Network operation track sample generation method, system, computer device and program product

Abstract

The application provides a network operation track sample generation method and system, computer equipment and program product, relates to the computer application technical field, and constructs a network topology state machine according to network topology state data;Based on the network topology state machine, a plurality of candidate troubleshooting tracks are generated;A one-to-one corresponding dynamic isolation network sandbox environment is constructed for each candidate troubleshooting track, and the candidate troubleshooting tracks are executed respectively after injecting corresponding network faults;According to the execution result, the target troubleshooting track is screened out, and the track that cannot be executed or cannot effectively recover the fault can be eliminated;The target troubleshooting track is associated with the network topology state data, the fault injection configuration and the execution result to form a structured track sample, which can improve the integrity and traceability of the sample.The application realizes the closed loop of network operation track generation, verification, screening and structured processing, improves the quality, executability, verifiability and training applicability of the network operation track sample.

Description

Technical Field

[0001] This application relates to the field of computer application technology, and in particular to a method, system, computer equipment, and program product for generating network operation and maintenance trajectory samples. Background Technology

[0002] As network systems continue to expand in scale and become increasingly complex, network operation and maintenance models need to be trained using operation and maintenance trajectory samples that include network faults, diagnostic processes, repair actions, and execution results in order to improve network fault diagnosis and repair capabilities.

[0003] Existing network operation and maintenance trajectory samples mainly come from historical fault records, manual compilation, or model generation. Because the fault scenarios and troubleshooting trajectories are not fully correlated with the actual network topology of the target network system, the generated fault injection configurations and troubleshooting operations may not match the target network system, making it difficult to reproduce fault scenarios and resulting in insufficient applicability of the troubleshooting trajectories.

[0004] Furthermore, existing methods typically remain at the level of text generation or static analysis, lacking an isolated network environment corresponding to each candidate troubleshooting trajectory. This makes it difficult to execute multiple candidate troubleshooting trajectories under the same network topology and fault conditions, and thus cannot verify whether the troubleshooting trajectory can be executed or whether the network fault can be recovered based on the actual execution situation.

[0005] For the multiple candidate troubleshooting trajectories generated, existing methods lack a mechanism for filtering based on actual execution results, which may allow troubleshooting trajectories that cannot be executed or whose faults cannot be recovered to enter the training data. At the same time, network topology state data, fault injection configuration, troubleshooting trajectories and their execution results are usually separated from each other, making it difficult to form structured training samples with complete correlation, which affects the verifiability, traceability and training applicability of network operation and maintenance trajectory samples.

[0006] Therefore, how to generate high-quality, executable, verifiable operation and maintenance trajectory samples containing complete logical reasoning chains has become a technical problem that needs to be solved by existing technologies. Summary of the Invention

[0007] The technical problem solved by this application is to provide a method, system, computer equipment and program product for generating network operation and maintenance trajectory samples, so as to overcome the defects of low quality, poor executability and insufficient verifiability of existing network operation and maintenance trajectory samples.

[0008] Firstly, this application proposes a method for generating network operation and maintenance trajectory samples, the method comprising: Obtain network topology state data of the target network system, and construct a network topology state machine based on the network topology state data; Based on the network topology state machine, an autoregressive language model is used to generate fault injection configurations and multiple candidate troubleshooting trajectories to characterize the network faults to be injected. Based on the network topology state machine, a dynamic isolated network sandbox environment is constructed that corresponds one-to-one with each of the candidate obstacle removal trajectories. According to the fault injection configuration, each of the dynamic isolation network sandbox environments is configured to the corresponding network fault state, and the corresponding candidate troubleshooting trajectory is executed in each of the configured dynamic isolation network sandbox environments to obtain the execution result; Based on the execution results, each candidate obstacle removal trajectory is filtered to obtain at least one target obstacle removal trajectory; The target troubleshooting trajectories are associated with the corresponding network topology status data, fault injection configurations, and execution results to generate structured trajectory samples for training network operation and maintenance models.

[0009] In one embodiment, the network topology state data includes at least one of a node set, a link set, a route set, a service policy set, an observability configuration, and a security boundary configuration; The step of constructing a network topology state machine based on the network topology state data includes: The network topology state data is parsed and structured to generate at least one type of state information that characterizes network node state, link connection relationship, routing and forwarding relationship, service operation constraints, observability information, and security control boundary. Based on the state information, network nodes, links, and service relationships are associated and modeled to generate a network topology state machine that represents the operating state of the target network system.

[0010] In one embodiment, the step of generating a fault injection configuration and multiple candidate troubleshooting trajectories to characterize the network fault to be injected, based on the network topology state machine and using an autoregressive language model, includes: A first conditional generation context is constructed based on the network topology state machine, and without inputting an artificial fault seed to specify a specific network fault, the autoregressive language model performs autoregressive decoding using a probability sampling strategy under the constraints of the first conditional generation context to generate the fault injection configuration. The network topology state machine is conditionally bound to the fault injection configuration to construct a second condition generation context. Under the constraints of the second condition generation context, the autoregressive language model performs multiple independent autoregressive decodings to obtain multiple candidate troubleshooting trajectories for the network fault represented by the fault injection configuration.

[0011] In one embodiment, constructing the first condition generation context based on the network topology state machine includes: The network topology state machine is serialized to generate a topology context sequence, which is used to represent at least one of network node state, link state, routing state and service state. Construct system prompt words; the system prompt words are used to limit the type range of network faults to be injected, the structured output format of fault injection configuration, the structured output format of candidate troubleshooting trajectories, and the network operation and maintenance operations that are allowed to be executed; The topology context sequence and the system prompt words are combined in a preset order, and a first structured identifier for identifying the fault injection configuration output position is inserted at the end of the combination result to form the first condition generation context.

[0012] In one embodiment, the fault injection configuration includes a fault injection object; the network topology state machine includes network nodes, links, and services, each with a network entity identifier. The step of conditionally binding the network topology state machine with the fault injection configuration to construct a second condition generation context includes: Based on the fault injection object, a target network entity with a matching network entity identifier is determined from the network topology state machine. The target network entity includes at least one of a target network node, a target link, and a target service. An association relationship is established between the fault injection configuration and the determined target network entity. The association, the fault injection configuration, the topology context sequence, and the system prompt words are combined in a preset order, and a second structured identifier for identifying the candidate obstacle removal trajectory output position is inserted at the end of the combination result to form the second condition generation context.

[0013] In one embodiment, the step of causing the autoregressive language model to perform multiple independent autoregressive decodings under the constraints of the second conditional generation context to obtain multiple candidate troubleshooting trajectories for the network fault represented by the fault injection configuration includes: Based on the complexity of the network fault represented by the fault injection configuration, the number of autoregressive decoding operations K is determined, where K is an integer greater than 1; While keeping the association, fault injection configuration, topology context sequence and system prompt words unchanged in the second condition generation context, different decoding control parameters are set for K autoregressive decodings. The decoding control parameters include at least one of sampling temperature and decoding random parameters. The second condition is generated by inputting the context into the autoregressive language model, and K independent autoregressive decodings are performed according to the corresponding decoding control parameters. A candidate obstacle removal trajectory is generated in each autoregressive decoding process. By aggregating the candidate troubleshooting trajectories obtained from K autoregressive decodings, multiple candidate troubleshooting trajectories are obtained for the network faults represented by the fault injection configuration.

[0014] In one embodiment, generating a candidate obstacle-clearing trajectory in each autoregressive decoding process includes: Based on the network topology state machine and the fault injection configuration, a diagnostic command for locating the network fault to be injected and the expected observation results corresponding to the diagnostic command are generated. Generate cause analysis information based on the diagnostic command and the expected observation results; Based on the cause analysis information, generate a fault repair action to eliminate the fault in the network to be injected; Based on the fault repair action, a verification action is generated to verify whether the fault in the network to be injected has been recovered; End this autoregressive decoding and take the output sequence generated by this autoregressive decoding, which includes the diagnostic command, the expected observation result, the cause analysis information, the fault repair action and the verification action, as a candidate troubleshooting trajectory.

[0015] In one embodiment, the step of constructing a dynamically isolated network sandbox environment corresponding one-to-one with each of the candidate obstacle-clearing trajectories based on the network topology state machine includes: For each of the aforementioned candidate obstacle removal trajectories, create a separate, isolated sandbox instance; In each sandbox instance, a container node is created according to the network topology state machine, a virtual network interface card, a network address and a routing table are configured for the container node, and the service process represented by the network topology state machine is started in the container node. In each sandbox instance, the service network, transmission network, control network, and observation network are configured according to the network topology state machine, and each container node is connected to the corresponding network according to the network connection relationship represented by the network topology state machine. Each sandbox instance that has completed the configuration of the container node and the corresponding network is initialized to the same initial topology state represented by the network topology state machine, so as to obtain a dynamic isolated network sandbox environment that corresponds one-to-one with each candidate obstacle removal trajectory and is isolated from each other.

[0016] In one embodiment, the fault injection configuration includes fault types and fault parameters, wherein the fault types include at least one of link-related faults, routing-related faults, service-related faults, resource-related faults, and security policy-related faults; The step of configuring each of the dynamically isolated network sandbox environments to a corresponding network fault state according to the fault injection configuration includes: Different injection methods are determined based on the fault type, and each of the dynamic isolated network sandbox environments is configured to the corresponding network fault state using the fault injection method according to the fault parameters. Wherein, when the fault type includes link-related faults, the injection method includes: injecting at least one of packet loss, latency, jitter, or bandwidth limitation into the target virtual link through a network traffic control tool, or simulating link interruption by disabling the target virtual network interface card or deleting the target route; When the fault type includes routing faults, the injection method includes at least one of: modifying route priority, adding an erroneous next hop, and creating a routing black hole; When the fault type includes service-related faults, the injection method includes at least one of the following: stopping the target service through a process management tool, restarting the target service, and loading an abnormal service configuration into the target service; When the fault type includes resource-related faults, the injection method includes: reducing the processor usage quota or memory usage limit of the target container through a resource control component or a container resource limiting component; or, applying processor computing load, memory usage load, or network connection load to the target container through a load generation component. When the fault type includes security policy-related faults, the injection method includes: adding at least one of access blocking rules, port blocking rules, and error forwarding rules through the firewall rule management component.

