AI-based intelligent pushing method and system for IT service knowledge base

Abstract

This invention relates to the field of intelligent knowledge base recommendation, providing an AI-based method and system for intelligent recommendation of IT service knowledge bases. The method includes: generating a set of operational entities and initial association triples based on the original dataset; constructing a knowledge graph topology and mapping it to a graph feature matrix; mapping real-time dialogue text streams to subjective semantic feature vectors; generating environmental state feature vectors by traversing target device identifiers based on the knowledge graph topology; extracting a preliminary candidate set by traversing the graph feature matrix; filtering and sorting the candidate set based on the knowledge graph topology to obtain a final knowledge recommendation list; generating intelligent recommendation cards based on the final knowledge recommendation list; collecting user operation behavior data and real-time environmental state streams to calculate a comprehensive reward value; and updating the parameter set using a policy gradient algorithm through backpropagation. This invention integrates subjective semantic features and objective environmental states to construct an intelligent recommendation and adaptive optimization mechanism based on a closed loop of topology reachability verification and physical verification.

Description

Technical Field

[0001] This invention relates to the field of intelligent knowledge base recommendation, and in particular to an AI-based method and system for intelligent recommendation of IT service knowledge bases. Background Technology

[0002] As enterprises deepen their digital transformation, the level of intelligent utilization of knowledge bases in IT service management systems has become a decisive factor in ensuring operational response speed and service quality. Existing knowledge base systems primarily rely on manual, passive retrieval, making it difficult to automatically match solutions based on the real-time context of work orders. This results in a large amount of historical experience not being effectively reused. Overcoming the limitations of passive querying and achieving a technological leap from passive retrieval to context-based proactive push has become a critical technical challenge that urgently needs to be addressed to improve the first-response resolution rate and operational response efficiency of IT service desks.

[0003] Chinese patent application CN121073401A discloses a fully automated method for generating and pushing work order processing information based on a fault knowledge base. The method includes: obtaining an initial decision tree; identifying core fault work orders among multiple fault work orders corresponding to a target fault type; pruning the initial decision tree based on the core fault work orders to obtain a target decision tree, wherein the number of nodes in the target decision tree is less than the number of nodes in the initial decision tree; and processing the multiple fault work orders corresponding to the target fault type separately according to the target decision tree to obtain work order handling information for each fault work order. By pruning the decision tree to obtain a more lightweight target decision tree, and using this tree to batch process multiple fault work orders corresponding to the target fault type, the overall processing time for multiple fault work orders corresponding to the target fault type is reduced, improving the overall response efficiency of a large number of fault work orders.

[0004] However, current technologies still face numerous challenges. In troubleshooting scenarios within IT service management systems, end users often use unstructured and ambiguous natural language, such as "system loop" or "unable to log in," to describe technical faults. Existing knowledge base retrieval primarily relies on keyword matching technology based on inverted indexes, which struggles to bridge the semantic gap between natural language and standardized technical terms in the knowledge base, such as "response timeout" or "authentication failure." Furthermore, traditional systems lack the ability to perceive the context of work orders and the dynamic operational environment, making it impossible to proactively establish logical connections between current faults and historical solutions. This passive query mechanism, relying on manual retrieval, renders a massive amount of valuable experience accumulated in historical work orders unstructured and unusable silent data that cannot be automatically parsed and reused by the system. This prevents operations personnel from obtaining immediate and actionable solutions when facing urgent faults, severely limiting the first-response resolution rate of the IT service desk, leading to delayed fault response, and consequently affecting the continuity and stability of enterprise business systems. Summary of the Invention

[0005] To achieve the above objectives, this invention provides an AI-based intelligent push method for IT service knowledge bases, the specific technical solution of which is as follows:

[0006] Based on the collected raw data set, an operation and maintenance entity set and an initial association triplet are generated. A knowledge graph topology structure is constructed based on the operation and maintenance entity set and the initial association triplet, and the knowledge graph topology structure is mapped to a graph feature matrix.

[0007] The real-time captured dialogue text stream is discretized and mapped into a subjective semantic feature vector, and the target device identifier pushed by the alarm interface of the external monitoring system is traversed based on the knowledge graph topology to generate an environmental state feature vector.

[0008] The subjective semantic feature vector and the environmental state feature vector are integrated to generate a hybrid query vector. Based on the hybrid query vector, the graph feature matrix is ​​traversed to extract a preliminary candidate set. The preliminary candidate set is then filtered and sorted in combination with the knowledge graph topology to obtain the final knowledge recommendation list.

[0009] Based on the final knowledge recommendation list, a pre-set automated tool library index is retrieved to encapsulate and generate a smart push card. User operation behavior data and real-time environmental state stream for the smart push card are collected to calculate the comprehensive reward value. The policy gradient algorithm is used to perform backpropagation update on the current parameter set.

[0010] Furthermore, the method for generating the spectral feature matrix includes:

[0011] A standard corpus sequence is generated based on the collected raw data set, and the standard corpus sequence is mapped into a state feature matrix through a sequence labeling calculation unit. Global path constraint optimization is performed on the state feature matrix to lock the global optimal label path. The standard corpus sequence is then cut according to the global optimal label path to generate a set of operation and maintenance entities and an initial association triplet.

[0012] Based on the set of operation and maintenance entities and the initial association triplet, a directed graph data structure is instantiated. Cross-modal logical association verification is performed on the directed graph data structure to establish the violating association edges. The dependent association edges are derived by passing closure operation to complete the derivation. The knowledge graph topology structure is then reconstructed.

[0013] Based on the knowledge graph topology, an initial state graph feature matrix is ​​instantiated. Then, parameter iterative optimization logic is executed to backpropagate and update the initial state graph feature matrix to obtain the graph feature matrix.

[0014] Furthermore, the method for locking the globally optimal label path includes:

[0015] A state selection topology graph is constructed based on the state feature matrix, which is defined by the temporal dimension and the state dimension. The temporal dimension corresponds to the discrete time step index of the standard corpus sequence, and the state dimension corresponds to the preset set of operation and maintenance entity labels.

