A Method and System for Constructing a High-Speed ​​Operation and Maintenance Knowledge Base Based on Generative AI

By binding the runtime status snapshot and target status fingerprint of operation and maintenance evidence fragments to the operation and maintenance knowledge base construction method of generative AI, inconsistent status evidence is isolated, and a fault causal closed-loop template is constructed. This solves the problem of inconsistent status of evidence fragments in the prior art and realizes the consistency of status and traceability of source of knowledge items.

CN123115294APending Publication Date: 2026-11-17ZHONGNAN TRANSPORT
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611235949.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for building operation and maintenance knowledge bases based on RAG and generative AI have several drawbacks. The recalled evidence fragments are not bound to the snapshot of the running state at the time of generation. Evidence from old versions, old topologies, old configurations, or old alarm rules can easily be mixed into the current knowledge entry generation process. Compressed evidence lacks causal loops and source tracing constraints between fault phenomena, root cause evidence, handling actions, and recovery verification.

Method used

By acquiring operational evidence fragments from RAG recall, binding and covering the corresponding service object's operational status snapshot according to event time, forming an evidence status fingerprint, and comparing it field by field with the target status fingerprint, isolating inconsistent status evidence, generating a fusionable evidence set, compressing the evidence set according to the fault causal closed loop template composed of fault phenomena, root cause evidence, handling actions, and recovery verification, and inputting it into a generative AI model for state constraint generation and field backtracking verification.

Benefits of technology

It improves the consistency of knowledge entries' status, closed-loop integrity, and traceability of sources, avoids the mixing of evidence from old versions, old topologies, old configurations, or old alarm rules, and ensures that the generated knowledge entries are applicable to the target operating environment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115294A_ABST
    Figure CN123115294A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for constructing a high-speed operation and maintenance knowledge base based on generative AI, relating to the field of intelligent operation and maintenance knowledge base construction technology. The method includes: acquiring operation and maintenance evidence fragments recalled by RAG; binding and covering the corresponding service object's operational status snapshot according to event time; reading state constraint elements from the operational status snapshot to form an evidence state fingerprint; extracting similar state constraint elements to form a target state fingerprint based on the target object and target time window of the knowledge entry to be constructed; comparing the evidence state fingerprint with the target state fingerprint field by field, isolating inconsistent state evidence and generating a fusionable evidence set; compressing the fusionable evidence set according to a fault causal closed-loop template composed of fault phenomena, root cause evidence, handling actions, and recovery verification; and inputting the compressed evidence package into a generative AI model to perform state constraint generation, field backtracking verification, and indexing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance knowledge base construction technology, specifically to a method and system for constructing a high-speed operation and maintenance knowledge base based on generative AI. Background Technology

[0002] With the development of cloud-native technologies, microservices, container orchestration, and distributed tracing, the sources of fault evidence in high-speed operation and maintenance systems have gradually expanded from single logs to multi-source data such as logs, alarms, work orders, configuration changes, release records, tracing data, and debriefing documents. To improve the efficiency of operation and maintenance knowledge accumulation, related technologies have begun to introduce RAG retrieval enhancement generation and generative AI models. These methods retrieve similar operation and maintenance fragments through vector retrieval, and then the generative AI model generates fault knowledge, handling suggestions, or debriefing summaries. While these methods can improve the utilization rate of unstructured operation and maintenance data, their processing focus is usually concentrated on semantic retrieval, text summarization, and knowledge item generation, with insufficient attention paid to the runtime constraints at the time the evidence fragments were generated.

[0003] Existing methods for building RAG-based operation and maintenance knowledge bases typically treat recalled fragments as static knowledge data, relying mainly on text similarity, keyword matching, time sorting, or manual labeling to determine usable evidence. This makes it difficult to identify whether the runtime version, topology version, configuration version, resource status, and alarm rule version at the time the evidence fragment was generated are consistent with the target runtime state of the knowledge entry to be constructed. In high-speed operation and maintenance scenarios, the same fault phenomenon may correspond to different root causes and handling actions due to service version changes, topology adjustments, configuration updates, or alarm rule changes. Effective handling experience in the old state may become invalid in the current state. If the RAG recall results are not bound to a runtime state snapshot and constrained for state consistency before evidence compression, generative AI models are prone to mixing evidence from old topologies, old configurations, or old alarm rules into the compression context, resulting in knowledge entries with unclear applicable states, untraceable sources, or inconsistencies with the current operation and maintenance environment. Furthermore, existing summary compression methods often prioritize text length or semantic coverage, lacking causal closed-loop constraints between fault phenomena, root cause evidence, handling actions, and recovery verification. This results in generated results that, while possessing textual integrity, lack a verifiable chain of technical evidence. Therefore, there is an urgent need for a generative AI constraint method for building a high-speed operation and maintenance knowledge base, which can complete evidence state binding, state consistency gating and closed-loop evidence compression before knowledge generation, and complete field-level source backtracking verification after generation. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing methods for constructing operation and maintenance knowledge bases based on RAG and generative AI have the following problems: the recalled evidence fragments are not bound to the snapshot of the running state at the time of generation; evidence from old versions, old topologies, old configurations, or old alarm rules is easily mixed into the current knowledge entry generation process; compressed evidence lacks causal closed loops and source backtracking constraints between fault phenomena, root cause evidence, handling actions, and recovery verification; and how to enable generative AI to generate state-bound operation and maintenance knowledge entries only based on operation and maintenance evidence that meets the conditions of state consistency and source backtracking.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for constructing a high-speed operation and maintenance knowledge base based on generative AI, comprising: acquiring operation and maintenance evidence fragments recalled by RAG; binding and covering the corresponding service object's operation status snapshot according to event time; reading state constraint elements from the operation status snapshot to form evidence state fingerprints; extracting similar state constraint elements to form target state fingerprints based on the target object and target time window of the knowledge entry to be constructed; comparing the evidence state fingerprints and target state fingerprints field by field, isolating inconsistent state evidence and generating a fusionable evidence set; compressing the fusionable evidence set according to the fault causal closed-loop template composed of fault phenomena, root cause evidence, handling actions, and recovery verification; and inputting the compressed evidence package into a generative AI model to perform state constraint generation, field backtracking verification, and indexing into the database.

[0007] As a preferred embodiment of the high-speed operation and maintenance knowledge base construction method based on generative AI described in this invention, the step of binding the operation status snapshot of the corresponding service object according to the event time includes: reading the fragment source identifier and candidate time field of the operation and maintenance evidence fragment recalled by RAG; the candidate time field includes the event occurrence time, the handling execution time, and the recovery verification time; when multiple candidate time fields coexist, the candidate time field closest to the action predicate in the evidence fragment text is selected as the state binding time according to the order of fault phenomenon occurrence, handling action execution, and recovery verification completion; when only one candidate time field exists, the candidate time field is used as the state binding time; the state change record saved by the operation and maintenance platform is retrieved by the state binding time, and the operation status snapshot with the effective time not later than the state binding time and the next effective time later than the state binding time is selected as the operation status snapshot covering the corresponding service object; when the evidence fragment does not carry a clear service object, the fragment source identifier is mapped to the service object according to the fragment source identifier and the object alias table in the text, and then the operation status snapshot is retrieved according to the service object; after the binding is completed, the operation status snapshot identifier, the state binding time, and the object identifier are written into the metadata record of the operation and maintenance evidence fragment.

[0008] As a preferred embodiment of the high-speed operation and maintenance knowledge base construction method based on generative AI described in this invention, the method for forming an evidence state fingerprint includes: reading state constraint elements from a bound operation state snapshot, performing normalized encoding, resource state level conversion, and combining them in a preset order to form an evidence state fingerprint; the state constraint elements include an object identifier representing the object to which the operation and maintenance evidence fragment belongs, and operation version, topology version, configuration version, resource state level, and alarm rule version representing the operation conditions when the operation and maintenance evidence fragment was generated; establishing an association record between the evidence state fingerprint and the operation and maintenance evidence fragment identifier, source anchor point, state binding time, and operation state snapshot identifier; when necessary state constraint elements are missing, searching within the same operation state snapshot for the most recently effective state constraint element value of the same object identifier before the state binding time to fill in the missing elements, and marking the corresponding operation and maintenance evidence fragment as incomplete state evidence when the filling fails.

[0009] As a preferred embodiment of the high-speed operation and maintenance knowledge base construction method based on generative AI described in this invention, the step of extracting similar state constraint elements to form a target state fingerprint includes: reading the target object and target time window from the construction parameters of the knowledge entry to be constructed; the target time window includes the start and end times to which the knowledge entry is intended to be applied; retrieving operation state snapshots that have an overlay relationship with the target object within the target time window; when more than two operation state snapshots are retrieved, selecting the operation state snapshot with an effective time no later than the end time of the target time window and the latest effective time in the order as the target operation state snapshot; reading state constraint elements with the same name as the evidence state fingerprint from the target operation state snapshot, and processing them according to normalized encoding, resource state level conversion, and fixed order combination; writing the processed state constraint elements into the target state fingerprint, and establishing an association record between the target state fingerprint and the target object, target time window, and target operation state snapshot identifier; when the target object covers multi-level call relationships, forming multiple target state fingerprints according to the call direction, and recording the call sequence relationship among the multiple target state fingerprints.