[0017] In one embodiment, executing the corresponding candidate obstacle-clearing trajectory in each configured dynamic isolation network sandbox environment to obtain the execution result includes: Perform security checks on the commands to be executed in the candidate troubleshooting trajectory, and execute the commands that have passed the security checks according to the candidate troubleshooting trajectory using a non-privileged user or a user granted the minimum operating privileges, and record the command execution information; During execution, monitoring probes are used to collect the running status data of the corresponding dynamic isolated network sandbox environment, and execution is terminated when the execution time exceeds the preset time. Based on the command execution information and the changes in the running status data before and after execution, execution results are generated to characterize the command execution status, fault recovery status, execution time, resource consumption status, operational safety status, and repair verification status. After completing the collection of the operational status data, the corresponding dynamic isolated network sandbox environment is destroyed or rolled back.

[0018] In one embodiment, the step of filtering each candidate obstacle removal trajectory based on the execution result includes: Based on the execution results corresponding to each candidate troubleshooting trajectory, indicator parameters are determined. The indicator parameters include at least one of the following: execution time indicator, resource consumption indicator, minimum change indicator, as well as recovery success indicator, command executability indicator, operation safety indicator, and verification closed loop indicator. Based on the preset weights corresponding to each indicator parameter, a comprehensive evaluation of each indicator parameter is performed to obtain the absolute evaluation result of each candidate obstacle removal trajectory. Candidate troubleshooting trajectories that do not meet at least one of the following preset quality conditions: recovery success index, command executability index, operational safety index, and verification closed-loop index; From the remaining candidate obstacle removal trajectories, the candidate obstacle removal trajectories whose absolute evaluation results reach the preset evaluation threshold are selected as the target obstacle removal trajectories.

[0019] In one embodiment, the step of filtering each candidate obstacle removal trajectory based on the execution result further includes: For multiple candidate troubleshooting trajectories corresponding to the same network fault to be injected, the average level and dispersion within the group are determined based on the absolute evaluation results of each candidate troubleshooting trajectory. Based on the degree of deviation of the absolute evaluation result of each candidate obstacle removal trajectory from the average level within the group, and in combination with the degree of dispersion within the group, the relative evaluation result within the group for each candidate obstacle removal trajectory is determined. From the candidate obstacle removal trajectories whose absolute evaluation results reach the preset evaluation threshold, further filter the candidate obstacle removal trajectories whose relative evaluation results within the group meet the preset relative evaluation conditions, and use them as the final target obstacle removal trajectories.

[0020] In one embodiment, the step of associating the target troubleshooting trajectory with the corresponding network topology status data, fault injection configuration, and execution results to generate a structured trajectory sample for training a network operation and maintenance model includes: Generate sample identifiers for each of the target obstacle removal trajectories, and obtain the topology identifiers of the corresponding network topology status data; Extract the fault type from the fault injection configuration, and generate a troubleshooting task to describe the network fault to be investigated based on the fault injection configuration; Diagnostic information and action sequences are extracted from the target obstacle removal trajectory, and the absolute evaluation results and intra-group relative evaluation results corresponding to the target obstacle removal trajectory are obtained. Generate sample labels based on the execution results, absolute evaluation results, and within-group relative evaluation results; According to the preset field mapping relationship, the sample identifier, topology identifier, network topology status data, fault type, fault injection configuration, troubleshooting task, diagnostic information, action sequence, execution result and sample label are written into the corresponding sample fields to form a sample record corresponding to the target troubleshooting trajectory. Each of the sample records is output in a structured data format that records line by line to generate the structured trajectory sample.

[0021] In one embodiment, the step of associating the target troubleshooting trajectory with the corresponding network topology status data, fault injection configuration, and execution results to generate structured trajectory samples for training the network operation and maintenance model further includes: The training purpose label of the sample record corresponding to the target obstacle removal trajectory whose execution result represents successful fault recovery and operational safety meets the preset safety conditions, whose absolute evaluation result reaches the preset evaluation threshold, and whose relative evaluation result within the group meets the preset relative evaluation conditions is set as a positive sample label to obtain supervised fine-tuning positive samples; From the candidate troubleshooting trajectories generated for the same network fault to be injected, select at least one of the following candidate troubleshooting trajectories as rejection trajectories: the absolute evaluation result does not reach the preset evaluation threshold, the relative evaluation result within the group is lower than the preset rejection threshold, and the execution result indicates that the fault has not been recovered. Generate rejection sample records corresponding to the rejection trajectory according to the preset field mapping relationship, and set the training purpose label of the rejection sample records as the preference optimization rejection sample label; Based on the same topology identifier and fault injection configuration, the supervised fine-tuning positive samples are used as preferred samples, and the rejected samples are recorded as rejected samples to generate sample pairs for preference optimization training.

[0022] Secondly, this application proposes a system for generating network operation and maintenance trajectory samples, the system comprising: The acquisition module is used to acquire network topology status data of the target network system; The first construction module is used to construct a network topology state machine based on the network topology state data; and based on the network topology state machine, to generate a fault injection configuration and multiple candidate troubleshooting trajectories to characterize the network fault to be injected through an autoregressive language model. The second construction module is used to construct a dynamic isolated network sandbox environment that corresponds one-to-one with each of the candidate obstacle removal trajectories based on the network topology state machine. The execution module is used to configure each of the dynamic isolation network sandbox environments to the corresponding network fault state according to the fault injection configuration, and to execute the corresponding candidate troubleshooting trajectory in each of the configured dynamic isolation network sandbox environments to obtain the execution result; The sample generation module is used to filter each candidate troubleshooting trajectory according to the execution result to obtain at least one target troubleshooting trajectory; and associate each target troubleshooting trajectory with the corresponding network topology status data, fault injection configuration and execution result to generate a structured trajectory sample for training the network operation and maintenance model.

[0023] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method of the first aspect.

[0024] Fourthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the method in the first aspect.

[0025] The methods, systems, computer equipment, and program products for generating the aforementioned network operation and maintenance trajectory samples have at least the following advantages: This application constructs a network topology state machine based on network topology state data, providing a unified topology basis for fault and troubleshooting trajectory generation, thus improving the matching degree between the generated content and the target network system. Based on the network topology state machine, it generates fault injection configurations and multiple candidate troubleshooting trajectories, improving trajectory generation efficiency and the richness of troubleshooting solutions. It constructs a one-to-one dynamically isolated network sandbox environment for each candidate troubleshooting trajectory, and executes the candidate troubleshooting trajectories separately after injecting the corresponding network fault, avoiding mutual interference between trajectories and obtaining actual execution results. It filters the target troubleshooting trajectory based on the execution results, eliminating trajectories that are unexecutable or cannot effectively recover from faults. It associates the target troubleshooting trajectory with network topology state data, fault injection configurations, and execution results to form structured trajectory samples, improving the integrity and traceability of the samples. Therefore, this application realizes a closed loop for the generation, verification, filtering, and structured processing of network operation and maintenance trajectories, improving the quality, executability, verifiability, and training applicability of network operation and maintenance trajectory samples. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a method for generating network operation and maintenance trajectory samples in one embodiment; Figure 2 This is a flowchart illustrating the steps involved in constructing a network topology state machine in one embodiment. Figure 3 This is a flowchart illustrating the steps for generating a fault injection configuration and multiple candidate troubleshooting trajectories in one embodiment. Figure 4 This is a flowchart illustrating the steps involved in constructing a dynamically isolated network sandbox environment in one embodiment. Figure 5 This is a flowchart illustrating the steps for obtaining the execution result in one embodiment; Figure 6 This is a block diagram of a system for generating network operation and maintenance trajectory samples in one embodiment. Detailed Implementation

[0027] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0028] Some exemplary embodiments of this application have been described for illustrative purposes. It should be understood that this application may be implemented in other ways not specifically shown in the accompanying drawings.

[0029] Please see Figure 1 In one exemplary embodiment, this application provides a method for generating network operation and maintenance trajectory samples, specifically including the following steps: Step 102: Obtain network topology status data of the target network system.

[0030] Specifically, the target network system refers to the physical network system, virtual network system, or its digital mapping system that requires the generation of network operation and maintenance trajectory samples. Network topology status data refers to data that reflects the composition, connection relationships, and operational status of the target network system. Generally, network topology status data can be obtained by collecting relevant data through at least one of the target network system's network management interface, configuration management interface, and status monitoring interface, and then performing format conversion and normalization processing on the collected data.

[0031] Step 104: Construct a network topology state machine based on the network topology state data; based on the network topology state machine, generate a fault injection configuration and multiple candidate troubleshooting trajectories to characterize the network faults to be injected using an autoregressive language model.

[0032] Specifically, a network topology state machine refers to a data model built based on network topology state data, used to uniformly represent the topology, operational status, and relationships between different network entities of a target network system, providing a unified data foundation for subsequent fault scenario self-generation, sandbox instantiation, and trajectory verification. An autoregressive language model is a language model that predicts and generates subsequent content sequentially based on existing input content. Fault injection configuration refers to data used to describe the network fault to be injected and can serve as the basis for fault configuration; it typically includes the fault injection object, fault type, and fault parameters. A candidate troubleshooting trajectory refers to an operation sequence generated from multiple network operation and maintenance operations in the execution order, created for the network fault to be injected.

[0033] In practical implementation, the network topology state data is processed to establish network entities and their relationships, resulting in a network topology state machine. The information represented by the network topology state machine is then provided to the autoregressive language model, enabling the autoregressive language model to generate a fault injection configuration that matches the target network system. Furthermore, multiple different candidate troubleshooting trajectories are generated for the network faults to be injected as represented by the fault injection configuration.

[0034] Step 106: Based on the network topology state machine, construct a dynamic isolated network sandbox environment that corresponds one-to-one with each candidate obstacle removal trajectory.