[0016] For each discrete time step, select each candidate state node in the topology graph, calculate the linear sum of the local observation potential of the candidate state node, the pre-optimal recursive potential pointing to the candidate state node, and the state transition potential of the candidate state node, and retain the local optimal sub-path pointer pointing to the current candidate state node that maximizes the linear sum.

[0017] The local observation potential is the non-normalized emission probability intensity determined based on the state feature matrix at the corresponding temporal dimension index and state dimension index; the state transition potential is the logical dependency weight determined based on the pre-stored global transition probability matrix.

[0018] In response to the recursive traversal to the last discrete time step, the candidate state node with the maximum accumulated potential energy value is selected. The reverse pointer backtracking is performed along the local optimal sub-path pointer in the state selection topology graph to lock the globally optimal label path.

[0019] Furthermore, the method for identifying violated associated edges includes:

[0020] Traverse the fault feature entity subset and the design baseline index entity subset in the operation and maintenance entity set to construct the entity pair to be verified, and perform semantic attribute consistency verification and numerical constraint violation verification on the entity pair to be verified.

[0021] The semantic attribute consistency verification includes calculating the text similarity value between the fault object name field parsed from the fault feature entity and the benchmark indicator name field parsed from the design benchmark indicator entity. If the text similarity value is greater than the preset semantic similarity threshold, the semantic attribute consistency verification is passed. The numerical constraint violation verification includes extracting the runtime observation value from the fault feature entity and the design limit value from the design benchmark indicator entity, and performing a binary logical operation on the runtime observation value and the design limit value.

[0022] If the entity pair to be verified passes the semantic attribute consistency check and numerical constraint violation check, and the result of the binary logic operation is true, then a violation association edge pointing from the fault feature entity to the design benchmark index entity is instantiated in the knowledge graph topology.

[0023] Furthermore, the method for generating the environmental state feature vector includes:

[0024] The real-time captured dialogue text stream is discretized into a valid text sequence within a sliding sampling time window. Based on the mapping correspondence in the pre-set IT operation and maintenance semantic parsing unit, the valid text sequence is mapped into a feature state matrix sequence. The feature state matrix sequence is then compressed with full sequence information to generate a subjective semantic feature vector.

[0025] In response to the target device identifier pushed by the alarm interface of the external monitoring system, a breadth-first search is performed in the knowledge graph topology to lock the set of related dependent components. The running index data of the set of related dependent components is collected within the sliding sampling time window. A time-series running status matrix is ​​generated based on the sliding sampling time window, the set of related dependent components, and the running index data.

[0026] Z-score normalization and first-order difference operation are performed on the time-series running state matrix to generate dimensionless running values ​​and time-series fluctuation rates, respectively. A joint feature input vector is constructed based on the dimensionless running values ​​and time-series fluctuation rates. The joint feature input vector is then input into a multilayer perceptron network to generate an environmental state feature vector.

[0027] Furthermore, the method for obtaining the final knowledge recommendation list includes:

[0028] The probability distribution of the subjective semantic feature vector is normalized to calculate the semantic uncertainty index. The semantic uncertainty index is mapped to the semantic feature weight coefficient using a dynamic gated attention arbitration unit. Based on the semantic feature weight coefficient, the subjective semantic feature vector and the environmental state feature vector are linearly weighted and fused to generate a hybrid query vector.

[0029] Extract candidate solution entity vectors from the graph feature matrix, perform complex similarity measurement operation to calculate the semantic cosine similarity value between the mixed query vector and the candidate solution entity vector, and perform result sorting and filtering logic based on the semantic cosine similarity value to construct a preliminary candidate set;

[0030] Using the target device identifier as the physical starting anchor, the system retrieves valid connected paths in the knowledge graph topology that connect the initial candidate set. Based on the valid connected paths, it performs topological association quantification to generate a topological logical association value, and combines it with semantic cosine similarity values ​​to perform linear weighted fusion to generate a comprehensive confidence score. Finally, it reorders the initial candidate set based on the comprehensive confidence score to output the final knowledge recommendation list.

[0031] Furthermore, the method for performing the topological association metric calculation includes:

[0032] Traverse the directed edges contained in the valid connected path. If the edge type attribute field of the directed edge is marked as a violation of association edge, then the preset edge weight coefficient is assigned to the first priority abnormal critical weight coefficient; if the edge type attribute field of the directed edge is marked as a normal association edge, then the edge weight coefficient is assigned to the second priority static association weight coefficient.

[0033] The edge weight coefficients of all directed edges in the effective connected path are accumulated to generate the path weighted cumulative value. The power of the path transmission level in the knowledge graph topology based on the effective connected path is used as the denominator, and the path weighted cumulative value is used as the numerator to perform a division operation to obtain the single path logical contribution value.

[0034] Perform a normalized summation operation on the single-path logical contribution values ​​of all valid connected paths in the preliminary candidate set to generate a topological logical correlation value.

[0035] Furthermore, the step of performing backpropagation update on the current parameter set includes:

[0036] The system iterates through the final knowledge recommendation list to extract a set of visual metadata. Based on the set of visual metadata, it retrieves the associated automated execution scripts from the pre-set automated tool library index. It then fills the automated execution scripts with static configuration parameters extracted based on the target device identifier and dynamic operation indicators extracted based on the environmental state feature vector to generate a one-click repair control. Finally, it combines the set of visual metadata and the comprehensive confidence score to perform structured encapsulation and rendering to generate a smart push card.

[0037] In response to user operation behavior data collected from smart push cards to obtain real-time environmental state stream, a comprehensive reward value is calculated by combining a physical verification closed-loop mechanism. Based on the comprehensive reward value, the current parameter set is updated by backpropagation using a policy gradient algorithm to generate a model update parameter set.

