication link status data to generate a text attribute graph of the network. This may further include:

[0033] Step A1: Collect device configuration data and generate a network device configuration data set. : V is the set of all network devices in the network. It is the configuration text for device v in V;

[0034] Device configuration data is text data based on network protocols (such as link-state routing protocols and border gateway protocols) and is used to record device operating parameters and setting rules.

[0035] Step A2: Collect status data such as link delay, transmission rate, and packet loss rate of all communication links between devices to generate a communication link status set. : , , , These are the link delay, transmission rate, and packet loss rate of the communication link between devices u and v in V, respectively.

[0036] Step A3, with All devices in the middle are nodes, Construct a weighted graph of the network topology, with all communication links as edges. ;

[0037] Step A4: Preprocess the collected multimodal data. , Integrate into a Text Attribute Graph (TAG) : .

[0038] Because it is necessary to combine the natural language queries from network maintenance personnel with the network abstract representation after preprocessing the multimodal data collected from the network. The data is transformed into a unified network representation. The input data is inherently modal: natural language requirements and device configurations are represented as textual information, network topology as structured information, and network performance metrics (latency, transmission rate, and packet loss rate) as numerical information. Therefore, this invention needs to encode the data of different modalities and map them to a unified low-dimensional embedding space, serving as the foundation for subsequent representation embedding and semantic alignment modules. To achieve this, this invention maps multimodal network data to a unified network representation space by configuring an encoder and a topology encoder. Specifically:

[0039] 1. A configuration encoder is used to process natural language queries and device configuration data in text form. Device configuration affects aspects such as packet forwarding plane and protocol deployment. Natural language queries clarify the network insights to be mined, and there is an inherent relationship between the two. Self-attention mechanisms have advantages in capturing long-distance dependencies and handling variable-length sequences. Therefore, the configuration encoder in this invention adopts this mechanism, and its core function is to extract semantic features from device configuration text and natural language queries.

[0040] In step two, the natural language query Q and the text attribute graph are combined. The configuration text of all devices is used to extract semantic features through a configuration encoder based on a self-attention mechanism to generate configuration-query association embeddings. It can further include:

[0041] By using a word segmenter, the natural language query Q and the text attribute graph are combined. The configuration text of all devices is converted into embedding vectors. Then, the natural language query Q and the embedding vector corresponding to the configuration text of each device are input into a configuration encoder based on a self-attention mechanism to capture long-distance dependencies within the text and the association between configuration data and query intent, thereby obtaining configuration-query association embeddings. : , This indicates the configuration of encoder operations. yes The transformed embedding vector, It is the embedding vector after Q transformation. These are parameters of a trainable self-attention mechanism.

[0042] The specific settings for configuring the encoder can be as follows: set the truncation length of the configured text, i.e., the seqlen parameter, to 1024, and the hidden layer dimension of the original large language model to 4096; set the number of self-attention layers to 8, and the hidden layer dimension to 768.

[0043] 2. A topology encoder is used to process multimodal network data using a graph neural network (GNN). Graph neural networks process topology data through message passing and node update mechanisms, generating node-level, node-pair-level, or global-level embedding representations as needed. The topology encoder outputs a global network embedding representation from a weighted graph of the network topology using a graph neural network, capturing comprehensive network structure and performance information. Using the message passing mechanism as input, each node aggregates the feature information and link performance indicators of its neighboring nodes to achieve the interaction and fusion of local features.

[0044] In step two, the text attribute map is... Global topological embedding is obtained by using a graph neural network-based topological encoder as the output. It can further include:

[0045] Text attribute map The input is a graph neural network-based topology encoder. Through the message passing mechanism of the graph neural network, each node aggregates the node attributes (i.e., device configuration text) and topological connectivity (i.e., communication link states) of its neighbors. The final output is a global topology embedding. : , This represents the topology encoder operation. These are the parameters of a trainable graph neural network.

[0046] The specific settings for the topology encoder can be as follows: the graph neural network can use the average pooling operation to obtain the representation of the overall topology structure, the relative random walk step number for implementing the message propagation mechanism is set to 8, and the hidden layer dimension is set to 768.