[0035] Specifically, a dynamic isolated network sandbox environment refers to a network testing environment established based on a network topology state machine, capable of simulating the topology and operating state of the target network system, and isolated from other execution environments. In this embodiment, each candidate troubleshooting trajectory corresponds one-to-one with a dynamic isolated network sandbox environment, so that each candidate troubleshooting trajectory has an independent execution environment.

[0036] In the specific implementation, network sandbox environments corresponding to the number of candidate troubleshooting trajectories are created according to the network topology state machine. The target network system represented by the network topology state machine is reproduced in each network sandbox environment. The correspondence between each candidate troubleshooting trajectory and the corresponding network sandbox environment is established, resulting in multiple mutually isolated dynamic network sandbox environments.

[0037] Step 108: Based on the fault injection configuration, configure each dynamic isolation network sandbox environment to the corresponding network fault state, and execute the corresponding candidate troubleshooting trajectory in each configured dynamic isolation network sandbox environment to obtain the execution result.

[0038] Specifically, network fault state refers to the abnormal operating state formed after the dynamic isolated network sandbox environment undergoes corresponding state changes according to the fault injection configuration. Execution result refers to the actual execution of the candidate troubleshooting trajectory in the corresponding dynamic isolated network sandbox environment and the state changes generated in the dynamic isolated network sandbox environment during the execution process.

[0039] In the specific implementation, the fault injection configuration is parsed, and the corresponding fault configuration operation is performed on each dynamic isolation network sandbox environment according to the parsing result, so that each dynamic isolation network sandbox environment enters the corresponding network fault state. Subsequently, the corresponding candidate troubleshooting trajectory is executed in each dynamic isolation network sandbox environment, and the execution process of the candidate troubleshooting trajectory and the state change of the corresponding dynamic isolation network sandbox environment are recorded to obtain the execution result corresponding to each candidate troubleshooting trajectory.

[0040] Step 110: Filter each candidate troubleshooting trajectory according to the execution result to obtain at least one target troubleshooting trajectory; associate each target troubleshooting trajectory with the corresponding network topology status data, fault injection configuration and execution result to generate a structured trajectory sample for training the network operation and maintenance model.

[0041] Specifically, the target troubleshooting trajectory refers to the troubleshooting trajectory that meets preset requirements and is selected from multiple candidate troubleshooting trajectories based on the execution results. Structured trajectory samples refer to machine-recognizable training data formed by associating and recording network topology status data, fault injection configurations, target troubleshooting trajectories, and execution results according to a preset data structure.

[0042] In the specific implementation, the execution results corresponding to each candidate troubleshooting trajectory are analyzed. Based on the analysis results, it is determined whether each candidate troubleshooting trajectory meets the preset screening conditions, and at least one candidate troubleshooting trajectory that meets the preset screening conditions is determined as the target troubleshooting trajectory. According to the correspondence between each target troubleshooting trajectory and network topology status data, fault injection configuration and execution results, the relevant data are associated and structured to generate structured trajectory samples that can be used to train the network operation and maintenance model.

[0043] The aforementioned method for generating network operation and maintenance trajectory samples constructs a network topology state machine based on network topology state data, providing a unified topology basis for fault and troubleshooting trajectory generation and improving the matching degree between the generated content and the target network system. Based on the network topology state machine, it generates fault injection configurations and multiple candidate troubleshooting trajectories, improving trajectory generation efficiency and the richness of troubleshooting solutions. It constructs a one-to-one dynamically isolated network sandbox environment for each candidate troubleshooting trajectory and executes the candidate troubleshooting trajectories after injecting the corresponding network fault, avoiding mutual interference between trajectories and obtaining actual execution results. It filters the target troubleshooting trajectory based on the execution results, eliminating trajectories that are unexecutable or cannot effectively recover from faults. It associates the target troubleshooting trajectory with network topology state data, fault injection configurations, and execution results to form a structured trajectory sample, improving the integrity and traceability of the sample. Therefore, this application achieves a closed loop for the generation, verification, filtering, and structured processing of network operation and maintenance trajectories, improving the quality, executability, verifiability, and training applicability of network operation and maintenance trajectory samples.

[0044] Please see Figure 2 Optionally, the network topology state data includes at least one of the following: node set, link set, route set, service policy set, observability configuration, and security boundary configuration; a network topology state machine is constructed based on the network topology state data, including: Step 202: Parse and structure the network topology state data to generate at least one state information that represents network node state, link connection relationship, routing and forwarding relationship, service operation constraints, observability information and security control boundary.

[0045] Step 204: Based on the state information, perform association modeling of network nodes, links and service relationships to generate a network topology state machine that represents the operating state of the target network system.

[0046] Specifically, the acquired network topology status data is first subjected to data format identification and field parsing, and the network entity identifier, status field and relationship field are formatted to convert network topology status data from different sources or with different data formats into a preset structured data format.

[0047] For a set of nodes, extract the network entity identifier and node status corresponding to each network node to generate network node status; for a set of links, extract the network entity identifiers and link status corresponding to both ends of the link to generate link connection relationships; for a set of routes, extract the starting object, target object, and forwarding object of the route to generate routing and forwarding relationships; for a set of service policies, extract the target object, applicable conditions, and policy content of the service policies to generate service operation constraints; for observability configuration, extract the monitoring objects and their corresponding status collection methods to generate observability information; for security boundary configuration, extract the security zone, controlled object, and their access control relationships to generate security control boundaries. When the network topology status data includes any of the above types of data, generate the status information corresponding to that data.

[0048] Furthermore, the generated state information is matched based on network entity identifiers, link connections are associated with corresponding network nodes, routing and forwarding relationships are associated with the network nodes and links they pass through or point to, and service operation constraints are associated with corresponding network nodes, links, or service objects. Simultaneously, observability information and security control boundaries are associated with their respective monitored or controlled objects, thereby establishing the relationships between network nodes, links, and service relationships.

[0049] Furthermore, each network node and its state, each relationship, and the corresponding business constraints are organized according to a preset state data structure to form a network topology state machine that can uniformly represent the target network system's topology, entity states, and relationships between entities. When the network topology state data changes, the states of the corresponding network entities and relationships in the network topology state machine are updated according to the changed data, so that the network topology state machine represents the current operating state of the target network system.

[0050] By adopting the above scheme, through parsing and structuring different types of network topology state data, and by associating and modeling network nodes, links and service relationships, the dispersed network entity states and network relationships can be transformed into a unified, associative and updatable network topology state machine. This enables accurate characterization of the target network system's operating state and provides a consistent topology data foundation for subsequent fault injection configuration generation, candidate troubleshooting trajectory generation, and the construction of a dynamic isolated network sandbox environment.

[0051] Please see Figure 3 Optionally, based on the network topology state machine, an autoregressive language model is used to generate a fault injection configuration and multiple candidate troubleshooting trajectories to characterize the network fault to be injected, including: Step 302: Construct a first condition generation context based on the network topology state machine, and without inputting artificial fault seeds to specify specific network faults, enable the autoregressive language model to perform autoregressive decoding using a probability sampling strategy under the constraints of the first condition generation context to generate fault injection configuration.

[0052] Step 304: Conditionally bind the network topology state machine with the fault injection configuration, construct a second condition generation context, and enable the autoregressive language model to perform multiple independent autoregressive decodings under the constraints of the second condition generation context to obtain multiple candidate troubleshooting trajectories for the network faults represented by the fault injection configuration.

[0053] Specifically, the first conditional generation context provides input information to the autoregressive language model regarding the network topology state and fault generation constraints. Artificial fault seeds refer to input content pre-specified by humans, specifying specific fault types, fault objects, or fault parameters; these differ from the parameters used to control sampling randomness during the autoregressive language model decoding process. The second conditional generation context uses the network topology state machine and the generated fault injection configuration together as generation conditions to form the input information.

[0054] Optionally, a first condition generation context is constructed based on the network topology state machine, including: The network topology state machine is serialized to generate a topology context sequence, which represents at least one of the network node state, link state, routing state, and service state. System prompt words are constructed to limit the type range of network faults to be injected, the structured output format of fault injection configuration, the structured output format of candidate troubleshooting trajectories, and the network operation and maintenance operations that are allowed to be executed. The topology context sequence and system prompt words are combined in a preset order, and a first structured identifier for identifying the output position of fault injection configuration is inserted at the end of the combination result to form a first condition generation context.

[0055] Specifically, serialization processing refers to converting structured data in the network topology state machine into sequential data that can be read in order by an autoregressive language model. In this embodiment, the topology context sequence can take the form of structured text, key-value pair sequences, or token sequences.

[0056] System prompts are instructions used to define the generation task and output requirements of the autoregressive language model. The first structured identifier is a predefined character, string, or token that can be distinguished from other input content and is used to indicate the starting position of the fault injection configuration output.

[0057] In the specific implementation, the network topology state machine is traversed according to the preset data reading order to read network entities, entity states and relationships between entities, and the read content is converted into a topology context sequence with a fixed arrangement order.

[0058] Furthermore, system prompt words are constructed based on the fault generation task, and the system prompt words are set with the type range of network faults to be injected, the fault injection configuration, the structured output format of the candidate troubleshooting trajectory, and the network operation and maintenance operations that are allowed to be executed, so as to constrain the content and data format that the autoregressive language model can generate.

[0059] Furthermore, the system prompts are concatenated with the topology context sequence in a preset order. For example, the system prompts can be placed before the topology context sequence, allowing the autoregressive language model to first acquire the generation rules, then read the topology state of the target network system, and insert a first structured identifier at the end of the concatenation result to form a first conditional generation context. After reading the first structured identifier, the autoregressive language model starts generating the fault injection configuration from the position indicated by that identifier.

[0060] By adopting the above scheme, serializing the network topology state machine and using system prompts and the first structured identifier to constrain the generated content and identify the output position, the autoregressive language model can accurately identify the network topology state and fault generation task, thereby improving the matching degree between fault injection configuration and network topology state and the consistency of output format.