[0038] Furthermore, the method for calculating the comprehensive reward value includes:

[0039] Boolean state determination is performed on user operation behavior data and real-time environmental state stream. If the user operation behavior data contains a confirmation execution command, the positive interaction behavior event is determined to be true. If the comparison between the real-time environmental state stream and the environmental state feature vector shows that the abnormal indicators have returned to the preset normal operating value range, the system state has returned to steady state event, which is determined to be true. If the user operation behavior data contains a "close" or "negative evaluation" command, the negative interaction behavior event is determined to be true.

[0040] The truth mapping result of positive interactive behavior events is multiplied by a preset first reward weight coefficient to generate a subjective feedback gain component; the truth mapping result of system state return to steady state events is multiplied by a preset second reward weight coefficient to generate an objective therapeutic effect gain component; the truth mapping result of negative interactive behavior events is multiplied by a preset third penalty weight coefficient to generate an interactive loss component.

[0041] Perform algebraic summation on the subjective feedback gain component, the objective therapeutic effect gain component, and the interaction loss component, and output the comprehensive reward value.

[0042] The AI-based IT service knowledge base intelligent push system is used to implement the above-mentioned AI-based IT service knowledge base intelligent push method. The system includes a knowledge graph construction module, a feature extraction and encoding module, an intelligent knowledge retrieval module, and an intelligent push feedback module.

[0043] The knowledge graph construction module is used to parse and generate a set of operation and maintenance entities and an initial association triplet based on the collected raw data set, construct a knowledge graph topology structure based on the set of operation and maintenance entities and the initial association triplet, and map the knowledge graph topology structure into a graph feature matrix.

[0044] The feature extraction and encoding module is used to discretize and map the real-time captured real-time dialogue text stream into a subjective semantic feature vector, and to traverse the target device identifier pushed by the alarm interface of the external monitoring system based on the knowledge graph topology to generate an environmental state feature vector.

[0045] The intelligent knowledge retrieval module is used to fuse subjective semantic feature vectors and environmental state feature vectors to generate a hybrid query vector, traverse the graph feature matrix based on the hybrid query vector to extract a preliminary candidate set, and filter and sort the preliminary candidate set in combination with the knowledge graph topology to obtain the final knowledge recommendation list.

[0046] The intelligent push feedback module is used to retrieve the pre-set automated tool library index based on the final knowledge recommendation list to encapsulate and generate intelligent push cards, collect user operation behavior data and real-time environmental state stream for intelligent push cards to calculate the comprehensive reward value, and use the policy gradient algorithm to perform backpropagation update on the current parameter set.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] This invention uses cross-modal logical association verification and transitive closure operation technology to reconstruct multi-source heterogeneous operation and maintenance data in a graph-like manner. It establishes explicit logical connections between the static design benchmarks during the construction period and the dynamic fault characteristics during the operation and maintenance period at the semantic and numerical levels, realizes the automatic derivation and completion of implicit dependencies between components, and effectively eliminates semantic silos and logical gaps between data throughout the entire system lifecycle.

[0049] This invention achieves orthogonal encoding of user subjective fault descriptions and objective operating status of IT facilities by simultaneously extracting subjective semantic features of instant messaging text streams and physical environment temporal features based on knowledge graph topology within a unified sliding sampling time window. This effectively compensates for the semantic ambiguity and lack of context caused by the lack of dynamic environment awareness in traditional single-modal retrieval.

[0050] This invention introduces a semantic entropy-driven dynamic gating mechanism and a graph topology-based multi-hop path verification strategy to achieve adaptive weighted fusion and logical reachability verification of the user's subjective fuzzy intent and objective physical operating status. This effectively solves the semantic ambiguity problem caused by the lack of physical context awareness when traditional retrieval technologies face unstructured fuzzy descriptions.

[0051] This invention encapsulates static operation and maintenance knowledge into intelligent push cards with automated execution capabilities, and constructs a closed-loop feedback mechanism by combining user operation behavior and real-time reset status of the physical environment. This enables the operation and maintenance solution to transform from a passive text reference to an active fault repair approach, effectively solving the problems of traditional knowledge bases lacking objective verification of the actual repair efficacy of recommended solutions and being unable to continuously iterate strategies based on real operation and maintenance results. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0053] Figure 1 This is a flowchart illustrating the principle of the AI-based intelligent push method for IT service knowledge bases according to the present invention.

[0054] Figure 2 This is a schematic diagram illustrating the principle of deriving dependent associated edges based on transitive closure operation in this invention.

[0055] Figure 3 This is a schematic diagram illustrating the generation logic of the set of valid connected paths in the knowledge graph of this invention.

[0056] Figure 4 This is a functional module diagram of the AI-based IT service knowledge base intelligent push system of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1:

[0059] Please see Figure 1 As shown, this embodiment provides an AI-based intelligent push method for IT service knowledge bases, including:

[0060] Step S1000, based on the collected raw data set Parse and generate a set of operation and maintenance entities and initial association triples Based on the set of operation and maintenance entities and initial association triples Constructing a knowledge graph topology and the knowledge graph topology Mapped to a graph feature matrix .

[0061] Specifically, this step aims to transform the unstructured historical work order dataset. A semi-structured standard operations and maintenance document set And the test document set during the construction period As the original corpus object, i.e., the original data set By utilizing natural language processing and graph neural network techniques, discretely distributed operational elements in text are mapped into a knowledge graph topology containing nodes and edges. And the topology of this knowledge graph The inherent logical deduction relationships are mapped to a graph feature matrix in a complex vector space. This eliminates the semantic gap between construction phase data and operation and maintenance phase data.

[0062] Further, step S1000 includes:

[0063] Step S1100, based on the collected raw data set Generate standard corpus sequence And the standard corpus sequence is processed by the sequence labeling calculation unit. The mapping is performed on the state feature matrix, and global path constraint optimization is applied to the state feature matrix to lock the globally optimal label path. Based on the globally optimal label path For standard corpus sequences Perform segmentation to generate a set of operation and maintenance entities. and initial association triples .