[0047] In generation and Subsequently, the low-dimensional embedding data needs to be embedded into the inference process of the large language model, so that the model can fully consider the background information of the current network during forward propagation to make accurate answers. This invention employs a low-rank tuning (LoRA) method to inject network representation information, mapping the network representation to a low-rank matrix before adding it to the model's inference process. To achieve the above function, this invention first concatenates the low-dimensional embeddings output by different encoders into a unified low-dimensional embedding representation, and sets the mapping matrix to conform to the latent space dimension of the low-rank matrix, using it as network representation information. In step two, the low-dimensional embeddings are then... and splicing into a network representation It can further include:

[0048] Will and The network representation is generated by concatenating data across batch sizes. The expression can be: ,in, This represents a vector concatenation operation. It is a LoRA module mapping matrix with trainable parameters.

[0049] Low-rank fit fine-tuning is an efficient fit method that introduces a low-rank matrix. To approximate the parameter updates in the pre-trained large language model matrix, i.e. .in, and The model uses a low-rank training matrix, with a rank much smaller than the original hidden layer space of a large language model. During fine-tuning, other model parameters remain frozen, ensuring model scalability while minimizing complexity, allowing for seamless integration into large language models. This invention utilizes network representation... Mapping the low-dimensional embeddings of multimodal network data to the Q and V layers of the attention mechanism allows the large language model to incorporate network context information during inference. The core idea stems from two key insights: (1) Mapping the low-dimensional embeddings of multimodal network data to the Q layer of the attention mechanism can endow the model with context information based on the network scene, enabling the model to filter relevant information for different query inference needs during training; (2) Mapping the low-dimensional embeddings of multimodal network data to the V layer of the attention mechanism will affect the calculation of the final attention score, enabling the model to learn the importance weights of different network information queries during training. In step two, the low-dimensional embeddings of multimodal network data are added to the Q and V layers of the attention mechanism. After mapping to a trainable low-rank matrix, it is added to the Q and V layers of the attention mechanism in a large language model, and can further include:

[0050] Will After mapping to a low-rank matrix, it is added to the Q and V layers of the attention mechanism in the large language model, and its expression can be: , ,in, and These are the original weights of the Q and V layers of the large language model. and It is a trainable low-rank matrix. and These are the updated weights of the Q and V layers of the large language model.

[0051] The rank of the low-rank fine-tuning matrix in this part is set to 4, the alpha parameter is set to 8, and the hidden layer dimension is 4096, the same as the hidden layer dimension of the original model.

[0052] The above methods can complete the process of combining data input from different modalities with the structure of a large language model. However, a problem still needs to be solved: the semantic space of the input data is not aligned with the semantic space of the original large language model. That is, the large language model "does not understand" the specific meaning of the input data. Therefore, semantic space alignment is still needed. To achieve semantic alignment between device configuration, which plays a decisive role in the packet forwarding plane, and network topology, which plays a decisive role in the physical interconnection of devices, this invention can also perform self-supervised random mask training. This aligns the semantic space of the introduced encoder with that of the original large language model, enabling the large language model to have a unified understanding of different modalities of data (textualized device configuration, natural language queries, structured network topology, and numerical performance indicators). Unlike conventional self-supervised learning tasks that rely on contextual semantics, device configuration has structured semantic features that cannot be accurately inferred from contextual information alone. In step two, this invention uses a self-supervised random mask training method, which further includes:

[0053] The IP address and interface information in the configuration text of the device in the training sample are masked. Then, the large language model learns the meaning of the input network representation by using the correspondence between network configuration and network topology, thereby aligning the semantic space of the encoder and the large language model. The training further includes two types of masking tasks: (1) Key field masking replacement, which performs masking replacement of key network fields such as IP address and interface information in the configuration text of the device by a preset ratio (e.g., replacing 15% of the characters with "[mask]"). The original fields are used as the true values ​​to train the prediction ability of the large language model, thereby enhancing the semantic alignment between the model's learning of device traffic forwarding and the actual physical connection of the topology; (2) Invalid placeholder replacement, which performs invalid placeholder replacement of key network fields such as IP address and interface information in the configuration text of the device by a preset ratio (e.g., replacing 20% ​​of the characters with "[replace]" which has no actual meaning), in order to train the ability of the large language model to restore the original effective information. By introducing structured noise and semantic interference terms, the robustness and reasoning ability of the large language model in complex real network environments are improved.