[0061] Optionally, the fault injection configuration includes a fault injection object; the network topology state machine includes network nodes, links, and services, each with a network entity identifier; the network topology state machine is conditionally bound to the fault injection configuration to construct a second condition generation context, including: Based on the fault injection object, a target network entity with a matching network entity identifier is determined from the network topology state machine. The target network entity includes at least one of a target network node, a target link, and a target service. An association relationship is established between the fault injection configuration and the determined target network entity. The association relationship, fault injection configuration, topology context sequence, and system prompt words are combined in a preset order, and a second structured identifier used to identify the candidate troubleshooting trajectory output position is inserted at the end of the combination result to form a second conditional generation context.

[0062] Specifically, a network entity identifier refers to the identification information used to uniquely identify network nodes, links, and services in the network topology state machine. A fault injection object refers to the object to which a network fault is applied, as specified in the fault injection configuration.

[0063] In practice, the network entity identifier corresponding to the fault injection object is extracted from the fault injection configuration, and this network entity identifier is compared with the network entity identifiers of each network entity in the network topology state machine. The network node, link, or service corresponding to the same network entity identifier is identified as the target network entity.

[0064] For each target network entity, an association is established between the fault injection configuration and that target network entity. This association represents the specific network entity affected by the fault content in the fault injection configuration. When the fault injection object includes multiple network entity identifiers, the target network entity corresponding to each network entity identifier is determined, and an association is established between the fault injection configuration and each target network entity.

[0065] Furthermore, the system prompts, topology context sequences, fault injection configurations, and relationships are combined in a preset order. The combined result simultaneously includes the troubleshooting trajectory generation rules, the topology status of the target network system, the network fault to be investigated, and the target network entity corresponding to that fault. A second structured identifier is inserted at the end of the combined result to form a second conditional generation context. After the autoregressive language model reads the second structured identifier, it starts generating candidate troubleshooting trajectories from the position indicated by the identifier.

[0066] By adopting the above scheme, the fault injection object is associated with the target network entity in the network topology state machine through network entity identification. This can clarify the position of the network fault in the target network system and make the candidate troubleshooting trajectory constrained by the network topology state, fault content and fault object at the same time, thereby reducing the situation where the troubleshooting trajectory does not match the actual fault object.

[0067] Optionally, the autoregressive language model is subjected to multiple independent autoregressive decodings under the constraints of the second conditional generation context to obtain multiple candidate troubleshooting trajectories for the network faults represented by the fault injection configuration, including: Based on the complexity of the network fault represented by the fault injection configuration, the number of autoregressive decoding operations K is determined, where K is an integer greater than 1. While maintaining the association relationships, fault injection configuration, topology context sequence, and system prompts in the second condition generation context, different decoding control parameters are set for the K autoregressive decoding operations. These decoding control parameters include at least one of sampling temperature and decoding random parameters. The second condition generation context is input into the autoregressive language model, and K independent autoregressive decoding operations are performed according to the corresponding decoding control parameters. A candidate troubleshooting trajectory is generated during each autoregressive decoding operation. The candidate troubleshooting trajectories obtained from the K autoregressive decoding operations are aggregated to obtain multiple candidate troubleshooting trajectories for the network fault represented by the fault injection configuration.

[0068] Specifically, the complexity of a network fault refers to parameters used to characterize the number of objects involved, the scope of the fault's impact, and the difficulty of troubleshooting. Decoding control parameters are parameters used to control the selection of output content by the autoregressive language model during the decoding process. Among these, sampling temperature is used to adjust the probability differences between different candidate output contents, and decoding random parameters are used to change the sampling path in different autoregressive decoding processes.

[0069] In practice, the fault injection object, fault content, and fault impact range are extracted from the fault injection configuration, and the complexity of the network fault is determined based on the extraction results.

[0070] By establishing a pre-defined correspondence between different complexity levels and the number of autoregressive decoding iterations, the system finds the corresponding number of autoregressive decoding iterations K based on the complexity level of the current network fault, where K is an integer greater than 1. The higher the complexity of the network fault, the more autoregressive decoding iterations are required to generate more candidate troubleshooting solutions.

[0071] During K autoregressive decoding iterations, the associations, fault injection configurations, topology context sequences, and system prompts in the second condition generation context remain unchanged, ensuring that each autoregressive decoding iteration corresponds to the same network topology state and network fault. Decoding control parameters are set for each autoregressive decoding iteration, with each parameter having a different value. For example, different sampling temperatures or different decoding random parameters can be set, or both can be set simultaneously.

[0072] Furthermore, the same second-condition generation context is input into the autoregressive language model, and K autoregressive decodings are performed according to the decoding control parameters corresponding to each decoding. At the start of each autoregressive decoding, the decoding state is reinitialized so that the generated content does not depend on the candidate troubleshooting trajectories generated by other autoregressive decodings. After each autoregressive decoding, its output is recorded as a candidate troubleshooting trajectory, and the outputs of K autoregressive decodings are aggregated to obtain multiple candidate troubleshooting trajectories for the same network fault.

[0073] By adopting the above scheme, the number of autoregressive decodings is determined according to the complexity of the network fault, and different decoding control parameters are used while keeping the network topology and network fault unchanged. This approach can take into account the consistency of fault conditions and the diversity of troubleshooting schemes among the candidate troubleshooting trajectories, and reduce the problem of simple network faults occupying too many decoding resources or the lack of candidate troubleshooting schemes for complex network faults.

[0074] Optionally, a candidate obstacle removal trajectory is generated in each autoregressive decoding process, including: Based on the network topology state machine and fault injection configuration, diagnostic commands for locating the network fault to be injected and the expected observation results corresponding to the diagnostic commands are generated; based on the diagnostic commands and expected observation results, cause analysis information is generated; based on the cause analysis information, fault repair actions for eliminating the network fault to be injected are generated; based on the fault repair actions, verification actions for verifying whether the network fault to be injected has been recovered are generated; the current autoregressive decoding ends, and the output sequence generated by this autoregressive decoding, which includes diagnostic commands, expected observation results, cause analysis information, fault repair actions, and verification actions, is taken as a candidate troubleshooting trajectory.

[0075] Specifically, diagnostic commands refer to network operation and maintenance commands used to obtain the status of the target network system and locate the network fault to be injected. Expected observation results refer to the status information or phenomena expected to be obtained after executing the diagnostic command. Root cause analysis information refers to the causes of the network fault determined based on the diagnostic command and its expected observation results. Fault repair actions refer to operations used to eliminate network faults or restore the target network system to normal operation. Verification actions refer to operations used to check whether the network fault has been restored after the fault repair actions are executed.

[0076] In its implementation, the autoregressive language model begins its autoregressive decoding from the position indicated by the second structured identifier. The model reads the network topology state represented by the network topology state machine and the network faults represented by the fault injection configuration, generates diagnostic commands to check the operational status of relevant network entities, and continues to generate the expected observations corresponding to those diagnostic commands.

[0077] Furthermore, the autoregressive language model uses the generated diagnostic commands and expected observations as prefixes for subsequent content generation, and continues to generate causal analysis information based on the correspondence between them. Then, it uses the causal analysis information as a subsequent generation condition to generate fault repair actions to address the analyzed fault causes. Following this, based on the fault repair actions and their expected state changes, it generates verification actions to determine whether the injected network fault has been recovered. After the verification actions are generated, if the autoregressive language model outputs a preset decoding termination content, or if the preset decoding termination condition is met, the current autoregressive decoding ends.

[0078] The diagnostic commands, expected observation results, cause analysis information, fault repair actions, and verification actions generated sequentially during this autoregressive decoding process constitute an output sequence, which is then used as a candidate troubleshooting trajectory.

[0079] By adopting the above scheme, corresponding content is generated sequentially according to the dependencies between diagnosis, analysis, repair and verification in the same autoregressive decoding process. This enables the various operations in the candidate troubleshooting trajectory to be connected and form a complete troubleshooting loop, thereby improving the logical integrity, executability and verifiability of the candidate troubleshooting trajectory.

[0080] Please see Figure 4 Based on the network topology state machine, a dynamically isolated network sandbox environment is constructed, corresponding one-to-one with each candidate obstacle-clearing trajectory, including: Step 402: For each candidate obstacle removal trajectory, create a sandbox instance that is isolated from each other.

[0081] Step 404: In each sandbox instance, create a container node based on the network topology state machine, configure virtual network interface cards, network addresses and routing tables for the container node, and start the service process represented by the network topology state machine in the container node.

[0082] Step 406: In each sandbox instance, configure the service network, transport network, control network, and observation network according to the network topology state machine, and connect each container node to the corresponding network according to the network connection relationship represented by the network topology state machine.

[0083] Step 408: Initialize each sandbox instance that has completed container node and corresponding network configuration to the same initial topology state represented by the network topology state machine, so as to obtain a dynamic isolated network sandbox environment that corresponds one-to-one with each candidate obstacle removal trajectory and is mutually isolated.

[0084] Specifically, a sandbox instance refers to an independent runtime space used to host simulated network nodes, network connections, and service processes. Network communication, process execution, and resource usage are independent of each other between different sandbox instances, and state changes in one sandbox instance will not affect other sandbox instances. In this embodiment, each candidate troubleshooting trajectory has a dedicated sandbox instance for executing that candidate troubleshooting trajectory.

[0085] In practice, the number of candidate obstacle removal trajectories is obtained, and a corresponding number of sandbox instances are created. Each sandbox instance is assigned a sandbox identifier, and each candidate obstacle removal trajectory is assigned a trajectory identifier, establishing a one-to-one correspondence between sandbox identifiers and trajectory identifiers.

[0086] Furthermore, a container node refers to a container runtime instance created within a sandbox instance to simulate the network nodes represented by the network topology state machine.

[0087] In practice, the network nodes represented by the network topology state machine are read, and container nodes corresponding to each network node are created in each sandbox instance. A correspondence between network entity identifiers and container node identifiers in the network topology state machine is established so that the corresponding container node can be located in the sandbox instance based on the network entity identifier.