[0010] As a preferred embodiment of the high-speed operation and maintenance knowledge base construction method based on generative AI described in this invention, the generation of a fusionable evidence set includes: comparing object identifiers in the evidence state fingerprint and the target state fingerprint; when object identifiers are inconsistent, writing the corresponding operation and maintenance evidence fragment into the isolated evidence list; when object identifiers are consistent, comparing fields in the order of running version, topology version, configuration version, resource status level, and alarm rule version to obtain the difference state field; when no difference state field exists, writing the corresponding operation and maintenance evidence fragment into the state-consistent evidence set; when a difference state field exists, reading the status change record corresponding to the difference state field, and extracting the difference state field name, the scope of the difference state field, and the object associated with the difference state field; and locating the fault phenomenon, root cause description, and handling in the operation and maintenance evidence fragment based on the source anchor point of the operation and maintenance evidence fragment. The evidence-bearing locations corresponding to actions and recovery verification are determined. The difference status field name, scope of effect, and associated object are matched with the object identifier and action object in the evidence-bearing location to form a difference status field influence chain. When the difference status field influence chain connects to any evidence-bearing location in the root cause description, action, or recovery verification, the corresponding maintenance evidence fragment is written to the isolated evidence list. When the difference status field influence chain does not connect to any evidence-bearing location in the root cause description, action, or recovery verification, the corresponding maintenance evidence fragment is written to the conditionally consistent evidence set, and the difference status field, its scope of effect, and applicable conditions are recorded. The state-consistent evidence set and the conditionally consistent evidence set are merged into a fusionable evidence set, and the source anchor point correspondence between the isolated evidence list and the fusionable evidence set is preserved.

[0011] As a preferred embodiment of the high-speed operation and maintenance knowledge base construction method based on generative AI described in this invention, the compressed and fusionable evidence set includes: establishing a fault causal closed-loop template containing fault phenomenon nodes, root cause evidence nodes, handling action nodes, and recovery verification nodes; writing the evidence fragments into the corresponding nodes in the fault causal closed-loop template according to the text tags, action predicates, and source anchors of each evidence fragment in the fusionable evidence set; when an evidence fragment covers the content of multiple nodes simultaneously, it is divided into multiple evidence fragments according to the node boundaries, and each fragment is written into only one node; and the source anchor is preferentially retained within each node. The system collects evidence status fingerprints, key action words, and key object words, and then merges evidence fragments with the same evidence status fingerprints, the same object identifiers, and the same action words. When evidence with consistent conditions enters the compression process, the difference status field and applicable conditions are written immediately after the corresponding compressed text. When any node in the root cause evidence node or the recovery verification node is empty, a supplementary evidence package is generated. When all nodes in the fault causal closed-loop template have been written with evidence fragments, a compressed evidence package is generated in the order of fault phenomenon node, root cause evidence node, handling action node, and recovery verification node, and the source anchor point correspondence between each text segment in the compressed evidence package and the original evidence fragment is preserved.

[0012] As a preferred embodiment of the high-speed operation and maintenance knowledge base construction method based on generative AI described in this invention, the following steps are included: generating execution state constraints, performing field backtracking verification, and indexing; constructing input prompts for the generative AI model based on compressed evidence packages; writing the input prompts into the target state fingerprint, the set of fusionable evidence, the list of isolated evidence, the fault causal closed-loop template, and knowledge item field constraints; limiting the root cause field to reference the root cause evidence node, the disposal field to reference the disposal action node, the recovery verification field to reference the recovery verification node, and the applicable condition field to reference the difference state field in the target state fingerprint and the condition consistency evidence; and the generative AI model is then processed according to... The system generates state-bound O&M knowledge entries based on input prompts. These entries include fault symptoms, root cause descriptions, handling actions, recovery verification methods, applicable state fingerprints, and source anchors. After generation, the system checks the compressed evidence package item by item according to the knowledge entry fields. If a knowledge field cannot be traced back to the corresponding node in the compressed evidence package, or if a knowledge field references an evidence fragment in the isolated evidence list, the corresponding knowledge field is deleted, and the generative AI model is re-triggered. When all necessary knowledge fields have been verified, the state-bound O&M knowledge entries are written into the high-speed O&M knowledge base, and a text vector index, a state fingerprint index, and a source anchor index are established simultaneously.

[0013] As a preferred embodiment of the high-speed operation and maintenance knowledge base construction system based on generative AI described in this invention, the system includes: a snapshot binding module, an evidence fingerprint module, a target fingerprint module, a state gating module, a closed-loop compression module, and a backtracking database module. The snapshot binding module is used to acquire operation and maintenance evidence fragments recalled by RAG and bind them to the corresponding service object's running state snapshot based on the event time. The evidence fingerprint module is used to read state constraint elements from the running state snapshot to form an evidence state fingerprint. The target fingerprint module is used to extract similar state constraint elements to form a target state fingerprint based on the target object and target time window of the knowledge entry to be constructed. The state gating module is used to compare the evidence state fingerprint and the target state fingerprint field by field, isolate inconsistent state evidence, and generate a fusionable evidence set. The closed-loop compression module is used to compress the fusionable evidence set according to the fault causal closed-loop template composed of fault phenomena, root cause evidence, handling actions, and recovery verification. The backtracking database module is used to input the compressed evidence package into the generative AI model and perform state constraint generation, field backtracking verification, and indexing database entry.

[0014] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program as a step in implementing a method for constructing a high-speed operation and maintenance knowledge base based on generative AI.

[0015] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for constructing a high-speed operation and maintenance knowledge base based on generative AI.

[0016] The beneficial effects of this invention are as follows: The high-speed operation and maintenance knowledge base construction method based on generative AI provided by this invention transforms RAG recall fragments from purely semantically related text into state-constrained evidence carrying the running version, topology version, configuration version, resource status level, and alarm rule version by constructing evidence state fingerprints and target state fingerprints. This allows for the determination of whether historical evidence is applicable to the target operating environment before knowledge entries are generated. Through the influence chain of difference state fields, field differences are further mapped to evidence-carrying locations such as root cause descriptions, handling actions, and recovery verifications, avoiding the simple deletion of all difference evidence and reducing the entry of inconsistent evidence of key states into the generation context. Through fault causal closed-loop compression and field-level source backtracking, the compressed evidence package retains the closed-loop structure between fault phenomena, root cause evidence, handling actions, and recovery verification even when length is limited, and requires generated fields to be traced back to permitted source anchors, thereby improving the state consistency, closed-loop integrity, and source traceability of knowledge entries. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The overall flowchart of the high-speed operation and maintenance knowledge base construction method based on generative AI provided by the present invention.

[0019] Figure 2 Evidence fusion diagram for the high-speed operation and maintenance knowledge base construction method based on generative AI provided by this invention.

[0020] Figure 3 A schematic diagram of a computer device provided by the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Reference Figures 1-2 As an embodiment of the present invention, a method for constructing a high-speed operation and maintenance knowledge base based on generative AI is provided, including:

[0023] S1: Obtain the operation and maintenance evidence fragments recalled by RAG, and bind and overwrite the corresponding service object's running status snapshot based on the event time.

[0024] Furthermore, the process of binding the runtime status snapshot of the corresponding service object based on the event time includes: reading the fragment source identifier and candidate time field of the RAG-recalled operation and maintenance evidence fragment; the candidate time field includes the event occurrence time, the handling execution time, and the recovery verification time; when multiple candidate time fields coexist, the candidate time field closest to the action predicate in the evidence fragment text is selected as the state binding time according to the order of fault occurrence, handling action execution, and recovery verification completion; when only one candidate time field exists, the candidate time field is used as the state binding time; the state change records saved by the operation and maintenance platform are retrieved based on the state binding time, and the runtime status snapshot with an effective time no later than the state binding time and the next effective time later than the state binding time is selected as the runtime status snapshot covering the corresponding service object; when the evidence fragment does not carry a specific service object, the fragment source identifier is mapped to the service object according to the fragment source identifier and the object alias table in the text, and then the runtime status snapshot is retrieved according to the service object; after the binding is completed, the runtime status snapshot identifier, the state binding time, and the object identifier are written into the metadata record of the operation and maintenance evidence fragment.