[0054] This invention can also, based on self-supervised random mask training, save the checkpoints of the model parameters with the best evaluation results. By focusing on the relevant needs of network operation and maintenance scenarios, it can continue to fine-tune the model training for specific downstream tasks, enabling the model to have the ability to handle specific downstream tasks. This invention may also include:

[0055] Step B1: Set up three types of fine-tuning tasks: route analysis, device management and performance evaluation, and generate training sample sets for the three types of fine-tuning tasks for model training: In advance, use a large language model (e.g., GPT-4o model) to expand the natural language requirement templates. Generate multiple (e.g., 100) natural language requirement templates for each type of fine-tuning task, and fill them with random algorithms to diversify the possible intentions of network administrators, enhance the robustness and usability of the model, and share other input and output information for multiple natural language requirements with the same intention, and set corresponding fine-tuning task labels for the training samples;

[0056] The core considerations for designing the above three types of fine-tuning tasks in this invention are as follows: Covering core operational needs: Route analysis aims to optimize data transmission paths and ensure transmission reliability; device management focuses on device status maintenance and prevents downtime due to faults; performance evaluation is used to comprehensively assess network service quality. These three elements together constitute the core scenario of network operation and maintenance. Requires multi-modal data collaborative analysis: Route analysis needs to integrate traffic characteristics and network protocol information; device management needs to associate device configuration and performance indicator data; performance evaluation needs to integrate multi-dimensional network operating parameters, all of which rely on joint inference from multi-modal data. Covering multiple network layers: Route analysis focuses on the network forwarding layer, device management targets specific device nodes, and performance evaluation is oriented towards overall network performance indicators, achieving full-level operational and maintenance coverage from local to global perspectives.

[0057] Step B2: During each instruction fine-tuning model training, set the data ratio of each type of fine-tuning task (e.g., 1:1:1), so that the large language model will automatically perform random mixed training on the three types of fine-tuning tasks during instruction fine-tuning model training. The loss function is calculated as the average cross-entropy loss of the three types of fine-tuning tasks to balance the model's capabilities.

[0058] To more clearly explain the technical effects of the method of the present invention, the following is a specific experimental embodiment of the application of the present invention:

[0059] 1. Experimental Environment Setup

[0060] The Llama2 model was chosen as the base large language model for the experiment, and its 7B parameter version was used. Specifically, the hidden layer dimension of Llama2 was set to 4096. The number of self-attention layers in the configuration encoding part was set to 8, and the hidden layer dimension was also 4096. The hidden layer dimension of the graph neural network in the topological encoding part was set to 768, and the message passing layer was set to 8. The rank of the low-rank matrix in the LoRA module was set to 4, and the alpha was set to 8. In the random mask training phase with data alignment across different modalities, the number of training epochs was set to 2, the batch size to 4, and LoRA fine-tuning was used. In the fine-tuning phase of the downstream task, the number of training epochs was set to 2, the batch size to 4, and LoRA fine-tuning was also used. In both training phases, the ratio of the training set, validation set, and test set was 7:1:2. The model was optimized using the AdamW optimizer, with a learning rate of [missing value]. All experiments were conducted on dual A100 GPUs equipped with 80GB of video memory.

[0061] 2. Implementation process and results

[0062] 1) Collect device configuration text, topology data, and link performance metrics (latency, transmission rate, packet loss rate) for 500 different topologies, and convert them into a unified text attribute graph representation; 2) Encode the text data and graph structure-numerical data in the TAG using a configuration encoder and a topology encoder, respectively, to generate... and ;3) will and After integration, it is adapted to a large language model, and multimodal semantic alignment is achieved through self-supervised mask training; 4) Based on the labeled dataset, the three core operation and maintenance tasks are fine-tuned to complete the model training.