[0088] Based on the network connection and communication requirements represented by the network topology state machine, a corresponding virtual network interface card (NIC) is created for each container node, and a network address is configured for the virtual NIC. Based on the routing state represented by the network topology state machine, a routing table is written into the corresponding container node, enabling data to be transmitted according to the forwarding relationships specified by the network topology state machine. Based on the service deployment state represented by the network topology state machine, the service runtime environment is loaded into the corresponding container node, and the corresponding service processes are started, giving each container node the service capabilities corresponding to the network node it simulates.

[0089] Furthermore, a business network, transport network, control network, and observation network are created in each sandbox instance, and corresponding network identifiers and address ranges are configured for each network. It should be understood that these networks in different sandbox instances are isolated from each other to avoid data transmission across sandbox instances. Specifically, the business network is used to simulate business data access paths; the transport network is used to simulate the link connections between different network nodes; the control network is used to issue pending commands from candidate troubleshooting trajectories to container nodes; and the observation network is used to collect runtime status data from each container node.

[0090] Based on the network connectivity represented by the network topology state machine, the network that each container node needs to access is determined, and the virtual network interface card (NIC) corresponding to that container node is connected to the appropriate network. Specifically, communication relationships between business access objects are established through the business network; communication relationships between container nodes with link connections are established through the transport network; command delivery channels between candidate troubleshooting trajectory execution components and container nodes are established through the control network; and status data acquisition channels between monitoring probes and container nodes are established through the observation network.

[0091] Furthermore, each sandbox instance is initialized, and the container nodes, virtual network cards, network addresses, routing tables, network connection relationships, and service processes in each sandbox instance are checked according to the network topology state machine to ensure that the network nodes, network connections, routing relationships, and service operation status in different sandbox instances are consistent.

[0092] After initialization, the status of container nodes and network connections is checked. When each sandbox instance reaches the initial topology state represented by the network topology state machine, the sandbox instance is marked as available, and each sandbox instance is determined as a dynamic isolated network sandbox environment for the corresponding candidate troubleshooting trajectory based on the correspondence between trajectory identifiers and sandbox identifiers.

[0093] By adopting the above scheme, and creating isolated sandbox instances for each candidate troubleshooting trajectory, and configuring container nodes, network connections, and service processes in each sandbox instance according to the network topology state machine, the corresponding network environment can be reproduced without affecting the target network system and other candidate troubleshooting trajectories. Furthermore, by initializing each sandbox instance to the same initial topology state, different candidate troubleshooting trajectories can be verified under consistent execution conditions. This improves the independence of candidate troubleshooting trajectory execution verification and the comparability of execution results.

[0094] Optionally, the fault injection configuration includes fault types and fault parameters. Fault types include at least one of link-related faults, routing faults, service-related faults, resource-related faults, and security policy-related faults. Based on the fault injection configuration, each dynamically isolated network sandbox environment is configured to the corresponding network fault state, including: Different injection methods are determined based on the fault type, and each dynamic isolated network sandbox environment is configured to the corresponding network fault state according to the fault parameters using the fault injection method.

[0095] In cases where the fault type includes link-related faults, the injection methods include: injecting at least one of packet loss, latency, jitter, or bandwidth limitation into the target virtual link through network traffic control tools, or simulating link interruption by disabling the target virtual network interface card or deleting the target route.

[0096] When the fault type includes routing faults, the injection methods include at least one of the following: modifying route priority, adding an erroneous next hop, and creating a routing black hole.

[0097] When the fault type includes service-related faults, the injection methods include at least one of the following: stopping the target service through process management tools, restarting the target service, and loading abnormal service configurations into the target service.

[0098] When the fault type includes resource-related faults, the injection methods include: reducing the processor usage quota or memory usage limit of the target container through resource control components or container resource limiting components; or, applying processor computing load, memory usage load or network connection load to the target container through load generation components.

[0099] In cases where the fault type includes security policy-related faults, the injection methods include: adding at least one of the following through the firewall rule management component: access blocking rules, port blocking rules, and error forwarding rules.

[0100] Specifically, fault type refers to the category to which the fault in the network to be injected belongs. Fault parameters are data used to determine the object affected by the fault, the form of the fault, the degree of the fault, and the duration of the fault.

[0101] Link-related failures refer to failures that affect the quality of data transmission or connectivity between network nodes; routing-related failures refer to failures that affect the network data forwarding path; service-related failures refer to failures that affect the operation of service processes or service configuration; resource-related failures refer to failures where the available computing resources, storage resources, or network connection resources of the target container are abnormal; and security policy-related failures refer to failures caused by abnormal access control or data forwarding rules.

[0102] In practical implementation, the fault injection configuration is parsed to extract the fault type and fault parameters. Based on the fault target indicated by the fault injection configuration, the corresponding target virtual link, target route, target service, target container, or target security boundary is determined in each dynamic isolated network sandbox environment. The appropriate fault injection component is selected based on the fault type, and the fault parameters are converted into control parameters that the fault injection component can recognize.

[0103] For the same fault injection configuration, the same target network entity is identified in each dynamic isolated network sandbox environment, and fault injection is performed using the same fault parameters, so that each dynamic isolated network sandbox environment forms a network fault state that is independent of each other and has consistent fault conditions. When the fault injection configuration includes multiple fault types, the injection method corresponding to each fault type is determined separately, and they are executed sequentially according to the preset injection order.

[0104] After fault injection is completed, the operational status of the target network entity is collected through monitoring probes, and the collected operational status is compared with the expected fault status represented by the fault injection configuration. When the operational status matches the expected fault status, the corresponding dynamic isolation network sandbox environment is marked as configured and ready for execution of the corresponding candidate troubleshooting trajectory.

[0105] By adopting the above scheme, and selecting the appropriate fault injection method according to the fault type, and using fault parameters to control the target and degree of anomaly of the fault, the structured fault injection configuration can be transformed into actual network state changes in a dynamically isolated network sandbox environment, supporting the controllable reproduction of different types of network faults. Furthermore, by injecting network faults with consistent conditions into each dynamically isolated network sandbox environment, a mutually isolated and comparable fault execution environment can be provided for multiple candidate troubleshooting trajectories, thereby improving the authenticity, reproducibility, and reliability of candidate troubleshooting trajectory verification.

[0106] Please see Figure 5 In each configured dynamic isolation network sandbox environment, the corresponding candidate obstacle removal trajectory is executed to obtain the execution results, including: Step 502: Perform security checks on the commands to be executed in the candidate troubleshooting trajectory, and execute the commands that have passed the security check according to the candidate troubleshooting trajectory using a non-privileged user or a user granted the minimum operating privileges, and record the command execution information.

[0107] Step 504: During the execution process, the running status data of the corresponding dynamic isolation network sandbox environment is collected by the monitoring probe, and the execution is terminated when the execution time exceeds the preset time.

[0108] Step 506: Based on the command execution information and the changes in the running status data before and after execution, generate execution results to characterize the command execution status, fault recovery status, execution time, resource consumption status, operational safety status, and repair verification status.

[0109] Step 508: After completing the collection of runtime status data, destroy or roll back the corresponding dynamic isolated network sandbox environment.

[0110] Specifically, commands to be executed refer to network operation and maintenance commands extracted from candidate troubleshooting trajectories that need to be actually executed in the dynamic isolated network sandbox environment. Security checks refer to the process of determining whether commands to be executed comply with preset operational security rules before execution.

[0111] In practice, commands to be executed are extracted sequentially according to their order in the candidate troubleshooting trajectory. The command type, target object, permissions, and scope of impact of each command are then checked to ensure they meet the preset security rules. Commands that fail the security check are prohibited from execution, and the reason for rejection is recorded. Commands that pass the security check are assigned to a user based on their required permissions.

[0112] In this embodiment, when the command to be executed does not require special operating permissions, it is executed by a non-privileged user; when the command to be executed requires specific operating permissions, the user is granted the minimum operating permissions required to complete the command, and the user executes the command. During execution, the execution order, start and end times, execution status, and command output of each command to be executed are recorded to form command execution information.

[0113] Furthermore, operational status data refers to data that reflects the operational status of networks, services, and resources in a dynamically isolated network sandbox environment. Monitoring probes continuously collect operational status data before, during, and after execution. When a candidate troubleshooting trajectory completes execution within a preset timeframe, its actual execution time is recorded; when the execution time exceeds the preset timeframe, the execution of the current candidate troubleshooting trajectory is terminated, and the execution timeout status and termination location are recorded in the command execution information.

[0114] Furthermore, based on the execution status, command output, and reasons for refusal of each command to be executed, the command execution status of the candidate troubleshooting trajectory is determined; based on the changes in the running status data before and after execution, and whether the running status after execution meets the preset normal operating conditions, the fault recovery status is determined; based on the start execution time, end execution time, or termination time of the candidate troubleshooting trajectory, the execution time is determined.

[0115] Based on the changes in the operational status data collected during execution relative to the operational status data before execution, the resource consumption during the execution of the candidate troubleshooting trajectory is determined; based on the security check results of the command to be executed, the actual execution permissions, and whether any unexpected status changes occur after execution, the operational safety status is determined; based on whether the execution status of the verification operation in the candidate troubleshooting trajectory and its corresponding operational status data meet the preset verification conditions, the repair verification status is determined.

[0116] The execution status of the above commands, fault recovery status, execution time, resource consumption, operational safety status, and repair verification status are associated with the corresponding candidate troubleshooting trajectories, and recorded according to the preset result data format to form the execution result corresponding to the candidate troubleshooting trajectory.

[0117] Furthermore, after confirming that command execution information, runtime status data, and execution results have all been recorded, the environment handling method is determined based on the subsequent usage requirements of the dynamic isolated network sandbox environment. For example, for dynamic isolated network sandbox environments that are no longer in use, their occupied resources are released and they are destroyed; for dynamic isolated network sandbox environments that need to continue to be used for verifying other candidate troubleshooting trajectories, they are rolled back to the preset state before fault injection or before the execution of candidate troubleshooting trajectories, to avoid the state changes caused by this execution affecting the subsequent verification process.