[0025] It should be noted that one scheme for binding the runtime status snapshot of the corresponding service object based on event time specifically includes, in this embodiment, the operational evidence fragments recalled by RAG refer to text fragments or structured fragments recalled from logs, alarms, work orders, fault replays, handling records, link tracing summaries, and configuration change records by the retrieval enhancement generation process. Each operational evidence fragment includes at least an operational evidence fragment identifier, a fragment source identifier, and the main text content. The fragment source identifier indicates the original record of the operational evidence fragment, and the main text content carries at least one of the following information: fault phenomenon, root cause description, handling action, and recovery verification.

[0026] The text fragments recalled by RAGs typically only have semantic relevance and cannot directly indicate the operational state in which the corresponding evidence was generated. In high-speed operation and maintenance scenarios, the running version, topology version, configuration version, resource status, and alarm rule version of a service object may change in a short period of time. Therefore, this embodiment first binds each operation and maintenance evidence fragment to a snapshot of the corresponding service object's operational state at the time the evidence was generated.

[0027] A runtime status snapshot is a structured record saved by the operations and maintenance platform when a status change takes effect. It includes at least the runtime status snapshot identifier, effective time, next effective time, object identifier, runtime version, topology version, configuration version, resource status, and alarm rule version. The effective time indicates when the corresponding runtime status snapshot begins to cover the service object, the next effective time indicates when the next runtime status snapshot begins to cover the same service object, and the object identifier is a unique identifier for the service object in the operations and maintenance platform.

[0028] For each operational evidence fragment, the fragment source identifier and candidate time field are read. The candidate time field includes the event occurrence time, the handling execution time, and the recovery verification time. The event occurrence time indicates the time when the fault phenomenon or alarm was triggered; the handling execution time indicates the time when a restart, rollback, expansion, configuration modification, or traffic switching was performed; and the recovery verification time indicates the time when the metric was restored, the alarm was cleared, or the work order was closed. If the candidate time field comes from a structured record, the field value is read directly; if the candidate time field is located in the main text, the time string in the main text is extracted using time expression recognition rules and uniformly converted into a timestamp.

[0029] When an operational evidence fragment has only one candidate time field, that candidate time field is used as the state-bound time. When an operational evidence fragment has two or more candidate time fields, the candidate time fields are categorized according to the chronological order of fault occurrence, action execution, and recovery verification completion. The candidate time field with the smallest distance to the action predicate text in the evidence fragment is selected as the state-bound time. Action predicates include trigger, escalation, timeout, failure, and exception indicating fault occurrence; restart, rollback, expansion, switchover, modification, and offline indicating action execution; and recovery, pass, decrease, elimination, and shutdown indicating recovery verification. The state-bound time satisfies:

[0030]

[0031] in, This represents the index of operational evidence fragments. Indicates the first The index set of candidate time fields in each piece of operational evidence. Indicates the first The selected segment from the maintenance evidence fragments The timestamps corresponding to the candidate time fields. Indicates the first The first piece of maintenance evidence fragment The word segmentation positions of the candidate time field in the main text content. Indicates the relationship with the first The word segmentation position in the main text of each operation and maintenance evidence fragment that matches the content of the action predicate. Indicates the distance between two locations. This indicates the index used to iterate over the candidate time field. Indicates the first Index of the selected candidate time field in each piece of operational evidence. Indicates the first The status binding time of each maintenance evidence fragment.

[0032] If two or more candidate time fields are at the same distance from the action predicate, they are selected in the order of priority: disposal execution time, event occurrence time, and recovery verification time.

[0033] After obtaining the status binding time, the status change records stored in the operations and maintenance platform are retrieved based on the status binding time. For operational evidence fragments where the service object is clearly identified, the object identifier and status binding time are used directly as search criteria. The operating status snapshot with an effective time no later than the status binding time and a next effective time later than the status binding time is selected as the operating status snapshot covering the corresponding service object. When there is no next effective time for the operating status snapshot, the corresponding operating status snapshot is considered to be valid continuously from the effective time until the current search time or the end of the access period. Therefore, even when the status binding time is later than the effective time, the corresponding operating status snapshot can still be selected. The selection of the operating status snapshot satisfies the following conditions:

[0034]

[0035] in, Indicates the relationship with the first A snapshot of the running status bound to a piece of operational evidence. Indicates the first in the operation and maintenance platform A snapshot of the running status. Represented as the first The selected runtime status snapshot index of the operation and maintenance evidence fragment. This represents the traversal index of the runtime snapshot. This represents a set of runtime status snapshot indexes that meet the conditions of service object and time interval. Indicates the first The object identifier covered by each runtime state snapshot Indicates the first The object identifier corresponding to each piece of maintenance evidence. Indicates the first The effective time of each runtime status snapshot. Indicates the first The next effective time for each runtime state snapshot. For the current runtime state snapshot that does not have a next effective time, Retrieve the current search time or the end time of the open period.

[0036] When an operational evidence fragment does not carry a clearly defined service object, the source object name is extracted from the fragment source identifier, and candidate object names are extracted from the text content. Candidate object names include the Chinese service name, English service name, application abbreviation, instance name prefix, interface name, and call chain node name. The object alias table includes an alias field and an object identifier field. The source object name and candidate object names are matched against the alias field in the object alias table. If a match is found, the object identifier from the same record is read as the object identifier for the corresponding operational evidence fragment. When two or more object identifiers are matched, the object identifier is selected in the order of fragment source identifier match, text title match, and text content match. If two or more object identifiers still exist, the object identifier that appears most frequently is selected. If a unique identifier cannot be determined, the corresponding operational evidence fragment is marked as evidence with an undetermined object, and the runtime snapshot binding is not performed. For other operational evidence fragments with determined object identifiers, runtime snapshots are retrieved according to the object identifier.

[0037] The object alias table stores the mapping between the Chinese service name, English service name, application abbreviation, instance name prefix, interface name, and call chain node name and the object identifier. After binding is completed, the runtime status snapshot identifier, status binding time, and object identifier are written to the metadata record of the operation and maintenance evidence fragment.

[0038] S2: Read state constraint elements from the running state snapshot to form evidence state fingerprints.

[0039] Furthermore, forming an evidence status fingerprint involves reading status constraint elements from bound operational status snapshots, performing normalized encoding, resource status level conversion, and combining them in a preset order to form an evidence status fingerprint. The status constraint elements include an object identifier representing the object to which the operational evidence fragment belongs, and operational version, topology version, configuration version, resource status level, and alarm rule version representing the operational conditions at the time the operational evidence fragment was generated. An association record is established between the evidence status fingerprint and the operational evidence fragment identifier, source anchor point, status binding time, and operational status snapshot identifier. When necessary status constraint elements are missing, the most recently effective status constraint element value for the same object identifier before the status binding time is retrieved within the same operational status snapshot to complete the missing elements. If completion fails, the corresponding operational evidence fragment is marked as incomplete status evidence.

[0040] It should be noted that one specific method for forming an evidence state fingerprint includes, after the operational state snapshot has been bound to the operational evidence fragment, reading state constraint elements from the bound operational state snapshot and forming the evidence state fingerprint. The evidence state fingerprint describes the operational state conditions at the time the corresponding operational evidence fragment was generated. Although the operational evidence fragment itself is text, whether the evidence can be used to construct the current knowledge entry depends on whether the corresponding operational environment behind the text is consistent. Therefore, this embodiment converts the key operational states in the operational state snapshot into comparable evidence state fingerprints.

[0041] The state constraint elements include the object identifier representing the object to which the operation and maintenance evidence fragment belongs, and the runtime version, topology version, configuration version, resource status level, and alarm rule version representing the operating conditions at the time the operation and maintenance evidence fragment was generated. The object identifier comes from the object code in the service registration, configuration center, or call chain record; the runtime version comes from the software version corresponding to the service object in the release platform; the topology version comes from the call relationship version recorded in the service governance platform; the configuration version comes from the configuration release version bound to the service object in the configuration center; and the alarm rule version comes from the rule version corresponding to the service object or alarm object in the alarm platform.

[0042] Resource status includes one or more of the following: CPU utilization, memory utilization, disk utilization, and connection pool occupancy. To avoid oversensitivity caused by directly using continuous values ​​for resource status in fingerprint comparison, this embodiment converts resource status into resource status levels. Preferably, resource status levels are set to three levels: low load, medium load, and high load. Resource utilization less than 60% is recorded as low load, resource utilization greater than or equal to 60% but less than 85% is recorded as medium load, and resource utilization greater than or equal to 85% is recorded as high load. When a runtime snapshot contains more than two resource status indicators, each resource status indicator is converted into a resource status level, and the highest resource status level is selected as the resource status level of the corresponding runtime snapshot.

[0043] For the A fragment of operational evidence, with its status fingerprint represented as follows:

[0044]

[0045] in, Indicates the first Evidence status fingerprint of a single maintenance evidence fragment Indicates the first The object identifier corresponding to each piece of maintenance evidence. Indicates the first The running version corresponding to each piece of maintenance evidence. Indicates the first The topology version corresponding to each piece of operational evidence. Indicates the first The configuration version corresponding to each piece of maintenance evidence. Indicates the first The resource status level corresponding to each piece of operational evidence. Indicates the first Each operational evidence fragment corresponds to an alarm rule version. Status constraint elements are combined in a preset order: object identifier, running version, topology version, configuration version, resource status level, and alarm rule version.