[0063] Figure 2 illustrates a schematic diagram of the present invention applied to a small network. In this small network, after autonomous network merging, the enhanced large language model achieves end-to-end network state understanding based on the intent of natural language. This topology contains multiple autonomous systems (AS1, AS2, AS3), interconnected via border routers. The operational intent is to ensure that all traffic destined for the network prefix 202.0.113.0 / 24 preferentially selects the transmission path via AS2.

[0064] 1. The test is initiated with a natural language query: "Verify that all border routers prioritize AS2 paths for network prefix 202.0.113.0 / 24?"

[0065] 2. First, import the device configuration and topology data, and then encode them into their respective low-dimensional embeddings. and ;

[0066] 3. By mapping the data to a lower-rank matrix using a mapping matrix, a unified network representation is achieved. Background information about the current network state is incorporated into the reasoning process of the large language model;

[0067] 4. Subsequently, using the forward propagation process of the large language model, joint reasoning is performed on the operational intent and network status. The final result shows that the model detected abnormal traffic between router 5 and router A that does not conform to the operational intent, and outputs the relevant results.

[0068] As shown in Figure 3, the system for enhancing the network operation and maintenance capabilities of large language models according to the present invention includes:

[0069] A sample acquisition device is used to collect multiple training samples to form a training sample set. The acquisition process for each training sample is as follows: Randomly construct simulated network data, which includes all network devices in the network, the configuration text of each network device, and the physical connection relationships between network devices. Then, input the simulated network data into a network simulator and collect device configuration data, network topology data, and communication link status data from the simulation data of the network simulator to generate a text attribute graph of the simulated network. At the same time, generate a natural language query based on a natural language requirement template. Finally, collect the corresponding network behavior data from the simulation data generated by the network simulator based on the natural language query. That is, the input of each training sample is a natural language query and a text attribute graph of the simulated network, and the expected output is the network behavior data collected from the simulation data generated by the network simulator based on the natural language query.

[0070] A model building apparatus is used to build and train an augmented large language model, which uses natural language queries Q and text attribute graphs. The input is the natural language query Q, and the output is the answer corresponding to the natural language query Q. The workflow is as follows: The natural language query Q and the text attribute graph are compared. The configuration text of all devices is used to extract semantic features through a configuration encoder based on a self-attention mechanism to generate configuration-query association embeddings. At the same time, the text attribute map Global topological embedding is obtained by using a graph neural network-based topological encoder as the output. Then and splicing into a network representation Then, through low-rank fine-tuning, After being mapped to a trainable low-rank matrix, it is added to the Q and V layers of the attention mechanism of the large language model. Finally, the large language model inferences and outputs the answer content corresponding to the natural language query Q.

[0071] The model application device is used to receive natural language queries input by users. At the same time, it collects device configuration data, network topology data, and communication link status data from the current network to generate a text attribute graph of the current network. Then, it inputs the natural language query input by users and the text attribute graph of the current network into the trained augmented big language model. Finally, the augmented big language model infers and outputs the answer content corresponding to the natural language query input by users.

[0072] Referring to Figure 4, which is a structural block diagram of a computing device 400 illustrating an exemplary embodiment of this specification, the components of the computing device 400 include, but are not limited to, a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and a database 450 is used to store data.

[0073] The computing device 400 also includes an access device 440, which enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 440 may include one or more of any type of wired or wireless network interface (e.g., Network Interface Card (NIC)), such as an IEEE 402.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0074] In one embodiment of this specification, the aforementioned components of the computing device 400, as well as other components not shown in FIG. 4, may be interconnected, for example, via a bus. It should be understood that the block diagram of the computing device shown in FIG. 4 is merely illustrative and not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0075] The computing device 400 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 400 can also be a mobile or stationary server or cloud server, etc.

[0076] The processor 420 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the method described above for enhancing the network operation and maintenance capabilities of large language models.

[0077] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the method for enhancing the network operation and maintenance capabilities of large language models described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the method for enhancing the network operation and maintenance capabilities of large language models described above.