[0064] Specifically, this step aims to transform the raw data set, which is unstructured and contains privacy noise, throughout the entire IT operations and maintenance lifecycle. As the original digital signal acquisition object, the text signal denoising and feature extraction capabilities of natural language processing technology are used to map the implicit, difficult-to-compute natural semantic fault descriptions and system design baseline parameters in the continuous text stream into a set of machine-computable operation and maintenance entities. And map the potential syntactic dependency features between operation and maintenance entities into initial association triples. This enables the digital translation of unstructured corpora into machine-readable semantic units at the source.

[0065] The original data set Includes unstructured historical work order datasets A semi-structured standard operations and maintenance document set And the test document set during the construction period The historical work order dataset is mentioned above. This involves configuring a database connector to access the underlying database of the IT service management system, and periodically and incrementally extracting unstructured records containing fault description text, processing logs, and classification codes; the standard operation and maintenance document set. These are technical disclosure documents and standard operating procedure manuals retrieved from the enterprise knowledge base server using optical character recognition or document parsing interfaces; the test document set during the construction period. These are system stress test reports and acceptance indicator documents obtained from the project lifecycle management system archive via file transfer protocol.

[0066] In the specific implementation process, firstly, this step establishes a data cleaning pipeline based on natural language processing, which cleans the acquired raw data set. The data signal to be processed is input into this data cleaning pipeline. Pre-defined regular expression matching logic removes non-semantic noise signals such as HTML tags, garbled characters, and system escape characters from the data signal. Simultaneously, for sensitive information contained in the IT operations data, specific entity recognition probes are deployed to scan the data signal after removing non-semantic noise signals, locating sensitive information fragments that pose data privacy risks, such as IP addresses, server hostnames, and user ID fields. Subsequently, the located sensitive information fragments are replaced with predefined general placeholders; for example, the detected specific IP address characters "192.168.1.10" are replaced with the abstract category tag "TOKEN_IP". The character data, after noise filtering and privacy replacement, is reconstructed into a standard corpus sequence while preserving the integrity of the original text's syntactic structure. This allows for the digital translation of raw, unstructured data into machine-readable, standardized input.

[0067] The generated standard corpus sequence The input character sequence is instantiated and fed into a pre-defined sequence labeling computation unit. This unit first activates the BERT pre-trained layer, utilizing its multi-head self-attention mechanism to perform context scanning on the input character sequence, mapping discrete character signals into continuous, high-dimensional deep semantic feature vectors. Subsequently, this deep semantic feature vector is input into a Bidirectional Long Short-Term Memory Network (BiLSTM) layer, which uses its bidirectional gating mechanism to extract the forward and backward long-distance dependencies of the deep semantic feature vectors in the time dimension, generating a state feature matrix corresponding to the time step of the input character sequence.

[0068] Building upon this, to address the issue of edge segmentation errors in unstructured IT texts, such as error codes and proprietary abbreviations, this step utilizes a Conditional Random Field (CRF) layer as a sequence decoder to perform global path constraint optimization on the state feature matrix output by the bidirectional Long Short-Term Memory (LSTM) network layer. The specific execution logic of this global path constraint optimization is as follows: A state selection topology fuzzy map is constructed based on the state feature matrix, orthogonally defined by both the temporal and state dimensions. The temporal dimension corresponds to the standard corpus sequence. The discrete-time step index represents the number of character signals. The state dimension corresponds to a pre-defined set of operation and maintenance entity labels. This set includes several logical state categories, such as "B-Error fault prefix" and "I-Metric indicator term," used to define the semantic segmentation boundaries of operation and maintenance. In the state selection topology graph, all possible directed connected paths that start from the initial state of the first time step, go through logical jumps at each time step, and finally reach the terminal state are traversed and evaluated; these are the candidate label path sets. A globally optimal label path is then searched and locked from this candidate label path set. The globally optimal tag path It is defined as the unique sequence of labels that maximizes the calculated value of the global path score function.

[0069] The calculation logic of the global path scoring function includes a linear superposition of two potential dimensions: local observation potential and state transition potential. The local observation potential represents the values ​​at the temporal and state dimension indices corresponding to the state feature matrix, used to quantify the non-normalized emission probability intensity of the input character signal at each discrete time step index, mapped to a specific operational entity label. The state transition potential represents the row and column values ​​of the state dimension from the pre-stored global transition probability matrix within the sequence decoder, used to quantify the logical dependency weights for the jump from the label state at the previous discrete time step index to the label state at the current discrete time step index. For example, a positive high logical dependency weight is given to the path from the fault word's first label "B-Error" to the fault word's middle label "I-Error" to strengthen the continuity of the fault entity; a negative penalty weight is given to the path from the irrelevant item label "O-Other" to the fault word's middle label "I-Error" to suppress illegal entity truncation that does not conform to IT operational syntax.

[0070] Based on the above globally optimal tag path The definition of dynamic programming recursive calculation in the state selection topology graph is as follows: At each discrete time step, for each candidate state node in that discrete time step, calculate the linear sum of the local observed potential energy value, the preceding optimal recursive potential energy value pointing to the candidate state node, and the state transition potential energy of the candidate state node. Maximize this linear sum to retain a locally optimal sub-path pointer pointing to the current candidate state node. The preceding optimal recursive potential energy value pointing to the candidate state node is the value carried by the candidate state node with the largest path cumulative potential energy value among all previous candidate state nodes that could potentially jump to the current candidate state node in the previous discrete time step. When the recursion reaches the last discrete time step, compare the cumulative potential energy values ​​of all terminating candidate state nodes. Starting from the terminating candidate state node with the largest calculated value of the global path scoring function, perform reverse pointer backtracking along the locally optimal sub-path pointer in the state selection topology graph to lock the globally optimal label path. Based on the globally optimal label path The entity start and end boundary identifiers marked in the standard corpus sequence Physical segmentation is performed, discretizing the continuous text stream into independent string fragments. Then, the globally optimal tag path is read. The system contains entity category tag metadata corresponding to each string fragment, such as "Error", "Metric", and "Action", and instantiates the string fragments into a structured collection of operation and maintenance entities. It conforms to the literal semantic features of the text and strictly follows the construction of standardized and structured operation and maintenance entities in IT operation and maintenance professional terminology, such as complete fault code segments and accurate indicator value segments.