[0046] Before forming the evidence status fingerprint, the status constraint elements are normalized and encoded. Normalization encoding includes object identifier normalization, version number normalization, topology version normalization, configuration version normalization, resource status level conversion, and alarm rule version normalization. During object identifier normalization, the name of the object to be normalized is matched against the alias field. If a match is found, the object identifier in the same record is read as the normalized object identifier. When two or more object identifiers are matched, the object identifier that matches the object identifier in the current running status snapshot is selected first. If a unique identifier still cannot be determined, the corresponding operation and maintenance evidence fragment is marked as undetermined object evidence. Undetermined object evidence does not enter the evidence status fingerprint formation process and is written into the isolated evidence list.

[0047] Version number normalization converts different version notation into a standard version number; topology version, configuration version, and alarm rule version adopt the version number in the corresponding operation and maintenance platform as the standard expression. After normalization, the state constraint elements are processed by removing spaces, unifying capitalization, and standardizing version numbers before being written into the evidence state fingerprint.

[0048] If necessary state constraint elements are missing, the system retrieves the most recently effective state constraint element value for the same object identifier before the state binding time within the same operational state snapshot to fill the gap. Only records with an effective time no later than the state binding time are used during the filling process to avoid using future states to fill historical evidence. If filling fails, the corresponding operational evidence fragment is marked as incomplete state evidence. Incomplete state evidence retains the operational evidence fragment identifier, source anchor, and the name of the missing state constraint element, but it is not included in the knowledge entry generation process.

[0049] After generating the evidence status fingerprint, associate the evidence status fingerprint with the operation and maintenance evidence fragment identifier, source anchor point, status binding time, and operation status snapshot identifier. The source anchor point indicates the position of the operation and maintenance evidence fragment in the original record, including the work order segment number, log time range, alarm record number, or replay document segment position.

[0050] Furthermore, after the evidence status fingerprint is formed, the evidence context relationship between the operation and maintenance evidence fragments is generated based on the operation and maintenance evidence fragment identifier, object identifier, status binding time, evidence status fingerprint, source anchor point, and action predicate. The evidence context relationship includes object association, status association, time sequence relationship, and closed-loop node relationship. Object association indicates that two or more operation and maintenance evidence fragments correspond to the same object identifier or belong to the same call chain object; status association indicates that two or more operation and maintenance evidence fragments have the same evidence status fingerprint, or the object identifier, running version, topology version, configuration version, resource status level, and alarm rule version in the evidence status fingerprint meet the matching conditions with the target status fingerprint; time sequence relationship indicates the sequential relationship of fault occurrence, handling execution, and recovery verification formed by two or more operation and maintenance evidence fragments according to the status binding time; closed-loop node relationship indicates the correspondence between operation and maintenance evidence fragments and fault phenomenon nodes, root cause evidence nodes, handling action nodes, or recovery verification nodes.

[0051] S3: Based on the target object and target time window of the knowledge entry to be constructed, extract similar state constraint elements to form the target state fingerprint.

[0052] Furthermore, the extraction of similar state constraint elements to form a target state fingerprint includes: reading the target object and target time window from the construction parameters of the knowledge entry to be constructed; the target time window contains the start and end times to which the knowledge entry is intended to be applied; retrieving runtime snapshots that have an overlay relationship with the target object within the target time window; when more than two runtime snapshots are retrieved, selecting the runtime snapshot with an effective time no later than the end time of the target time window and the latest effective time in the order as the target runtime snapshot; reading state constraint elements with the same name as the evidence state fingerprint from the target runtime snapshot, and processing them according to normalized encoding, resource state level conversion, and fixed order combination; writing the processed state constraint elements into the target state fingerprint, and establishing an association record between the target state fingerprint and the target object, target time window, and target runtime snapshot identifier; when the target object covers multi-level call relationships, forming multiple target state fingerprints according to the call direction, and recording the call sequence relationship among the multiple target state fingerprints.

[0053] It should be noted that one specific approach to extracting similar state constraint elements to form a target state fingerprint involves, after forming the evidence state fingerprint, forming the target state fingerprint based on the target object and target time window of the knowledge entry to be constructed. The evidence state fingerprint describes the state at the time the evidence was generated, while the target state fingerprint describes the state to which the knowledge entry is intended to be applied. To ensure comparability between the two, the target state fingerprint uses the same state constraint elements and encoding rules as the evidence state fingerprint.

[0054] The target object and target time window are read from the construction parameters of the knowledge entry to be constructed. The construction parameters of the knowledge entry to be constructed are the target constraint information formed when the knowledge entry construction is triggered, including at least the target object and the target time window. The target object represents the service object or service chain object corresponding to the knowledge entry to be constructed, and the target time window represents the applicable time range of the knowledge entry to be constructed. The target time window includes the start and end times to which the knowledge entry is intended to be applied. The target object can be a service object or a service chain object covered by multi-level call relationships.

[0055] Within the target time window, retrieve runtime snapshots that have an overriding relationship with the target object. An overriding relationship means that the object identifier in the runtime snapshot is identical to the target object, or that the object identifier in the runtime snapshot belongs to a call node overridden by the target object. When a runtime snapshot is retrieved within the target time window, it is used as the target runtime snapshot. When more than two runtime snapshots are retrieved, the runtime snapshot with an effective time no later than the end time of the target time window and the latest effective time in the order is selected as the target runtime snapshot.

[0056] The selection of the target runtime state snapshot satisfies:

[0057]

[0058] in, Represents the target object The corresponding target running state snapshot, Indicates the first in the operation and maintenance platform A snapshot of the running status. Indicates the first in the operation and maintenance platform A snapshot of the running status. Represented as target object Selected target runtime snapshot index, This represents a set of runtime snapshot indexes that meet the conditions of the target object and the target time window. Represents the target object within the target time window A collection of runtime state snapshots in memory with overlapping relationships. Indicates the start time of the target time window. Indicates the end time of the target time window. Indicates the first The effective time of each runtime state snapshot. If no runtime state snapshot meets the conditions, a target state missing marker is generated, and the construction of formal knowledge entries is stopped.

[0059] Read the state constraint elements with the same names as the evidence state fingerprint from the target's operational state snapshot, and process them according to the same normalized encoding, resource state level conversion, and fixed-order combination as the evidence state fingerprint to form the target state fingerprint. Target object The corresponding target state fingerprint is represented as:

[0060]

[0061] in, Represents the target object The corresponding target state fingerprint, Represents the target object The corresponding object identifier, Represents the target object The corresponding target runtime version, Represents the target object The corresponding target topology version, Represents the target object The corresponding target configuration version, Represents the target object The corresponding target resource status level, Represents the target object The corresponding target alarm rule version.

[0062] The processed state constraint elements are written into the target state fingerprint, and an association record is established between the target state fingerprint and the target object, target time window, and target runtime snapshot identifier. When the target object covers multi-level call relationships, multiple target state fingerprints are formed according to the call direction, and the call sequence relationship is recorded among the multiple target state fingerprints. When multiple target state fingerprints exist, the target state fingerprint with the same object identifier is selected for field comparison based on the object identifier in the evidence state fingerprint; if no target state fingerprint with the same object identifier exists, the corresponding operation and maintenance evidence fragment is written into the isolated evidence list.

[0063] S4: Compare the evidence state fingerprint with the target state fingerprint field by field, isolate evidence with inconsistent states and generate a set of evidence that can be fused.

[0064] Furthermore, refer to Figure 2The process of generating a fusionable evidence set includes: comparing object identifiers in the evidence state fingerprint and the target state fingerprint; when object identifiers are inconsistent, writing the corresponding operation and maintenance evidence fragments into the isolated evidence list; when object identifiers are consistent, comparing fields in the order of running version, topology version, configuration version, resource status level, and alarm rule version to obtain the difference state fields; when no difference state fields exist, writing the corresponding operation and maintenance evidence fragments into the state-consistent evidence set; when difference state fields exist, reading the status change records corresponding to the difference state fields and extracting the difference state field name, the scope of the difference state field, and the objects associated with the difference state fields; based on the source anchor point of the operation and maintenance evidence fragments, locating the evidence-bearing positions corresponding to the fault phenomenon, root cause description, handling actions, and recovery verification in the operation and maintenance evidence fragments; and combining the difference state fields into the fusionable evidence set. The names of the different state fields, the scope of the different state fields, and the objects associated with the different state fields are matched with the object identifiers and action objects in the evidence-bearing locations to form the influence chain of the different state fields. When the influence chain of the different state fields connects to any evidence-bearing location in the root cause description, handling action, or recovery verification, the corresponding operation and maintenance evidence fragment is written into the isolated evidence list. When the influence chain of the different state fields does not connect to any evidence-bearing location in the root cause description, handling action, or recovery verification, the corresponding operation and maintenance evidence fragment is written into the conditionally consistent evidence set, and the different state fields, their scope of application, and applicable conditions are recorded. The state-consistent evidence set and the conditionally consistent evidence set are merged into a fusionable evidence set, and the source anchor correspondence between the isolated evidence list and the fusionable evidence set is preserved.