[0078] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the method described above for enhancing the network operation and maintenance capabilities of large language models.

[0079] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the method for enhancing the network operation and maintenance capabilities of large language models described above. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the method or system for enhancing the network operation and maintenance capabilities of large language models described above.

[0080] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the method for enhancing the network operation and maintenance capabilities of large language models described above.

[0081] The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program and the technical solution of the method for enhancing the network operation and maintenance capabilities of large language models described above belong to the same concept. For details not described in detail in the technical solution of the computer program, please refer to the description of the technical solution of the method or system for enhancing the network operation and maintenance capabilities of large language models described above.

[0082] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0083] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0084] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for enhancing the network operation and maintenance capabilities assisted by large language models, characterized in that, The process includes: Step 1: Collecting multiple training samples to form a training sample set. The collection process for each training sample is as follows: Randomly construct simulated network data, then input the simulated network data into a network simulator, and collect device configuration data, network topology data, and communication link status data from the simulation data of the network simulator to generate a text attribute graph of the simulated network. At the same time, generate a natural language query based on the natural language requirement template, and finally collect the corresponding network behavior data from the simulation data generated by the network simulator based on the natural language query; Step 2: Constructing and training an enhanced large language model. The enhanced large language model uses the natural language query Q and the text attribute graph... The input is the natural language query Q, and the output is the answer corresponding to the natural language query Q. The workflow is as follows: The natural language query Q and the text attribute graph are used as input. The configuration text of all devices is used to extract semantic features through a configuration encoder based on a self-attention mechanism to generate configuration-query association embeddings. At the same time, the text attribute map Global topological embedding is obtained by using a graph neural network-based topological encoder as the output. Then and splicing into a network representation Then, through low-rank fine-tuning, After mapping to a trainable low-rank matrix, it is added to the Q and V layers of the attention mechanism of the large language model. Finally, the large language model infers and outputs the answer content corresponding to the natural language query Q. Step 3: Receive the natural language query input by the user. At the same time, collect device configuration data, network topology data and communication link status data from the current network to generate the text attribute graph of the current network. Then, input the natural language query input by the user and the text attribute graph of the current network into the trained augmented large language model. Finally, the augmented large language model infers and outputs the answer content corresponding to the natural language query input by the user.

2. The method according to claim 1, characterized in that, Steps one and three involve collecting device configuration data, network topology data, and communication link status data to generate a text attribute graph of the network. This further includes: Step A1, collecting device configuration data to generate a set of network device configuration data. ： V is the set of all network devices in the network. This is the configuration text of device v in V; Step A2: Collect link delay, transmission rate, and packet loss rate status data of all inter-device communication links to generate a communication link status set. ： ， 、 、 These are the link delay, transmission rate, and packet loss rate of the communication link between devices u and v in V; Step A3, with All devices in the middle are nodes, Construct a weighted graph of the network topology, with all communication links as edges. Step A4: Preprocess the collected multimodal data. 、 Integrate into a text attribute graph ： 。 3. The method according to claim 1, characterized in that, In step two, the natural language query Q and the text attribute graph are combined. The configuration text of all devices is used to extract semantic features through a configuration encoder based on a self-attention mechanism to generate configuration-query association embeddings. This further includes: using a tokenizer to segment the natural language query Q and the text attribute graph. The configuration text of all devices is converted into embedding vectors. Then, the natural language query Q and the embedding vector corresponding to the configuration text of each device are input into a configuration encoder based on a self-attention mechanism to obtain the configuration-query related embedding. ： ， This indicates the configuration of encoder operations. yes The transformed embedding vector, It is the embedding vector after Q transformation. These are parameters of a trainable self-attention mechanism.

4. The method according to claim 1, characterized in that, In step two, the text attribute map is... Global topological embedding is obtained by using a graph neural network-based topological encoder as the output. This further includes: text attribute graphs The input is a graph neural network-based topology encoder. Through the message passing mechanism of the graph neural network, the topology encoder enables each node to aggregate the node attributes and topological connections of its neighbors, ultimately outputting a global topology embedding. ： ， This represents the topology encoder operation. These are the parameters of a trainable graph neural network.