[0071] To achieve logical decoupling and source tracing of multi-source data, this step performs a traffic mapping operation based on entity category label metadata. The specific execution logic is as follows: When the entity category label metadata is identified as fault type Error, the corresponding string fragment is mapped to the historical work order dataset that logically belongs to it. Fault feature entity subset This subset is specifically designed to characterize dynamic abnormal states during IT system operation, such as connection timeouts and memory overflows; when an entity category tag metadata is identified as a metric type, the corresponding string fragment is mapped to a logically belonging test document set. Design benchmark entity subset This subset is specifically used to characterize static design constraints during the IT system construction phase, such as maximum concurrency thresholds and heap memory limits; when entity category tag metadata is identified as an action type, the corresponding string fragment is mapped to a logically belonging standard operations and maintenance document set. A subset of operation and maintenance entities .

[0072] Finally, to prevent the loss of contextual semantics due to entity discretization, this step calls the syntactic dependency analyzer to process the standard corpus sequence. The original sentence fragments containing the extracted structured operation and maintenance entities are scanned to extract the direct syntactic dependency paths of the structured operation and maintenance entities in the original sentence fragments, such as subject-verb relations, verb-object relations, and modification relations. The direct syntactic dependency paths are then structured and encapsulated into initial association triples. This completes the transformation from a one-dimensional linear sequence of semantic tags to atomic units of a knowledge graph with a topological structure.

[0073] Step S1200, based on the set of operation and maintenance entities and initial association triples Instantiate a directed graph data structure, perform cross-modal logical association checks on the directed graph data structure to identify violating association edges, and derive dependent association edges by using transitive closure operations to complete the derivation, thus reconstructing the knowledge graph topology. .

[0074] Specifically, this step aims to integrate the set of operation and maintenance entities that are discretely distributed in step S1100 and originate from different spatiotemporal dimensions such as the construction phase and the operation and maintenance phase. and initial association triples As discrete semantic data objects, leveraging the topological connectivity characteristics of computer graph data structures, the numerical constraint violation judgment results between static design benchmark entity and dynamic fault feature entity are mapped to violation association edges in the graph. Furthermore, the implicit logical transitivity relationships between entities are mapped to deductive subordinate association edges, generating a knowledge graph topology structure with complete semantic paths. This enables the topological reconstruction of discrete data points into a global knowledge network at the logical level.

[0075] In the specific implementation process, firstly, this step initializes an empty directed graph data structure, i.e., the knowledge graph topology, in the graph database storage unit of computer memory. Next, the node instantiation mapping operation is performed: traversing the input set of operation and maintenance entities. This set of operation and maintenance entities Each discrete operation and maintenance entity object in the process is instantiated as a knowledge graph topology. The unique node object in the knowledge graph is identified as the knowledge graph node; and the attribute information set extracted along with the operation and maintenance entity object is written into the data attribute field of the knowledge graph node. Specifically, the attribute information set includes a unique identifier for the source file identifying the origin of the operation and maintenance entity object and an entity type label used to distinguish the business attributes of the operation and maintenance entity object. Simultaneously, a topological connection operation for edge relationships is performed: traversing the initial association triples of the input... For each association triple containing a head entity, a tail entity, and a relation type, a directed edge is created between the existing head entity node and the tail entity node, and the relation type is written into the edge type attribute field of the directed edge, thus completing the knowledge graph topology. The construction of.

[0076] Building upon this, to break down data silos between construction phase documents and operation and maintenance phase logs, this step performs cross-modal logical association verification based on both semantic and numerical validation. The specific execution logic is as follows: Traverse the subset of fault feature entities. Each fault characteristic entity and a subset of design baseline index entities in the table. For each design benchmark entity in the dataset, for any pair of entities to be verified consisting of the fault feature entity and the design benchmark entity, calculate the association matching degree between them. If and only if the pair of entities to be verified simultaneously satisfies both semantic attribute consistency verification and numerical constraint violation verification, instantiate a violation association edge between the fault feature entity and the design benchmark entity between the pair of entities to be verified.

[0077] The semantic attribute consistency verification is performed by using a string similarity algorithm, such as the Jaccard similarity coefficient, to calculate the consistency of historical work order datasets for fault feature entities. The fault object name field parsed from the test document set from the design baseline index entity The text similarity between the parsed benchmark indicator name fields is calculated. If the text similarity is greater than a preset semantic similarity threshold, it is determined that the two semantically refer to the same physical IT resource object. The semantic similarity threshold is a floating-point constant pre-set in the system configuration items, used to filter out invalid entity pairs to be checked during semantic attribute consistency verification that have significantly different names and descriptions and clearly do not belong to the same monitoring object, thereby reducing the resource overhead of subsequent numerical calculations.

[0078] The numerical constraint violation check is performed by analyzing fault characteristic entities in the historical work order dataset. Extract runtime observation values ​​from the corresponding unstructured log text stream, and extract them from the design baseline metric entity in the test document set. The design limit values ​​and logic comparison operators are extracted from the corresponding structured parameter table. Then, the three extracted parameters are input into the logic judgment function. Binary logic operations are performed to verify whether the quantitative relationship between dynamic runtime observed values ​​and static design limit values ​​satisfies the violation conditions defined by the logic comparison operator. If the logic judgment function... If the calculation result is True, the current operating state is determined to violate the preset design baseline in mathematical logic. A violation association edge "Violation_of" is then established and stored between the fault feature entity and the design baseline indicator entity. Conversely, if the calculation result is False, no violation is determined, and no violation association edge is established. The runtime observed values ​​are discrete floating-point or integer data variables extracted from the log text stream of the fault feature entity using a regular expression algorithm. These variables are used to quantify the dynamic physical state of IT components in the operating environment, such as the instantaneous CPU utilization value of 85%. The design limit values ​​are derived from the test document set during the construction phase. The static baseline constants parsed and stored in the middle are used to define the numerical boundary constraints for the normal operation of IT components, such as the safe upper limit of CPU utilization of 80%; the logical comparison operators are relational operators defined in the entity attributes of the design baseline index, and their values ​​include standard mathematical symbols such as greater than, less than, greater than or equal to, less than or equal to, etc., which are used to specify the directional logical rules for triggering the determination of violations of observed values ​​during runtime.