[0065] It should be noted that one specific scheme for generating a fusionable evidence set includes, after obtaining the evidence state fingerprint and the target state fingerprint, comparing the evidence state fingerprint with the target state fingerprint field by field, and, based on whether the state difference affects the location of key evidence, writing the operation and maintenance evidence fragments into an isolated evidence list, a state-consistent evidence set, or a condition-consistent evidence set. Ordinary RAGs typically retain evidence based on semantic similarity, while this embodiment further determines whether the operational state at the time the evidence occurred can support the root cause description, handling actions, and recovery verification of the target knowledge entry, thereby avoiding the compression of experience from old versions, old topologies, old configurations, or old alarm rules into the knowledge entry.

[0066] For the Each piece of operational evidence is first identified based on the object identifier in the evidence status fingerprint. Select the target state fingerprint that matches the object identifier from the target state fingerprint set, and denote it as... , Indicates the first A target state fingerprint, Indicates the relationship with the first The index of the target object whose object identifier matches the first maintenance evidence fragment. If no target state fingerprint with a matching object identifier exists, then the first... Each piece of maintenance evidence is written into the isolated evidence list.

[0067] The isolated evidence list records operation and maintenance evidence fragments that are not allowed to participate in the generation of compressed evidence packages. Each isolated record includes at least the operation and maintenance evidence fragment identifier, source anchor point, evidence status fingerprint, target status fingerprint, and isolation reason. When writing to the isolated evidence list, the operation and maintenance evidence fragment identifier, source anchor point, evidence status fingerprint, target status fingerprint, and isolation reason are recorded. The isolation reason is recorded as object identifier inconsistency.

[0068] When object identifiers match, fields are compared in the following order: runtime version, topology version, configuration version, resource status level, and alarm rule version, to obtain the difference status fields. The set of difference status fields is represented as follows:

[0069]

[0070] in, Indicates the first The set of difference status fields corresponding to each piece of operation and maintenance evidence. This represents the set of status fields that are being compared. This indicates the running version field. This indicates the topology version field. This indicates the configuration version field. This field represents the resource status level. This field indicates the alarm rule version. Represents any field in the status field set. Indicates the first Fields in the fingerprint of evidence status The value of , Indicates the relationship with the first The target state fingerprint contains fields that are consistent with the object identifier of each maintenance evidence fragment. The value of . If If empty, write the corresponding operation and maintenance evidence fragment into the state consistency evidence set.

[0071] When a differential status field exists, the corresponding status change record is read, and the differential status field name, scope of effect, and associated objects are extracted. The status change record stores the field name, value before the change, value after the change, effective time, scope of effect, associated objects, and the anchor point from which the change originated. The scope of effect of a differential status field indicates the range of service objects or call relationships affected by the corresponding differential status field; the associated objects of a differential status field represent the object identifier, configuration item, topology edge, resource object, or alarm object directly associated with the corresponding differential status field.

[0072] Simultaneously, based on the source anchor point of the maintenance evidence fragment, the evidence-bearing positions corresponding to the fault phenomenon, root cause description, handling action, and recovery verification are located within the maintenance evidence fragment. The evidence-bearing position refers to the text range or structured field location within the maintenance evidence fragment that carries the content of the closing point. For structured work orders, it is located based on field names, paragraph titles, and form items; for unstructured text, it is located based on action predicates, object aliases, and source anchor points. Each evidence-bearing position records at least the node type, object identifier, action object, and source anchor point. Node types include fault phenomenon, root cause description, handling action, and recovery verification; the action object represents the object on which the action predicate acts.

[0073] When forming an influence chain for a difference status field, the difference status field name, its scope of application, and associated objects are matched with the object identifiers and action objects in the evidence-bearing location. If the associated object or scope of application of a difference status field intersects with an object identifier or action object in a certain evidence-bearing location, the difference status field influence chain is considered to be connected to that evidence-bearing location. The connection relationship is represented as follows:

[0074]

[0075] in, Indicates the first The difference status field of each maintenance evidence fragment Is it connected to the first Each piece of evidence is located at a position where a value of 1 indicates a connection and a value of 0 indicates no connection. The traversal index indicating the location where the evidence is located; Indicates the first The difference status field of each maintenance evidence fragment The corresponding scope of the difference status field and the collection of objects composed of the objects associated with the difference status field; Indicates the first The first piece of maintenance evidence. Each piece of evidence consists of a set of object identifiers and action objects within its carrying location. Through this connection, state differences are transformed into an executable judgment of whether the difference falls on a key object or key action in the evidence text.

[0076] When the difference status field affects any evidence location in the root cause description, action, or recovery verification, the corresponding operational evidence fragment is written to the isolated evidence list. The isolation determination is represented as follows:

[0077]

[0078] in, Indicates the first Whether a piece of maintenance evidence is included in the isolated evidence list, a value of 1 indicates that it is included in the isolated evidence list; Indicates the first A collection of evidence locations in an operation and maintenance evidence fragment whose node type is root cause description, handling action, or recovery verification. This indicates the existence of at least one. When writing to the isolated evidence list, record the operation and maintenance evidence fragment identifier, source anchor point, difference status field, scope of effect of difference status field, impact chain of difference status field, and isolation reason. The isolation reason is recorded as the key evidence carrying location being affected by the difference status field.

[0079] When the impact chain of a differing state field is not connected to any evidence-bearing location in the root cause description, action, or recovery verification, the corresponding operational evidence fragment is written into the conditionally consistent evidence set, and the differing state field, its scope, and applicable conditions are recorded. Finally, the state-consistent evidence set and the conditionally consistent evidence set are merged into a fusionable evidence set, while preserving the source anchor correspondence between the isolated evidence list and the fusionable evidence set. This source anchor correspondence provides prohibited and permitted reference boundaries for state constraint generation and field backtracking verification.

[0080] It should also be noted that simply filtering based on identical or different fields is too coarse, because state differences may only affect alarm display without affecting root cause and handling; while other configuration or topology differences can directly change the cause of the fault and the recovery path. By first comparing object identifiers, and then comparing fields such as running version, topology version, configuration version, resource status level, and alarm rule version, the difference state fields are obtained. For segments with differences, the scope and associated objects in the state change record are further read, and the evidence carrying the fault phenomenon, root cause description, handling action, and recovery verification is located based on the source anchor point. When the scope of the difference state field intersects with the object identifier and action object in the evidence carrying location, a difference state field influence chain is formed, and the evidence affecting the root cause, handling, or recovery verification is written into the isolated evidence list. This avoids deleting all difference evidence and also avoids allowing key state inconsistency evidence to enter the compressed context, reducing the probability that the root cause and handling actions in the knowledge entries are biased by old state experience.

[0081] S5: Compress and fuse evidence sets based on the fault causal closed-loop template consisting of fault phenomena, root cause evidence, handling actions, and recovery verification.

[0082] Furthermore, compressing the fusionable evidence set includes: establishing a fault causal closed-loop template containing fault phenomenon nodes, root cause evidence nodes, handling action nodes, and recovery verification nodes; writing the evidence fragments into the corresponding nodes in the fault causal closed-loop template based on the text tags, action predicates, and source anchors of each evidence fragment in the fusionable evidence set; when an evidence fragment covers the content of multiple nodes simultaneously, it is divided into multiple evidence fragments according to the node boundaries, and each fragment is written into only one node; within each node, the source anchor, evidence state fingerprint, key action words, and related information are preferentially retained. Keyword objects are used to merge evidence fragments with the same evidence state fingerprint, the same object identifier, and the same action words. When evidence with consistent conditions enters the compression, the difference state field and applicable conditions are written immediately after the corresponding compressed text. When any node in the root cause evidence node and the recovery verification node is empty, an evidence package to be completed is generated. When all nodes in the fault causal closed loop template have been written with evidence fragments, a compressed evidence package is generated in the order of fault phenomenon node, root cause evidence node, disposal action node, and recovery verification node, and the source anchor point correspondence between each text segment in the compressed evidence package and the original evidence fragment is preserved.

[0083] It should be noted that one specific approach to compressing the fusionable evidence set includes, after generating the fusionable evidence set, compressing it according to a fault causal closed-loop template consisting of fault phenomena, root cause evidence, handling actions, and recovery verification. This embodiment uses a fault causal closed-loop template to ensure that the compressed evidence package retains the key structure required for the knowledge entries to be valid even when the context window is limited.