[0118] By employing the above scheme, the risks of dangerous and unauthorized operations can be reduced by performing command security checks before execution and using restricted permissions to execute the commands to be executed. By continuously collecting operational status data through monitoring probes and setting execution time limits, the actual execution status of candidate troubleshooting trajectories can be obtained, and abnormal execution can be prevented from continuously consuming resources. By combining command execution information and changes in operational status to generate execution results, the executability, fault recovery effect, and operational safety of candidate troubleshooting trajectories can be objectively characterized. By destroying or rolling back the sandbox environment after data collection, mutual interference between different verification processes can be avoided. Therefore, this application improves the security of the candidate troubleshooting trajectory execution verification process and the reliability of the execution results.

[0119] Optionally, the candidate troubleshooting trajectories can be filtered based on the execution results, including: Based on the execution results corresponding to each candidate troubleshooting trajectory, indicator parameters are determined. These parameters include at least one of the following: execution time, resource consumption, minimum change, recovery success, command executability, operational safety, and verification loop closure. Each indicator parameter is comprehensively evaluated according to its preset weight to obtain an absolute evaluation result for each candidate troubleshooting trajectory. Candidate troubleshooting trajectories that do not meet at least one of the preset quality conditions for recovery success, command executability, operational safety, and verification loop closure are eliminated. From the remaining candidate troubleshooting trajectories, those whose absolute evaluation results reach a preset evaluation threshold are selected as the target troubleshooting trajectories.

[0120] Specifically, the execution results corresponding to each candidate troubleshooting trajectory are read separately. Based on the fault recovery status represented by the execution results, the recovery success index is determined. Based on the command execution status, the command executability index is determined. Based on the operation safety status, the operation safety index is determined. Based on the repair verification status, the verification closed-loop index is determined.

[0121] Furthermore, based on the execution time, resource consumption, and state changes before and after execution as represented by the execution results, at least one of the following indicators is determined: execution time indicator, resource consumption indicator, and minimum change indicator. Different indicator parameters are then converted to a unified evaluation scale, so that higher-quality indicator parameters correspond to better evaluation values.

[0122] Each indicator parameter is pre-set with a corresponding weight, and a comprehensive evaluation is performed based on each indicator parameter and its weight to obtain the absolute evaluation result of each candidate obstacle removal trajectory. The pre-set weights can be determined according to the degree of influence of different indicator parameters on the quality of the obstacle removal trajectory.

[0123] Each of the following criteria is evaluated: recovery success, command executability, operational safety, and verification loop closure. The criteria are then assessed to determine if they meet their respective preset quality conditions. If any of these criteria fails to meet the preset quality condition, the candidate troubleshooting trajectory is removed. For the remaining candidate troubleshooting trajectories, their absolute evaluation results are compared with preset evaluation thresholds. Candidate troubleshooting trajectories whose absolute evaluation results reach the preset evaluation thresholds are designated as target troubleshooting trajectories.

[0124] By adopting the above scheme, and setting successful recovery, executable commands, safe operation, and closed-loop verification as quality screening conditions, and combining them with execution time, resource consumption, or scope of change for comprehensive evaluation, it is possible to avoid the problem of single advantages masking issues such as unrecovered faults, unexecutable commands, or unsafe operation, thereby improving the overall quality and reliability of the target troubleshooting trajectory.

[0125] Optionally, filtering candidate obstacle removal trajectories based on execution results also includes: For multiple candidate troubleshooting trajectories corresponding to the same network fault to be injected, the average level and dispersion within the group are determined based on the absolute evaluation results of each candidate troubleshooting trajectory. Based on the deviation of the absolute evaluation result of each candidate troubleshooting trajectory from the average level within the group, and combined with the dispersion within the group, the relative evaluation result within the group of each candidate troubleshooting trajectory is determined. From the candidate troubleshooting trajectories whose absolute evaluation results reach the preset evaluation threshold, the candidate troubleshooting trajectories whose relative evaluation results within the group meet the preset relative evaluation conditions are further screened as the final target troubleshooting trajectories.

[0126] Specifically, multiple candidate troubleshooting trajectories corresponding to the same network fault to be injected refer to candidate troubleshooting trajectories generated under the same network topology and the same fault injection configuration constraints. The intra-group average level characterizes the overall level of the absolute evaluation results of the candidate troubleshooting trajectories in that group; the intra-group dispersion characterizes the degree of fluctuation of the absolute evaluation results of that group relative to the intra-group average level. The intra-group relative evaluation result characterizes the quality of a candidate troubleshooting trajectory relative to other candidate troubleshooting trajectories in the same group.

[0127] In practice, based on network topology status data and fault injection configuration, candidate troubleshooting trajectories corresponding to the same network fault to be injected are grouped into the same trajectory group. The absolute evaluation results of each candidate troubleshooting trajectory in the trajectory group are obtained, and the average level and dispersion within the trajectory group are determined based on all absolute evaluation results.

[0128] For each candidate obstacle removal trajectory in the trajectory group, its absolute evaluation result is determined to be higher than, lower than, or close to the group average level, and its deviation from the group average level is also determined. The degree of dispersion within the group is used as the intra-group evaluation metric to measure this deviation, and the direction and degree of deviation are converted into a relative intra-group evaluation result for that candidate obstacle removal trajectory. Specifically, candidate obstacle removal trajectories with absolute evaluation results higher than the group average level receive relatively better relative intra-group evaluation results, while those with absolute evaluation results lower than the group average level receive relatively lower relative intra-group evaluation results.

[0129] After completing the relative evaluation within the group, from the candidate obstacle removal trajectories that have passed the preset quality conditions and whose absolute evaluation results have reached the preset evaluation threshold, it is further determined whether their relative evaluation results within the group meet the preset relative evaluation conditions, and the candidate obstacle removal trajectories that meet the preset relative evaluation conditions are determined as the final target obstacle removal trajectories.

[0130] By adopting the above scheme, and further comparing the candidate troubleshooting trajectories corresponding to the same network fault within the same group on the basis of absolute evaluation, it is possible to ensure that the candidate troubleshooting trajectories meet the basic quality requirements, while further retaining the relatively better troubleshooting trajectories in the same group, thereby reducing the impact of different network fault difficulties and different network topology states on the fixed evaluation criteria.

[0131] Optionally, the target troubleshooting trajectory is associated with the corresponding network topology status data, fault injection configuration, and execution results to generate structured trajectory samples for training the network operation and maintenance model, including: Generate sample identifiers for each target troubleshooting trajectory and obtain the topology identifiers of the corresponding network topology status data; extract fault types from the fault injection configuration and generate troubleshooting tasks to describe the network faults to be investigated based on the fault injection configuration; extract diagnostic information and action sequences from the target troubleshooting trajectory and obtain the absolute evaluation results and intra-group relative evaluation results corresponding to the target troubleshooting trajectory; generate sample labels based on the execution results, absolute evaluation results, and intra-group relative evaluation results.

[0132] According to the preset field mapping relationship, the sample identifier, topology identifier, network topology status data, fault type, fault injection configuration, troubleshooting task, diagnostic information, action sequence, execution result and sample label are written into the corresponding sample fields to form a sample record corresponding to the target troubleshooting trajectory; each sample record is output in a structured data format of line-by-line recording to generate a structured trajectory sample.

[0133] Specifically, a sample identifier refers to the identification information used to uniquely or relatively uniquely identify a sample record; a topology identifier refers to the identification information used to identify the network topology status data corresponding to the target troubleshooting trajectory; and a troubleshooting task refers to task information generated based on the fault injection configuration, used to describe the network faults to be troubleshooted and the troubleshooting target.

[0134] Diagnostic information refers to the information used to determine the location and cause of network faults in the target troubleshooting trajectory; action sequence refers to the network operation and maintenance operations arranged in the execution order in the target troubleshooting trajectory.

[0135] In practice, a corresponding sample identifier is assigned to each target obstacle removal trajectory, and the topology identifier is read from the network topology status data corresponding to the target obstacle removal trajectory, so as to establish the correspondence between the target obstacle removal trajectory and the network topology status data through the sample identifier and the topology identifier.

[0136] The system analyzes the fault injection configuration corresponding to the target troubleshooting trajectory, extracts the fault type, and converts the network faults and troubleshooting objectives described in the fault injection configuration into troubleshooting tasks. It also analyzes the output of the target troubleshooting trajectory, extracting diagnostic information and the action sequence arranged in execution order. Simultaneously, it obtains the absolute evaluation result and the relative evaluation result within the group corresponding to the target troubleshooting trajectory.

[0137] Based on the fault recovery status, operational safety status, and repair verification status represented by the execution results, and combined with the absolute evaluation results and the relative evaluation results within the group, the quality and training purpose of the target obstacle removal trajectory are determined, and corresponding sample labels are generated.

[0138] A field mapping relationship between data content and sample fields is established in advance. According to this field mapping relationship, the sample identifier, topology identifier, network topology status data, fault type, fault injection configuration, troubleshooting task, diagnostic information, action sequence, execution result, and sample label are written into the corresponding sample fields to form a sample record corresponding to the target troubleshooting trajectory.

[0139] After completing the sample recording of each target obstacle clearance trajectory, each sample record is output as an independent line of structured data, and different sample records are separated by line breaks, generating a structured trajectory sample that can be read line by line by the network operation and maintenance model training program.

[0140] By adopting the above scheme, the correspondence between various types of data in the training samples can be maintained by associating network topology status, network faults, troubleshooting tasks, troubleshooting processes, actual execution results and evaluation information with the same sample record, thereby improving the integrity of the structured trajectory samples and making them easy to use for training network operation and maintenance models.

[0141] Optionally, the target troubleshooting trajectory is associated with the corresponding network topology status data, fault injection configuration, and execution results to generate structured trajectory samples for training the network operation and maintenance model, and further includes: The training purpose label of the sample record corresponding to the target obstacle removal trajectory, whose execution result represents successful fault recovery and operational safety meets the preset safety conditions, whose absolute evaluation result reaches the preset evaluation threshold, and whose relative evaluation result within the group meets the preset relative evaluation conditions, is set as the positive sample label, thus obtaining supervised fine-tuning positive samples.