[0079] Ultimately, in order to mine the topology of the knowledge graph This step addresses deep dependencies not directly manifested in the existing knowledge graph topology. The rule-based transitive closure operation is performed on the knowledge graph topology. The specific execution logic of the transitive closure operation is as follows: a set of pre-defined reasoning rules is defined in memory, and the knowledge graph topology is... Perform a path scan on any three knowledge graph nodes. If the scan reveals a knowledge graph topology... The scan reveals a first directed edge connecting knowledge graph node A to knowledge graph node B, with the relationship type recorded as "Deployed_On" in the edge type attribute field. Simultaneously, a second directed edge connecting knowledge graph node B to knowledge graph node C is found, with the relationship type recorded as "Belongs_To" in the edge type attribute field. Based on the principle of logical transitivity, it is determined that there is an implicit logical connection between knowledge graph node A and knowledge graph node C within the knowledge graph topology. In the first step, a directed edge is instantiated and inserted, directly pointing from knowledge graph node A to knowledge graph node C. This is called an inferred dependency edge, and the relation type field of the edge type attribute is marked as "Inferred_Belonging". This process is iterated through multiple rounds using the transitive property principle until no new directed edges are generated after the knowledge graph is traversed and scanned. Finally, the complete knowledge graph topology is output. This enables the logical closed-loop completion of relationships between discrete IT architectures.

[0080] Further, please refer to Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the principle of deriving dependent and related edges based on transitive closure operation in this invention. For example... Figure 2 As shown, in the basic knowledge graph topology The diagram illustrates three exemplary nodes: knowledge graph node A (black), knowledge graph node B (purple), and knowledge graph node C (green). Knowledge graph node A is connected to knowledge graph node B via a solid directed edge marked "Deployed at"; similarly, knowledge graph node B is connected to knowledge graph node C via a solid directed edge marked "Belongs to". The gear icon at the top of the diagram represents a pre-defined set of inference rules. These rules are used to logically scan and verify the connection paths and edge attributes between the three knowledge graph nodes. When a pre-defined transitivity condition is met, the knowledge graph topology... Instantiate and insert a thick dashed directed edge from knowledge graph node A to knowledge graph node C. This dashed directed edge is defined as a deduced dependency association edge, and its edge type attribute field is recorded as "deduced dependency", which is used to establish the implicit logical dependency relationship between knowledge graph node A and knowledge graph node C.

[0081] For example, taking the fault diagnosis scenario of an order system as an example, assume that the test document set during the construction period is... The extracted design benchmark entity is labeled as knowledge graph node C. The core attribute of knowledge graph node C defines "the maximum number of connections to the order database is 2000". Simultaneously, the historical work order dataset from the operation and maintenance period is... The extracted fault feature entity is labeled as knowledge graph node A. This knowledge graph node A contains the text description of the dynamically observed value "The current number of connections for the order service is 2500, and the response timed out." During the construction of logical associations, semantic attribute consistency checks identify that "current connection for the order service" in knowledge graph node A and "maximum number of connections for the order database" in node C point to the same physical object. Numerical constraint violation checks extract numerical inputs to the logical judgment function. Performing binary logic operations, since 2500 is greater than 2000, it satisfies the violation condition of the logic comparison operator being greater than. The logic judgment function... The calculation result is true. Based on this, the system automatically establishes a violation association edge between knowledge graph node A and knowledge graph node C. Further, assuming knowledge graph node A is deployed on knowledge graph node B representing host resources, and knowledge graph node B belongs to knowledge graph node D representing the transaction subsystem, the system automatically derives the dependent association edge through transitive closure operations. Finally, when operations personnel retrieve faults related to the "transaction subsystem," they quickly locate knowledge graph node A by following the derived dependent association edge, and continue tracing along the violation association edge to the root cause node, i.e., knowledge graph node C, thus directly arriving at the diagnostic conclusion that the transaction subsystem fault is caused by an excessive number of connections.

[0082] Step S1300, based on the knowledge graph topology Instantiate the initial state of the graph feature matrix The execution parameter iterative optimization logic optimizes the graph feature matrix of the initial state. Perform backpropagation to update and obtain the graph feature matrix. .

[0083] Specifically, this step aims to transform the discretely distributed knowledge graph topology from step S1200. As a topological data source, it utilizes geometric mapping logic within the vector space to transform the knowledge graph topology. Each knowledge graph node is mapped to a position vector in a multidimensional semantic vector space, and the edge type attribute fields between knowledge graph nodes are mapped to logical orientation deflection operators. This mathematically achieves a feature mapping from a discrete symbolic logic space to a continuous geometric vector space, providing a machine-computable graph feature matrix while preserving the original graph's logical deduction capabilities such as transitivity and symmetry. .

[0084] In the specific implementation process, firstly, this step allocates and initializes a dimension in the computer memory. A multidimensional semantic vector space. Traversing the knowledge graph topology. For each discrete knowledge graph node and directed edge, a randomly initialized semantic feature state vector is assigned in the multidimensional semantic vector space. For each directed edge's corresponding edge type attribute field, a randomly initialized relational logic mapping vector is assigned in the same multidimensional semantic vector space. All assigned semantic feature state vectors and relational logic mapping vectors are then structurally aggregated, assembled, and instantiated into an initial state graph feature matrix. This matrix was established as the set of parameters to be optimized in subsequent iterative calculations.