[0084] Establish a closed-loop template for fault cause and effect, including fault phenomenon nodes, root cause evidence nodes, handling action nodes, and recovery verification nodes. Fault phenomenon nodes carry alarm manifestations, abnormal indicators, and user impact; root cause evidence nodes carry abnormal logs, dependency anomalies, configuration changes, resource bottlenecks, or topology anomalies; handling action nodes carry restarts, rollbacks, expansions, configuration modifications, traffic switching, and rate limiting; recovery verification nodes carry alarm clearance, indicator recovery, link recovery, or work order confirmation.

[0085] For each piece of evidence in the fusionable evidence set, the evidence piece is written into the corresponding node in the fault causal loop template based on the text tag, action predicate, and source anchor. To avoid incorrectly writing evidence pieces into nodes based solely on keywords, this embodiment uses the text tag, action predicate, and source anchor together as the node matching criteria. The text tag reflects the structural position of the evidence piece in the original work order or review record, the action predicate reflects the technical action described by the evidence piece, and the source anchor reflects the specific location of the evidence piece from the alarm, log, work order field, or review paragraph.

[0086] For the first in the set of fusing evidence Each piece of evidence is written to a node based on the degree of matching between the text tags, action predicates, source anchors, and each node, satisfying the following conditions:

[0087]

[0088] in, This represents a set of nodes consisting of fault phenomenon nodes, root cause evidence nodes, handling action nodes, and recovery verification nodes. Represents a set of nodes Any node in, Indicates the first The node where a piece of evidence is written. Indicates the first Evidence fragments and nodes The matching value, Indicates the first The text tags of each piece of evidence. Indicates the first Action predicates in a piece of evidence; Indicates the first The source anchor point of each piece of evidence, , , Representing nodes respectively The corresponding set of tags, action predicates, and source anchor types. This means that the value is 1 if the condition is true and 0 if the condition is false. , , These represent the matching weights for text tags, action predicates, and source anchors, respectively.

[0089] Preferably, , , Text tags and action predicates typically determine the semantic nodes of evidence fragments, while source anchors primarily serve as auxiliary verification information. If two or more nodes have the same matching value, the node to be written is determined according to the priority order of text tag match, action predicate match, and source anchor match; if it still cannot be uniquely determined, the corresponding evidence fragment is written to the evidence package to be supplemented, but not to the compressed evidence package.

[0090] When an evidence fragment covers the content of multiple nodes, it is segmented into multiple evidence fragments according to node boundaries, and each segmented evidence fragment is written to only one node. Node boundaries are determined based on paragraph titles, period separators, action predicate categories, and source anchor point positions; when node boundaries cannot be determined by titles, action predicate categories are used as the segmentation basis. Specifically, the evidence fragment is segmented into multiple sentence segments by periods, semicolons, line breaks, or table fields, and the node type of each sentence segment is determined according to node matching rules. If the node types corresponding to adjacent sentence segments are different, the position between them is determined as the node boundary. Node boundaries satisfy:

[0091]

[0092] in, Indicates the first The set of node boundaries in a piece of evidence. Indicates the first Segment position index in each piece of evidence Indicates the first The first piece of evidence A sentence or paragraph, Indicates the relationship with the first The next sentence adjacent to each other in a sentence segment. Sentence segment The node type is determined according to the node matching rules. Sentence segment The node type is determined according to the node matching rules. If the node types of adjacent sentence segments are different, the position between them is determined as the node boundary, and an evidence segment is divided into multiple evidence segments according to the node boundary, so that each segmented evidence segment is written into only one node in the fault causal closed loop template.

[0093] Within each node, priority is given to retaining the source anchor, evidence status fingerprint, key action words, and key object words. Evidence fragments with the same evidence status fingerprint, object identifier, and action words are then merged. Key action words are predicates in the evidence fragment that represent operational actions or status changes, while key object words are words in the evidence fragment that represent service objects, configuration items, topology nodes, alarm objects, or resource objects. During merging, the source anchor with the earliest timestamp or the earliest segment number is retained, along with the set of key object words containing the largest number of key object words. When conditionally consistent evidence is compressed, the difference status field and applicable conditions are written immediately after the corresponding compressed text to prevent the conditionally consistent evidence from losing its applicable boundaries after compression.

[0094] To ensure that the compression has executable boundaries, this embodiment sets an upper limit on the target character length for the compressed evidence package. , , This indicates the maximum character length of the input window for the generative AI model. Each closed node retains at least one piece of evidence; when multiple pieces exist within a node, state-consistent evidence is retained first, followed by condition-consistent evidence; among the same type of evidence, fragments with complete source anchors and where both key action words and key object words are present are retained first. A complete source anchor means that the source anchor simultaneously contains the original record identifier and the paragraph position, field position, or time range in the original record.

[0095] The compressed length satisfies:

[0096]

[0097] in, This represents a set of nodes consisting of fault phenomenon nodes, root cause evidence nodes, handling action nodes, and recovery verification nodes. express Any node in, Represents a node The compressed text content, Represents a node The length of the compressed characters.

[0098] When the total length of the text before compression is greater than First, merge fragments with the same evidence state fingerprint, the same object identifier, and the same action word within each node; if the total length of the text before compression is still greater than [a certain value], then [further steps are needed]. At that time, delete duplicate segments with incomplete source anchors within the same node; if the total length of the text before compression is still greater than When compressing, the abbreviation should be based on the condition that at least one piece of evidence is retained for each of the fault phenomenon node, root cause evidence node, handling action node, and recovery verification node. Root cause evidence nodes and recovery verification nodes cannot be compressed to empty.

[0099] When either the root cause evidence node or the recovery verification node is empty, a supplementary evidence package is generated. This supplementary evidence package records existing nodes, missing nodes, source anchors, and the target state fingerprint, but is not written to the high-speed operation and maintenance knowledge base. When all nodes in the fault causal closed-loop template have evidence fragments written to them, a compressed evidence package is generated in the order of fault phenomenon nodes, root cause evidence nodes, handling action nodes, and recovery verification nodes, while preserving the source anchor correspondence between each text segment in the compressed evidence package and the original evidence fragments.

[0100] It should also be noted that ordinary summary compression typically targets text length, keyword coverage, or semantic similarity. This can easily preserve fault phenomena and handling actions during compression, but lose root cause evidence or recovery verification. Generative AI then fills in the missing links, making the source unverifiable. By establishing fault phenomenon nodes, root cause evidence nodes, handling action nodes, and recovery verification nodes, evidence fragments are written into corresponding nodes according to text tags, action predicates, and source anchors, ensuring the causal chain of fault analysis is maintained during compression. When a fragment covers the content of multiple nodes, it is segmented according to node boundaries to avoid the same text segment being incorrectly assigned to a single node. Within a node, source anchors, evidence state fingerprints, key action words, and key object words are prioritized for retention. Duplicate fragments with the same state fingerprints, object identifiers, and action words are merged to reduce input length while retaining the object, action, state, and source information required to generate knowledge entries. When a root cause evidence node or recovery verification node is empty, a supplementary evidence package is generated and not directly added to the database, reducing the number of knowledge entries lacking closed-loop evidence entering the high-speed operation and maintenance knowledge base.

[0101] S6: Input the compressed evidence package into the generative AI model to perform state constraint generation, field backtracking verification, and indexing into the database.

[0102] Furthermore, the execution of state constraint generation, field backtracking verification, and indexing includes: constructing input prompts for the generative AI model based on the compressed evidence package; writing the input prompts into the target state fingerprint, fusionable evidence set, isolated evidence list, fault causal closed-loop template, and knowledge item field constraints; limiting the root cause field to reference the root cause evidence node, the handling field to reference the handling action node, the recovery verification field to reference the recovery verification node, and the applicable condition field to reference the difference state field in the target state fingerprint and condition consistency evidence; the generative AI model generates state-bound operation and maintenance knowledge items according to the input prompts; state-bound operation and maintenance knowledge items include fault phenomena, root cause descriptions, handling actions, recovery verification methods, applicable state fingerprints, and source anchors; after generation, the compressed evidence package is checked item by item according to the knowledge item fields; when a knowledge field cannot be backtracked to the corresponding node in the compressed evidence package, or when a knowledge field references an evidence fragment in the isolated evidence list, the corresponding knowledge field is deleted and the generative AI model generation is retried; when all necessary knowledge fields have completed backtracking verification, the state-bound operation and maintenance knowledge items are written into the high-speed operation and maintenance knowledge base, and a text vector index, a state fingerprint index, and a source anchor index are established simultaneously.

[0103] It should be noted that one specific approach to generating state constraints, backtracking and validating fields, and storing indexes includes, after obtaining the compressed evidence package, inputting the compressed evidence package into the generative AI model, and then performing state constraint generation, backtracking and validating fields, and storing indexes.