5. The method according to claim 1, characterized in that, In step two, and splicing into a network representation Further including: and The network representation is generated by concatenating the batch data. The expression is: ,in, This represents a vector concatenation operation. It is a LoRA module mapping matrix with trainable parameters.

6. The method according to claim 1, characterized in that, In step two, After mapping to a trainable low-rank matrix, it is added to the Q and V layers of the attention mechanism in a large language model, further including: After mapping to a low-rank matrix, it is added to the Q-layer and V-layer of the attention mechanism in the large language model, and its expression is: ， ,in, and These are the original weights of the Q and V layers of the large language model. and It is a trainable low-rank matrix. and These are the updated weights of the Q and V layers of the large language model.

7. The method according to claim 1, characterized in that, In step two, a self-supervised random masking training method is also adopted, including: masking the IP address and interface information in the configuration text of the device in the training sample, and then enabling the large language model to learn the meaning of the input network representation by using the correspondence between network configuration and network topology. The training further includes two types of masking tasks: (1) Key field masking replacement, which performs masking replacement of the IP address and interface information network fields in the device configuration text at a preset ratio, and uses the original fields as the true value to train the prediction ability of the large language model; (2) Invalid placeholder replacement, which performs invalid placeholder replacement of the IP address and interface information network fields in the device configuration text at a preset ratio, in order to train the ability of the large language model to restore the original information.

8. The method according to claim 7, characterized in that, Based on self-supervised random mask training, the instruction fine-tuning model training continues, including: Step B1, setting three types of fine-tuning tasks: route analysis, device management, and performance evaluation, and generating training sample sets for the three types of fine-tuning tasks for model training: the natural language requirement templates are pre-expanded using the large language model, and multiple natural language requirement templates are generated for each type of fine-tuning task, which are then filled using a random algorithm. For multiple natural language requirements with the same intent, other input and output information is shared, and corresponding fine-tuning task labels are set for the training samples; Step B2, during each instruction fine-tuning model training, the data ratio of each type of fine-tuning task is set, so that the large language model automatically performs random mixed training on the three types of fine-tuning tasks during instruction fine-tuning model training. The loss function is calculated as the average cross-entropy loss of the three types of fine-tuning tasks to balance the model's capabilities.

9. A system for enhancing the network operation and maintenance capabilities of large language models, characterized in that, It includes: a sample acquisition device that collects multiple training samples to form a training sample set. The acquisition process for each training sample is as follows: randomly construct simulated network data, then input the simulated network data into a network simulator, and collect device configuration data, network topology data, and communication link status data from the simulation data of the network simulator to generate a text attribute graph of the simulated network. At the same time, a natural language query is generated based on the natural language requirement template. Finally, the corresponding network behavior data is collected from the simulation data generated by the network simulator based on the natural language query. Model building apparatus, constructs and trains augmented large language models, augmented large language models with natural language query Q and text attribute graphs. The input is the natural language query Q, and the output is the answer corresponding to the natural language query Q. The workflow is as follows: The natural language query Q and the text attribute graph are compared. The configuration text of all devices is used to extract semantic features through a configuration encoder based on a self-attention mechanism to generate configuration-query association embeddings. At the same time, the text attribute map Global topological embedding is obtained by using a graph neural network-based topological encoder as the output. Then and splicing into a network representation Then, through low-rank fine-tuning, After being mapped to a trainable low-rank matrix, the data is added to the Q and V layers of the attention mechanism of the large language model. Finally, the large language model infers and outputs the answer content corresponding to the natural language query Q. The model application device receives the natural language query input by the user and simultaneously collects device configuration data, network topology data, and communication link status data from the current network to generate a text attribute graph of the current network. Then, the natural language query input by the user and the text attribute graph of the current network are input into the trained augmented large language model. Finally, the augmented large language model infers and outputs the answer content corresponding to the natural language query input by the user.

10. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the method for enhancing the network operation and maintenance capabilities of large language models as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed by a processor, implement the steps of the method for enhancing the network operation and maintenance capabilities of large language models as described in any one of claims 1-8.

Images