[0085] In order to preserve the topological structure of the knowledge graph in the multidimensional semantic vector space The logical deduction capability, especially transitivity, is crucial. For example, if knowledge graph node A points to knowledge graph node B, and knowledge graph node B points to knowledge graph node C, then there is a logical rule between knowledge graph node A and knowledge graph node C. This step employs geometric mapping logic based on logical orientation bias. The specific execution logic is as follows: Read the graph feature matrix already stored in the initial state. The semantic feature state vector and relational logical mapping vector are used as the objects of operation. Specifically, the relational logical mapping vector representing the logical relationship is defined as: in the multidimensional semantic vector space, the logical orientation deflection operator maps the semantic feature vector of the head entity to the semantic feature vector of the tail entity. The semantic feature vector of the head entity is the semantic feature state vector used to represent the starting point, and the semantic feature vector of the tail entity is the semantic feature state vector used to represent the ending point. The relational logical mapping vector is a complex vector of unit modulus in the complex vector space, so the Hadamard product operation in the complex field is equivalent to performing a pure phase rotation on each dimension of the semantic feature vector of the head entity.

[0086] The physical meaning of this process is that in an ideal multidimensional semantic vector space, each semantic feature state vector occupies a fixed semantic anchor point coordinate. When the head entity semantic feature vector and the relational logical mapping vector perform a Hadamard product operation, it is equivalent to forcing each dimensional component of the head entity semantic feature vector to undergo a directional rotation transformation according to the logical deflection angle defined in the relational logical mapping vector. For example, for reciprocal relationships in IT architecture, such as "Server A hosts application B" and "Application B is deployed on server A," they represent opposite rotational movements in geometric space, that is, clockwise rotations respectively. Angle and counterclockwise rotation For symmetrical relationships in IT architecture, such as "dual-machine hot standby," the transformation is represented by a 180-degree rotation. This means that a vector rotates once to reach the opposite end, and then rotates by the same angle again to return to its initial position. Through this logical directional deflection, the semantic feature vector of the head entity, after undergoing a directional rotation transformation based on the Hadamard product, can point to the semantic anchor point coordinates of the semantic feature vector of the tail entity, thereby achieving the coincidence of logical relationships in geometric space.

[0087] This step constructs and executes parameter iterative optimization logic based on geometric semantic distance to generate the final graph feature matrix. The specific execution logic of the parameter iterative optimization logic is as follows: The initial state's graph feature matrix... As a parameter update object. In each iteration of computation, for the knowledge graph topology... For any association triplet consisting of a head entity node, a relation type, and a tail entity node, the semantic deviation of this association triplet in the multidimensional semantic vector space is quantified, i.e., the geometric semantic distance difference, based on the geometric potential difference calculation logic. The specific execution logic of the geometric potential difference calculation logic is as follows: calculate the Euclidean distance between the predicted coordinate vector obtained after the head entity semantic feature vector undergoes a directional rotation transformation defined by the relation logic mapping vector and the actual tail entity semantic feature vector. With minimizing the geometric semantic distance difference as the optimization objective, the stochastic gradient descent algorithm is used to optimize the initial state of the graph feature matrix. The numerical values ​​of the elements in the matrix are updated through backpropagation. As the number of iterations increases, the iteration stops when the geometric semantic distance difference converges to a preset convergence criterion. At this point, the set of vectors stored in memory that have achieved numerical convergence constitutes the final output graph feature matrix. .

[0088] Step S2000: The real-time captured dialogue text stream Discretization mapping to subjective semantic feature vectors And based on the knowledge graph topology Target device identifier pushed to the alarm interface of the external monitoring system Perform a traversal to generate environment state feature vectors .

[0089] Specifically, this step aims to transform the unstructured, streaming real-time conversational text streams in instant messaging collaboration platforms. and target device identifiers with logical topological dependencies in the IT infrastructure. As the original data collection object, combined with the knowledge graph topology from step S1200 By leveraging the modal differences between the natural language semantic space and the physical operational signal space, the sliding sampling time window is... The subjective intent description of the user is mapped to a subjective semantic feature vector. The runtime timing metrics of the target device and its topological neighboring devices at the physical layer are mapped to environmental state feature vectors. This enables the orthogonal extraction of semantic and physical information.

[0090] Further, step S2000 includes:

[0091] Step S2100: The real-time captured dialogue text stream In the sliding sampling time window Discretize the internal text into an effective text sequence Based on the mapping relationships in the pre-built IT operations and maintenance semantic parsing unit, the effective text sequence will be... The mapping is performed on a sequence of feature state matrices, and full-sequence information compression is performed on the feature state matrix sequence to generate subjective semantic feature vectors. .

[0092] Specifically, this step aims to process the unstructured, streaming real-time conversational text generated by operations and maintenance personnel when a fault occurs. As the original signal source, it follows a preset sliding sampling time window. Real-time dialogue text stream Discretize and slice the text, transforming it into a valid text sequence that can be processed by a computer. The IT operations and maintenance semantic parsing unit is used to extract the cleaned and effective text sequences. The latent, non-standardized user description characters are mapped to subjective semantic feature vectors. This enables digital quantization from the natural language symbol space to the computer-computable vector space at the data source.

[0093] In the specific implementation process, this step first uses the application programming interface of an instant messaging collaboration platform that has established a data connection with the IT operations and maintenance management system to listen to and capture the real-time text stream of conversations generated by operations and maintenance personnel in the fault handling group. To address the issue of vague end-user descriptions and difficulty in actively associating them with the work order context, a sliding sampling time window is established that maintains frequency synchronization with the physical monitoring cycle of the underlying IT infrastructure. This window advances forward at fixed steps on the timeline to capture real-time conversational text streams. All dialogue message records falling within the current sampling time interval are processed. For the captured dialogue message records, a regular expression-based filter is applied for noise reduction, removing noisy data that does not contain substantial fault description information. Specifically, this includes removing emoji encodings, consecutive non-semantic punctuation strings, and common stop words, thereby obtaining a structured, cleaned, and effective text sequence. The regular expression-based filter is a data cleaning function based on a pre-set set of regular expressions and an IT-related stop word library.