[0104] In this embodiment, the generative AI model is a pre-trained natural language generation model. This model is not the sole basis for determining evidence usability and does not participate in the binding of operational state snapshots, the formation of evidence state fingerprints, the formation of target state fingerprints, the formation of influence chains for differing state fields, or the generation of fusionable evidence sets. The input to the generative AI model consists of a compressed evidence package, a target state fingerprint, an isolated evidence list, a fault causal closed-loop template, and knowledge item field constraints. The output of the generative AI model is state-bound operational knowledge items, which include fault phenomena, root cause descriptions, handling actions, recovery verification methods, applicable state fingerprints, and source anchors. Each knowledge field output by the generative AI model must carry a source anchor and be written into the high-speed operational knowledge base only after field backtracking verification.

[0105] The operational evidence fragments and operational status snapshots used in this invention originate from operational systems, monitoring systems, alarm systems, configuration systems, release systems, work order systems, and fault review systems that the implementing entity has legitimate management authority over. For operational evidence fragments that may contain account information, mobile phone numbers, email addresses, access tokens, keys, user identifiers, or other sensitive information, desensitization processing is performed before entering RAG retrieval, evidence compression, and generative AI model input prompts. Desensitization processing does not change the source anchor point, state binding time, or the association between the evidence state fingerprint and the target state fingerprint. State-bound operational knowledge entries output by the generative AI model are only used for operational knowledge base construction and retrieval, and do not directly execute production environment commands; knowledge entries involving disposal scripts, rollback operations, or production configurations are subject to access control according to operational permissions.

[0106] The input prompts for the generative AI model are constructed based on the compressed evidence package. These prompts include the target state fingerprint, the set of fusionable evidence, the list of isolated evidence, the fault causal loop template, and knowledge entry field constraints. To ensure the generative AI model outputs verifiable content, the input prompts require each generated field to carry a source anchor. Knowledge entry field constraints limit the root cause field to reference the root cause evidence node, the disposition field to reference the disposition action node, the recovery verification field to reference the recovery verification node, and the applicable condition field to reference the difference state field in the target state fingerprint and the conditionally consistent evidence. The input prompts also specify that source anchors in the isolated evidence list must not appear in the generated results.

[0107] The generative AI model generates state-bound O&M knowledge entries based on input prompts. These entries include fault symptoms, root cause descriptions, remedial actions, recovery verification methods, applicable state fingerprints, and source anchors. A field-source mapping is established between each field and its source anchor. The root cause description field corresponds to the source anchor in the root cause evidence node, the remedial action field corresponds to the source anchor in the remedial action node, and the recovery verification method field corresponds to the source anchor in the recovery verification node. If a field output by the generative AI model does not carry a source anchor, the corresponding field is marked as an un-backtracked field and is not written to the high-speed O&M knowledge base.

[0108] Field backtracking verification involves reviewing and compressing the evidence package item by item according to the knowledge field. For knowledge fields... The source anchor set is denoted as The set of source anchor points corresponding to the field types in the compressed evidence package is denoted as... The set of source anchors in the isolated evidence list is denoted as Field backtracking validation satisfies:

[0109]

[0110] in, Represents knowledge fields Whether the backtracking check passed or not, a value of 1 indicates that it passed. This indicates that the generative AI model is a knowledge field. The set of source anchor points for the output. This indicates the knowledge field in the compressed evidence package. The set of anchor points from which the closing loop point corresponds to the field type. This represents the set of source anchors in the list of isolated evidence. For the root cause field, This is the set of source anchor points for root cause evidence nodes; for the disposal field, The set of source anchor points for handling action nodes; for the recovery verification field, To restore the source anchor set of the verification node; for applicable condition fields, This is the set of anchor points for the source of fingerprints in the target state and the source of evidence consistent with the conditions.

[0111] When a knowledge field cannot be traced back to the corresponding node in the compressed evidence package, or when a knowledge field references evidence fragments in the isolated evidence list, the corresponding knowledge field is deleted. The deleted field name, allowed closed-loop nodes, and prohibited source anchors are then rewritten into the input prompt, triggering the generative AI model to regenerate the corresponding field. Preferably, a single knowledge field is regenerated twice. If the field fails the backtracking verification after two regenerations, the corresponding state-bound maintenance knowledge item is marked as an item awaiting manual review. The number of regenerations can be configured by the maintenance knowledge base construction task.

[0112] Once all necessary knowledge fields have undergone backtracking verification, state-bound O&M knowledge entries are written to the high-speed O&M knowledge base, and text vector indexes, state fingerprint indexes, and source anchor point indexes are simultaneously established. The text vector index is used to semantically recall knowledge entries, the state fingerprint index is used to filter knowledge entries according to the target operating state, and the source anchor point index is used to retrieve the knowledge source based on the location of the original evidence. During writing, state-bound O&M knowledge entries must at least store the knowledge entry identifier, fault phenomenon, root cause description, handling action, recovery verification method, applicable state fingerprint, source anchor point set, and knowledge state. The initial value of the knowledge state is "available"; if there are differing state fields in the applicable conditions fields, the initial value of the knowledge state is "restricted availability," and the applicable conditions are also stored.

[0113] Furthermore, the evidence context is written into the high-speed operation and maintenance knowledge base, and a context index is established. The context index is used to retrieve state-bound operation and maintenance knowledge entries and corresponding operation and maintenance evidence fragments according to object association, state association, time sequence, and closed-loop point relationship. When the generative AI model calls the high-speed operation and maintenance knowledge base, it first filters candidate knowledge entries from the state fingerprint index based on the target object and target state fingerprint, and then locates the evidence fragments corresponding to the fault phenomenon, root cause evidence, handling actions, and recovery verification based on the context index, and writes the location results into the input prompts of the generative AI model.

[0114] It should also be noted that generative AI models, after receiving compressed evidence, may expand upon root causes, actions, or verification conclusions not explicitly stated in the evidence. Without field-level validation, while the knowledge entries may be linguistically complete, their origins are untraceable. By incorporating the target state fingerprint, fusionable evidence set, isolated evidence list, fault causal loop template, and knowledge entry field constraints into the input prompts, different knowledge fields are limited to referencing their corresponding closed-loop nodes. The root cause field, action field, recovery verification field, and applicable condition field in the generated results must carry the source anchor, and each item must be back-checked against the compressed evidence package; fields that cannot be back-checked to the corresponding node or reference the isolated evidence list are deleted and regenerated. Through post-generation validation, the model output no longer relies solely on linguistic probabilities but must be consistent with the source anchor, state fingerprint, and closed-loop nodes. By establishing text vector indexes, state fingerprint indexes, and source anchor indexes, the knowledge entries stored in the database can simultaneously support semantic retrieval, operational status filtering, and original evidence back-checking, reducing the disconnect between generated content and actual operational status.

[0115] One embodiment of the present invention provides a high-speed operation and maintenance knowledge base construction system based on generative AI, including a snapshot binding module, an evidence fingerprint module, a target fingerprint module, a state gating module, a closed-loop compression module, and a backtracking database module.

[0116] The system comprises the following modules: a snapshot binding module for acquiring operational evidence fragments retrieved by RAG and binding them to the corresponding service object's operational status snapshot based on the event time; an evidence fingerprint module for reading state constraint elements from the operational status snapshot to form an evidence state fingerprint; a target fingerprint module for extracting similar state constraint elements to form a target state fingerprint based on the target object and target time window of the knowledge entry to be constructed; a state gating module for comparing the evidence state fingerprint with the target state fingerprint field by field, isolating inconsistent state evidence and generating a fusionable evidence set; a closed-loop compression module for compressing the fusionable evidence set according to the fault causal closed-loop template composed of fault phenomena, root cause evidence, handling actions, and recovery verification; and a backtracking and database entry module for inputting the compressed evidence package into the generative AI model to perform state constraint generation, field backtracking verification, and indexing and database entry.

[0117] Reference Figure 3 This embodiment also provides a computer device applicable to the high-speed operation and maintenance knowledge base construction method based on generative AI, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the high-speed operation and maintenance knowledge base construction method based on generative AI proposed in the above embodiment.

[0118] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0119] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for constructing a high-speed operation and maintenance knowledge base based on generative AI as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

Claims

1. A high-speed operation and maintenance knowledge base construction method based on generative AI, characterized by, include: Obtain operational evidence fragments from RAG recalls and bind and overwrite the corresponding service object's runtime status snapshot based on the event time; Read state constraint elements from the runtime state snapshot to form an evidence state fingerprint; Based on the target object and target time window of the knowledge entry to be constructed, extract similar state constraint elements to form a target state fingerprint; The evidence status fingerprint is compared with the target status fingerprint field by field to isolate inconsistent evidence and generate a set of fusionable evidence. The evidence set that can be fused can be compressed based on the fault causal closed-loop template consisting of fault phenomena, root cause evidence, handling actions, and recovery verification. The compressed evidence package is input into the generative AI model to perform state constraint generation, field backtracking verification, and indexing into the database.