[0142] From the candidate troubleshooting trajectories generated for the same network fault to be injected, the candidate troubleshooting trajectory that has at least one of the following characteristics is selected as the rejection trajectory: the absolute evaluation result does not reach the preset evaluation threshold, the relative evaluation result within the group is lower than the preset rejection threshold, and the execution result indicates that the fault has not been recovered.

[0143] Rejection sample records corresponding to rejection trajectories are generated according to preset field mapping relationships, and the training purpose labels of the rejection sample records are set as preference optimization rejection sample labels.

[0144] Based on the same topology identifier and fault injection configuration, supervised fine-tuning positive samples are used as preferred samples, and rejected samples are recorded as rejected samples to generate sample pairs for preference optimization training.

[0145] Specifically, supervised fine-tuning positive samples refer to samples whose trajectory quality meets preset requirements and can be used as the expected output to train the network operation and maintenance model. Rejected trajectories refer to candidate troubleshooting trajectories that correspond to the same network fault as the supervised fine-tuning positive samples, but whose trajectory quality is relatively low or fails to recover the network fault. Preferred samples refer to samples that the network operation and maintenance model is expected to generate preferentially during preference optimization training; rejected samples refer to samples that the network operation and maintenance model is not expected to generate preferentially during preference optimization training. Sample pairs refer to preferred and rejected samples that are associated under the same network topology and the same network fault conditions.

[0146] In practice, the execution results, absolute evaluation results, and intra-group relative evaluation results are read from the sample records corresponding to the target obstacle removal trajectory. The system determines whether the execution results indicate successful fault recovery, whether the operational safety meets preset safety conditions, whether the absolute evaluation results reach preset evaluation thresholds, and whether the intra-group relative evaluation results meet preset relative evaluation conditions. For target obstacle removal trajectories that simultaneously meet the above conditions, the training purpose label in their corresponding sample records is set to a positive sample label, and the set sample records are designated as supervised fine-tuning positive samples.

[0147] Furthermore, from the candidate troubleshooting trajectories generated for the same network fault to be injected, it is determined whether any of the candidate troubleshooting trajectories have an absolute evaluation result that does not reach a preset evaluation threshold, a relative evaluation result within the group that is lower than a preset rejection threshold, or an execution result indicating that the fault has not been recovered. If at least one of the above conditions exists, the corresponding candidate troubleshooting trajectory is determined as a rejection trajectory.

[0148] Furthermore, according to the same preset field mapping relationship as the target obstacle removal trajectory, the network topology status data, fault injection configuration, diagnostic information, action sequence, execution result and evaluation information corresponding to the rejection trajectory are written into the corresponding sample fields to form a rejection sample record, and the training purpose label in the rejection sample record is set as the preference optimization rejection sample label.

[0149] Furthermore, supervised fine-tuning positive samples and rejection sample records are matched based on topology identifiers and fault injection configurations. Supervised fine-tuning positive samples with the same topology identifier and the same fault injection configuration are designated as preferred samples, and their corresponding rejection sample records are designated as rejection samples. These two are then associated to form sample pairs for preference optimization training. Where one preferred sample corresponds to multiple rejection samples, multiple sample pairs can be established separately.

[0150] By adopting the above scheme, supervised fine-tuning positive samples are generated from trajectories that meet the requirements of both actual execution results and evaluation results. Under the same network topology and fault conditions, preference optimization sample pairs of preferred and rejected samples are constructed. This enables the network operation and maintenance model to learn correct, effective and safe troubleshooting trajectories, while identifying troubleshooting trajectories with relatively low quality or unrecoverable faults. This improves the accuracy and safety of the troubleshooting trajectories generated by the network operation and maintenance model.

[0151] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0152] Based on the same inventive concept, this application also provides a network operation and maintenance trajectory sample generation system. This system is applicable to the above-mentioned network operation and maintenance trajectory sample generation method. The solution provided by this system is similar to the solution described in the above-mentioned method. Therefore, the specific limitations of one or more system embodiments provided below can be found in the limitations of the method above, and will not be repeated here.

[0153] Please see Figure 6 In one embodiment, the network operation and maintenance trajectory sample generation system includes: an acquisition module, a first construction module, a second construction module, an execution module, and a sample generation module.

[0154] The acquisition module is used to acquire network topology status data of the target network system.

[0155] The first construction module is used to construct a network topology state machine based on network topology state data; based on the network topology state machine, it generates a fault injection configuration and multiple candidate troubleshooting trajectories to characterize the network faults to be injected through an autoregressive language model.

[0156] The second building module is used to construct a dynamic isolated network sandbox environment that corresponds one-to-one with each candidate obstacle removal trajectory based on the network topology state machine.

[0157] The execution module is used to configure each dynamic isolation network sandbox environment to the corresponding network fault state according to the fault injection configuration, and to execute the corresponding candidate troubleshooting trajectory in each configured dynamic isolation network sandbox environment to obtain the execution result.

[0158] The sample generation module is used to filter each candidate troubleshooting trajectory based on the execution results to obtain at least one target troubleshooting trajectory; and associate each target troubleshooting trajectory with the corresponding network topology status data, fault injection configuration and execution results to generate structured trajectory samples for training network operation and maintenance models.

[0159] The aforementioned network operation and maintenance trajectory sample generation system constructs a network topology state machine based on network topology state data, providing a unified topology basis for fault and troubleshooting trajectory generation, thus improving the matching degree between the generated content and the target network system. Based on the network topology state machine, it generates fault injection configurations and multiple candidate troubleshooting trajectories, improving trajectory generation efficiency and the richness of troubleshooting solutions. It constructs a one-to-one dynamically isolated network sandbox environment for each candidate troubleshooting trajectory, and executes the candidate troubleshooting trajectories after injecting the corresponding network fault, avoiding mutual interference between trajectories and obtaining actual execution results. It filters target troubleshooting trajectories based on execution results, eliminating trajectories that are unexecutable or cannot effectively recover from faults. It associates the target troubleshooting trajectory with network topology state data, fault injection configurations, and execution results to form structured trajectory samples, improving the integrity and traceability of the samples. Therefore, this application realizes a closed loop of network operation and maintenance trajectory generation, verification, filtering, and structured processing, improving the quality, executability, verifiability, and training applicability of network operation and maintenance trajectory samples.

[0160] Each module in the aforementioned network operation and maintenance trajectory sample generation system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0161] In one feasible embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for generating network operation and maintenance trajectory samples.

[0162] In one feasible embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method for generating network operation and maintenance trajectory samples.

[0163] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0164] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating network operation and maintenance trajectory samples, characterized in that, The method includes: Obtain network topology state data of the target network system, and construct a network topology state machine based on the network topology state data; Based on the network topology state machine, an autoregressive language model is used to generate fault injection configurations and multiple candidate troubleshooting trajectories to characterize the network faults to be injected. Based on the network topology state machine, a dynamic isolated network sandbox environment is constructed that corresponds one-to-one with each of the candidate obstacle removal trajectories. According to the fault injection configuration, each of the dynamic isolation network sandbox environments is configured to the corresponding network fault state, and the corresponding candidate troubleshooting trajectory is executed in each of the configured dynamic isolation network sandbox environments to obtain the execution result; Based on the execution results, each candidate obstacle removal trajectory is filtered to obtain at least one target obstacle removal trajectory; The target troubleshooting trajectories are associated with the corresponding network topology status data, fault injection configurations, and execution results to generate structured trajectory samples for training network operation and maintenance models.

2. The method according to claim 1, characterized in that, The network topology status data includes at least one of the following: node set, link set, route set, service policy set, observability configuration, and security boundary configuration. The step of constructing a network topology state machine based on the network topology state data includes: The network topology state data is parsed and structured to generate at least one type of state information that characterizes network node state, link connection relationship, routing and forwarding relationship, service operation constraints, observability information, and security control boundary. Based on the state information, network nodes, links, and service relationships are associated and modeled to generate a network topology state machine that represents the operating state of the target network system.

3. The method according to claim 1, characterized in that, The step of generating fault injection configurations and multiple candidate troubleshooting trajectories to characterize the network faults to be injected, based on the network topology state machine and using an autoregressive language model, includes: A first conditional generation context is constructed based on the network topology state machine, and without inputting an artificial fault seed to specify a specific network fault, the autoregressive language model performs autoregressive decoding using a probability sampling strategy under the constraints of the first conditional generation context to generate the fault injection configuration. The network topology state machine is conditionally bound to the fault injection configuration to construct a second condition generation context. Under the constraints of the second condition generation context, the autoregressive language model performs multiple independent autoregressive decodings to obtain multiple candidate troubleshooting trajectories for the network fault represented by the fault injection configuration.

4. The method according to claim 3, characterized in that, The step of constructing a first condition generation context based on the network topology state machine includes: The network topology state machine is serialized to generate a topology context sequence, which is used to represent at least one of network node state, link state, routing state and service state. Construct system prompt words; the system prompt words are used to limit the type range of network faults to be injected, the structured output format of fault injection configuration, the structured output format of candidate troubleshooting trajectories, and the network operation and maintenance operations that are allowed to be executed; The topology context sequence and the system prompt words are combined in a preset order, and a first structured identifier for identifying the fault injection configuration output position is inserted at the end of the combination result to form the first condition generation context.

5. The method according to claim 4, characterized in that, The fault injection configuration includes a fault injection object; the network topology state machine includes network nodes, links, and services, each with a network entity identifier. The step of conditionally binding the network topology state machine with the fault injection configuration to construct a second condition generation context includes: Based on the fault injection object, a target network entity with a matching network entity identifier is determined from the network topology state machine. The target network entity includes at least one of a target network node, a target link, and a target service. An association relationship is established between the fault injection configuration and the determined target network entity. The association, the fault injection configuration, the topology context sequence, and the system prompt words are combined in a preset order, and a second structured identifier for identifying the candidate obstacle removal trajectory output position is inserted at the end of the combination result to form the second condition generation context.