2. The generative AI-based high-speed operation and maintenance knowledge base construction method of claim 1, wherein: The method of binding and overriding the runtime status snapshot of the corresponding service object based on event time includes... The source identifier and candidate time field of the operation and maintenance evidence fragments recalled by RAG are read; the candidate time field includes the event occurrence time, the handling execution time, and the recovery verification time; When multiple candidate time fields coexist, the candidate time field closest to the action predicate in the evidence fragment text is selected as the state binding time, according to the order of fault occurrence, handling action execution, and recovery verification completion. When only one candidate time field exists, the candidate time field is used as the state binding time; Retrieve status change records stored in the operation and maintenance platform by status binding time, select the running status snapshot whose effective time is no later than the status binding time and whose next effective time is later than the status binding time, and use it as the running status snapshot to cover the corresponding service object. When the evidence fragment does not carry a clear service object, the fragment source identifier is mapped to the service object based on the fragment source identifier and the object alias table in the text, and then the running status snapshot is retrieved according to the service object; After binding is completed, the runtime status snapshot identifier, status binding time, and object identifier will be written into the metadata record of the operation and maintenance evidence fragment.

3. The high-speed operation and maintenance knowledge base construction method based on generative AI according to claim 2, characterized in that: The fingerprint representing the evidence state includes, Read the state constraint elements from the bound running state snapshot, perform normalized encoding, resource state level conversion, and combine them in a preset order to form evidence state fingerprints; The state constraint elements include the object identifier that represents the object to which the operation and maintenance evidence fragment belongs, and the running version, topology version, configuration version, resource status level and alarm rule version that represent the running conditions when the operation and maintenance evidence fragment is generated. Establish a record linking the evidence status fingerprint with the operation and maintenance evidence fragment identifier, source anchor point, status binding time, and operation status snapshot identifier; When necessary state constraint elements are missing, the state constraint element value that was most recently effective before the state binding time is retrieved from the same running state snapshot to fill in the missing elements. If the filling fails, the corresponding operation and maintenance evidence fragment is marked as incomplete state evidence.

4. The high-speed operation and maintenance knowledge base construction method based on generative AI of claim 3, wherein: The step of extracting similar state constraint elements to form a target state fingerprint includes... Read the target object and target time window from the construction parameters of the knowledge entry to be constructed; the target time window includes the start and end times to which the knowledge entry is to be applied. Retrieve runtime snapshots that overlap with the target object within the target time window; When more than two runtime snapshots are retrieved, the runtime snapshot with the latest effective time sorted and no later than the end time of the target time window is selected as the target runtime snapshot. Read the state constraint elements with the same name as the evidence state fingerprint from the target's operational state snapshot, and process them according to normalized coding, resource state level conversion, and fixed order combination. Write the processed state constraint elements into the target state fingerprint, and establish an association record between the target state fingerprint and the target object, the target time window, and the target running state snapshot identifier; When the target object covers a multi-level call relationship, multiple target state fingerprints are formed according to the call direction, and the call sequence relationship is recorded among the multiple target state fingerprints.

5. The generative AI-based high-speed operation and maintenance knowledge base construction method of claim 4, wherein: The generated set of fusionable evidence includes, Compare the object identifiers in the evidence status fingerprint and the target status fingerprint. If the object identifiers are inconsistent, write the corresponding operation and maintenance evidence fragments into the isolated evidence list. When the object identifiers are consistent, the fields are compared in the order of running version, topology version, configuration version, resource status level, and alarm rule version to obtain the difference status fields; When there is no difference status field, the corresponding operation and maintenance evidence fragment will be written into the status consistency evidence set; When a difference status field exists, read the status change record corresponding to the difference status field and extract the difference status field name, the scope of the difference status field, and the object associated with the difference status field. Based on the source anchor point of the operation and maintenance evidence fragment, locate the evidence carrying position corresponding to the fault phenomenon, root cause description, handling action and recovery verification in the operation and maintenance evidence fragment; The difference status field name, the scope of the difference status field, and the object associated with the difference status field are matched with the object identifier and action object in the evidence carrying location to form the difference status field influence chain; When the difference status field affects any evidence-bearing location in the root cause description, handling action, and recovery verification, the corresponding operation and maintenance evidence fragment will be written into the isolated evidence list. When the difference status field's impact chain is not connected to any evidence-bearing location in the root cause description, handling action, or recovery verification, the corresponding operation and maintenance evidence fragment will be written into the conditionally consistent evidence set, and the difference status field, the scope of the difference status field, and the applicable conditions will be recorded. The state-consistent evidence set and the condition-consistent evidence set are merged into a fusionable evidence set, while preserving the source anchor correspondence between the isolated evidence list and the fusionable evidence set.

6. The generative AI-based high-speed operation and maintenance knowledge base construction method of claim 5, wherein: The compressed and fusionable evidence set includes, Establish a closed-loop template for the cause and effect of a fault, which includes fault phenomenon nodes, root cause evidence nodes, handling action nodes, and recovery verification nodes. Based on the text tags, action predicates, and source anchors of each evidence fragment in the fusionable evidence set, the evidence fragments are written into the corresponding nodes in the fault causal closed loop template. When a piece of evidence covers the content of multiple nodes, it is divided into multiple pieces of evidence according to the node boundaries, and each piece of evidence is written to only one node. Within each node, the source anchor point, evidence status fingerprint, key action words, and key object words are retained first, and then evidence fragments with the same evidence status fingerprint, the same object identifier, and the same action words are merged. When evidence of consistent conditions is compressed, the difference status field and applicable conditions are written immediately after the corresponding compressed text. When either the root cause evidence node or the recovery verification node is empty, a supplementary evidence package is generated. When all nodes in the fault causal closed-loop template are written with evidence fragments, a compressed evidence package is generated in the order of fault phenomenon node, root cause evidence node, handling action node, and recovery verification node, and the source anchor point correspondence between each text segment in the compressed evidence package and the original evidence fragment is preserved.

7. The generative AI-based high-speed operation and maintenance knowledge base construction method of claim 6, wherein: The execution status constraint generation, field backtracking verification, and index insertion include, Input prompts are constructed based on the compressed evidence package to create a generative AI model. Input prompts are written into the target state fingerprint, the set of evidence that can be fused, the list of isolated evidence, the fault causal closed-loop template, and the knowledge entry field constraints; The knowledge entry field constraints limit the root cause field to reference the root cause evidence node, the disposition field to reference the disposition action node, the recovery verification field to reference the recovery verification node, and the applicable condition field to reference the difference state field in the target state fingerprint and condition consistency evidence. The generative AI model generates state-bound operation and maintenance knowledge entries based on input prompts; the state-bound operation and maintenance knowledge entries include fault phenomena, root cause descriptions, handling actions, recovery verification methods, applicable state fingerprints, and source anchor points. After generation, the compressed evidence package is checked item by item according to the knowledge field. When a knowledge field cannot be traced back to the corresponding node in the compressed evidence package, or when a knowledge field references an evidence fragment in the isolated evidence list, the corresponding knowledge field is deleted and the generative AI model generation is retried. Once all necessary knowledge fields have completed backtracking verification, state-bound operation and maintenance knowledge entries are written into the high-speed operation and maintenance knowledge base, and text vector indexes, state fingerprint indexes, and source anchor point indexes are established simultaneously.

8. A high-speed operation and maintenance knowledge base construction system based on generative AI, which adopts the high-speed operation and maintenance knowledge base construction method based on generative AI according to any one of claims 1 to 7, characterized in that: Includes a snapshot binding module, an evidence fingerprint module, a target fingerprint module, a state gating module, a closed-loop compression module, and a backtracking and data entry module; The snapshot binding module is used to obtain operation and maintenance evidence fragments from RAG recall and bind and cover the running status snapshots of the corresponding service objects according to the event time. The evidence fingerprint module is used to read state constraint elements from the running state snapshot to form an evidence state fingerprint; The target fingerprint module is used to extract similar state constraint elements to form a target state fingerprint based on the target object and target time window of the knowledge entry to be constructed. The state gating module is used to compare the evidence state fingerprint with the target state fingerprint field by field, isolate evidence with inconsistent states and generate a set of evidence that can be fused. The closed-loop compression module is used to compress a set of evidence that can be fused according to a fault causal closed-loop template consisting of fault phenomena, root cause evidence, handling actions, and recovery verification. The backtracking and data entry module is used to input the compressed evidence package into the generative AI model, and perform state constraint generation, field backtracking verification, and index data entry. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8. When the processor executes the computer program, it implements the steps of the method for constructing a high-speed operation and maintenance knowledge base based on generative AI as described in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for constructing a high-speed operation and maintenance knowledge base based on generative AI as described in any one of claims 1 to 7.