6. The method according to claim 5, characterized in that, The step of performing multiple independent autoregressive decodings on the autoregressive language model under the constraints of the second conditional generation context to obtain multiple candidate troubleshooting trajectories for the network fault represented by the fault injection configuration includes: Based on the complexity of the network fault represented by the fault injection configuration, the number of autoregressive decoding operations K is determined, where K is an integer greater than 1; While keeping the association, fault injection configuration, topology context sequence and system prompt words unchanged in the second condition generation context, different decoding control parameters are set for K autoregressive decodings. The decoding control parameters include at least one of sampling temperature and decoding random parameters. The second condition is generated by inputting the context into the autoregressive language model, and K independent autoregressive decodings are performed according to the corresponding decoding control parameters. A candidate obstacle removal trajectory is generated in each autoregressive decoding process. By aggregating the candidate troubleshooting trajectories obtained from K autoregressive decodings, multiple candidate troubleshooting trajectories are obtained for the network faults represented by the fault injection configuration.

7. The method according to claim 6, characterized in that, The process of generating a candidate obstacle-clearing trajectory in each autoregressive decoding process includes: Based on the network topology state machine and the fault injection configuration, a diagnostic command for locating the network fault to be injected and the expected observation results corresponding to the diagnostic command are generated. Generate cause analysis information based on the diagnostic command and the expected observation results; Based on the cause analysis information, generate a fault repair action to eliminate the fault in the network to be injected; Based on the fault repair action, a verification action is generated to verify whether the fault in the network to be injected has been recovered; End this autoregressive decoding and take the output sequence generated by this autoregressive decoding, which includes the diagnostic command, the expected observation result, the cause analysis information, the fault repair action and the verification action, as a candidate troubleshooting trajectory.

8. The method according to claim 1, characterized in that, The step of constructing a dynamic isolated network sandbox environment corresponding one-to-one with each of the candidate obstacle-clearing trajectories based on the network topology state machine includes: For each of the aforementioned candidate obstacle removal trajectories, create a separate, isolated sandbox instance; In each sandbox instance, a container node is created according to the network topology state machine, a virtual network interface card, a network address and a routing table are configured for the container node, and the service process represented by the network topology state machine is started in the container node. In each sandbox instance, the service network, transmission network, control network, and observation network are configured according to the network topology state machine, and each container node is connected to the corresponding network according to the network connection relationship represented by the network topology state machine. Each sandbox instance that has completed the configuration of the container node and the corresponding network is initialized to the same initial topology state represented by the network topology state machine, so as to obtain a dynamic isolated network sandbox environment that corresponds one-to-one with each candidate obstacle removal trajectory and is mutually isolated.

9. The method according to claim 1, characterized in that, The fault injection configuration includes fault types and fault parameters. The fault types include at least one of the following: link faults, routing faults, service faults, resource faults, and security policy faults. The step of configuring each of the dynamically isolated network sandbox environments to a corresponding network fault state according to the fault injection configuration includes: Different injection methods are determined based on the fault type, and each of the dynamic isolated network sandbox environments is configured to the corresponding network fault state using the fault injection method according to the fault parameters. Wherein, when the fault type includes link-related faults, the injection method includes: injecting at least one of packet loss, latency, jitter, or bandwidth limitation into the target virtual link through a network traffic control tool, or simulating link interruption by disabling the target virtual network interface card or deleting the target route; When the fault type includes routing faults, the injection method includes at least one of: modifying route priority, adding an erroneous next hop, and creating a routing black hole; When the fault type includes service-related faults, the injection method includes at least one of the following: stopping the target service through a process management tool, restarting the target service, and loading an abnormal service configuration into the target service; When the fault type includes resource-related faults, the injection method includes: reducing the processor usage quota or memory usage limit of the target container through a resource control component or a container resource limiting component; or, applying processor computing load, memory usage load, or network connection load to the target container through a load generation component. When the fault type includes security policy-related faults, the injection method includes: adding at least one of access blocking rules, port blocking rules, and error forwarding rules through the firewall rule management component.

10. The method according to claim 1, characterized in that, The step of executing the corresponding candidate obstacle-clearing trajectory in each of the configured dynamic isolation network sandbox environments to obtain the execution result includes: Perform security checks on the commands to be executed in the candidate troubleshooting trajectory, and execute the commands that have passed the security checks according to the candidate troubleshooting trajectory using a non-privileged user or a user granted the minimum operating privileges, and record the command execution information; During execution, monitoring probes are used to collect operational status data of the corresponding dynamic isolated network sandbox environment, and execution is terminated when the execution time exceeds the preset time. Based on the command execution information and the changes in the running status data before and after execution, execution results are generated to characterize the command execution status, fault recovery status, execution time, resource consumption status, operational safety status, and repair verification status. After completing the collection of the operational status data, the corresponding dynamic isolated network sandbox environment is destroyed or rolled back.

11. The method according to claim 1, characterized in that, The step of filtering each candidate obstacle removal trajectory based on the execution result includes: Based on the execution results corresponding to each candidate troubleshooting trajectory, indicator parameters are determined. The indicator parameters include at least one of the following: execution time indicator, resource consumption indicator, minimum change indicator, as well as recovery success indicator, command executability indicator, operation safety indicator, and verification closed loop indicator. Based on the preset weights corresponding to each indicator parameter, a comprehensive evaluation of each indicator parameter is performed to obtain the absolute evaluation result of each candidate obstacle removal trajectory. Candidate troubleshooting trajectories that do not meet at least one of the following preset quality conditions: recovery success index, command executability index, operational safety index, and verification closed-loop index; From the remaining candidate obstacle removal trajectories, the candidate obstacle removal trajectories whose absolute evaluation results reach the preset evaluation threshold are selected as the target obstacle removal trajectories.

12. The method according to claim 11, characterized in that, The step of filtering each candidate obstacle removal trajectory based on the execution result also includes: For multiple candidate troubleshooting trajectories corresponding to the same network fault to be injected, the average level and dispersion within the group are determined based on the absolute evaluation results of each candidate troubleshooting trajectory. Based on the degree of deviation of the absolute evaluation result of each candidate obstacle removal trajectory from the average level within the group, and in combination with the degree of dispersion within the group, the relative evaluation result within the group for each candidate obstacle removal trajectory is determined. From the candidate obstacle removal trajectories whose absolute evaluation results reach the preset evaluation threshold, further filter the candidate obstacle removal trajectories whose relative evaluation results within the group meet the preset relative evaluation conditions, and use them as the final target obstacle removal trajectories.

13. The method according to claim 12, characterized in that, The step of associating the target troubleshooting trajectory with the corresponding network topology status data, fault injection configuration, and execution results to generate structured trajectory samples for training the network operation and maintenance model includes: Generate sample identifiers for each of the target obstacle removal trajectories, and obtain the topology identifiers of the corresponding network topology status data; Extract the fault type from the fault injection configuration, and generate a troubleshooting task to describe the network fault to be investigated based on the fault injection configuration; Diagnostic information and action sequences are extracted from the target obstacle removal trajectory, and the absolute evaluation results and intra-group relative evaluation results corresponding to the target obstacle removal trajectory are obtained. Generate sample labels based on the execution results, absolute evaluation results, and within-group relative evaluation results; According to the preset field mapping relationship, the sample identifier, topology identifier, network topology status data, fault type, fault injection configuration, troubleshooting task, diagnostic information, action sequence, execution result and sample label are written into the corresponding sample fields to form a sample record corresponding to the target troubleshooting trajectory. Each of the sample records is output in a structured data format that records line by line to generate the structured trajectory sample.

14. The method according to claim 13, characterized in that, The step of associating the target obstacle removal trajectory with the corresponding network topology status data, fault injection configuration, and execution results to generate structured trajectory samples for training network operation and maintenance models also includes: The training purpose label of the sample record corresponding to the target obstacle removal trajectory whose execution result represents successful fault recovery and operational safety meets the preset safety conditions, whose absolute evaluation result reaches the preset evaluation threshold, and whose relative evaluation result within the group meets the preset relative evaluation conditions is set as a positive sample label to obtain supervised fine-tuning positive samples; From the candidate troubleshooting trajectories generated for the same network fault to be injected, select at least one of the following candidate troubleshooting trajectories as rejection trajectories: the absolute evaluation result does not reach the preset evaluation threshold, the relative evaluation result within the group is lower than the preset rejection threshold, and the execution result indicates that the fault has not been recovered. Generate rejection sample records corresponding to the rejection trajectory according to the preset field mapping relationship, and set the training purpose label of the rejection sample records as the preference optimization rejection sample label; Based on the same topology identifier and fault injection configuration, the supervised fine-tuning positive samples are used as preferred samples, and the rejected samples are recorded as rejected samples to generate sample pairs for preference optimization training.

15. A system for generating network operation and maintenance trajectory samples, characterized in that, The system includes: The acquisition module is used to acquire network topology status data of the target network system; The first construction module is used to construct a network topology state machine based on the network topology state data; and based on the network topology state machine, to generate a fault injection configuration and multiple candidate troubleshooting trajectories to characterize the network fault to be injected through an autoregressive language model. The second construction module is used to construct a dynamic isolated network sandbox environment that corresponds one-to-one with each of the candidate obstacle removal trajectories based on the network topology state machine. The execution module is used to configure each of the dynamic isolation network sandbox environments to the corresponding network fault state according to the fault injection configuration, and to execute the corresponding candidate troubleshooting trajectory in each of the configured dynamic isolation network sandbox environments to obtain the execution result; The sample generation module is used to filter each candidate troubleshooting trajectory according to the execution result to obtain at least one target troubleshooting trajectory; and associate each target troubleshooting trajectory with the corresponding network topology status data, fault injection configuration and execution result to generate a structured trajectory sample for training the network operation and maintenance model.

16. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1-14.

17. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-14.

Images