[0094] Subsequently, the cleaned valid text sequence The input is fed into a pre-configured IT operations and maintenance semantic parsing unit. This unit is a dedicated semantic feature extraction device initialized based on historical work order logs, standard operating manuals, and error code libraries from the IT service management domain. Within the logical layer of this parsing unit, a pre-configured IT operations and maintenance domain terminology association rule base stores mapping coefficients between user fuzzy descriptions and standard technical terms in the IT operations and maintenance management system. Valid text sequences are calculated using this IT operations and maintenance domain terminology association rule base. The semantic topology projection values ​​between each discrete character data node and specific operational concepts such as "database deadlock", "service restart", and "memory overflow" are used to generate a feature state matrix sequence containing feature information of all character data nodes based on these semantic topology projection values.

[0095] Building upon this foundation, to centralize the effective information scattered across natural language descriptions of varying lengths, this step performs full-sequence information compression. First, a character data node located at the first address of the physical storage space within the feature state matrix sequence is identified and defined as a global semantic node. This global semantic node does not carry the meaning of a specific character but serves as a data container for the overall fault description intent. The semantic association strength distribution coefficients of each discrete character data node (excluding the first one) relative to this global semantic node are calculated. These coefficients are obtained by performing a dot product operation between the feature vectors of each character data node and the feature vector of the global semantic node, followed by Softmax normalization. The initial feature vector of the global semantic node is taken from the output vector corresponding to the special classification label at the first position of the pre-trained language model's input sequence. These semantic association strength distribution coefficients characterize the contribution of each discrete character to determining the overall fault type. Based on these coefficients, the local semantic information scattered across the discrete character data nodes is weighted, aggregated, and superimposed onto the global semantic node, achieving information flow and compression from local to global. Subsequently, the feature coordinate set stored within the global semantic node is directly accessed and read. This feature coordinate set is an ordered set of real numbers formed after weighted aggregation. This feature coordinate set is directly defined as representing the current sliding sampling time window. Subjective semantic feature vector of the user's subjective fault description intent .

[0096] Step S2200: Responding to the target device identifier pushed by the alarm interface of the external monitoring system. In the knowledge graph topology Perform a breadth-first search to locate the set of related dependent components. During the sliding sampling time window Internal collection of related dependent components The operational metrics data are based on the sliding sampling time window. Related dependent component collection Combine operational metrics data to generate a time-series operational status matrix .

[0097] Specifically, this step aims to identify target device identifiers that are isolated and distributed throughout the IT infrastructure. As a topological positioning anchor point in physical space, the knowledge graph topology structure from step S1200 is utilized. The logical connection attributes in the data define the potential fault propagation range in the logical space and drive the data acquisition engine to measure the data of all IT component nodes within that fault propagation range within a sliding sampling time window. The system captures resource consumption signals and business response signals generated internally in parallel, and finally aggregates discrete single-point monitoring indicators into a time-series operational status matrix that can characterize the overall local health. This enables a dimensional upgrade at the physical perception level, moving from single-point monitoring to collaborative detection of related and dependent environments.

[0098] In the specific implementation process, this step first performs a topology structure retrieval operation. The specific execution logic is as follows: using the target device identifier pushed by the alarm interface of the external monitoring system. As the starting root node for breadth-first search, in the knowledge graph topology... A breadth-first search is performed. Based on a preset topology association depth threshold, upstream dependent nodes with a direct logical connection to the starting root node are identified, such as database servers providing data support for the business system. Downstream affected nodes whose state fluctuates with the starting root node are also identified, such as web application servers that depend on data interfaces. The topology association depth threshold is a pre-configured integer logical constraint constant used to define the target device identifier. The maximum logical hop count radiating outwards from the center limits the fault propagation scope to local dependency environments where cascading effects might occur. All the locked nodes are aggregated to construct a set of dependent components containing the fault source and its directly affected scope. The topology retrieval operation expands the scope of troubleshooting from a single physical device to a cluster of logically related components that include its entire dependent environment.

[0099] Determine the set of related dependent components Next, this step starts the concurrent data acquisition engine. This engine utilizes a pre-configured monitoring interface adapter, within a sliding sampling time window preset in step S2100. Within the same defined physical time base, for the set of related dependent components Each IT component node executes status polling requests in parallel. The operational metrics data returned by the status polling requests specifically cover two types of key physical characteristic signals: first, resource consumption physical signals, which characterize the load of the underlying computing capabilities of the IT components, specifically including the instruction cycle utilization of the central processing unit and the address space utilization of random access memory; second, service response physical signals, which characterize the quality of services provided by the IT components, specifically including the round-trip latency of data packets at the network transport layer and the transaction count per second at the application layer.

[0100] Finally, the discrete indicator values ​​collected above and distributed across various IT component nodes are structurally reorganized. Tensor reorganization is performed according to the logical order of time, device, and indicator dimensions to construct a time-series runtime state matrix. The time dimension corresponds to the sliding sampling time window. Each consecutive sampling time within; the device dimension corresponds to the set of associated dependent components. Each IT component node in the matrix; the metric dimension corresponds to the specific type index of the collected resource consumption physical signals and business response physical signals. Time-series operational status matrix. Each row corresponds to a sampling time, and each column corresponds to the physical index value of the IT component node in a specific dimension, thus forming a holographic digital snapshot of the dynamic operating environment within the current fault-related impact domain.

[0101] Step S2300, process the timing operation state matrix. Perform Z-score normalization and first-order difference operations to generate dimensionless numerical values. Time series fluctuation rate Numerical operation based on dimensionless quantities and time series fluctuation rate Construct a joint feature input vector and input it into a multilayer perceptron network to generate an environmental state feature vector. .

[0102] Specifically, this step aims to transform the time-series operational state matrix, which is discretely distributed and in a dynamic flow state from the IT operations and maintenance environment in step S2200, into a matrix. As a digital signal source, statistical transformations are used to eliminate measurement differences between different physical dimensions, and the sliding sampling time window is used. The underlying, static physical properties of internal components are mapped to dimensionless operational values ​​based on a unified benchmark. and the dimensionless quantity will run numerically. The dynamic characteristics that evolve over time are mapped to the temporal fluctuation rate. Finally, an environmental state feature vector is generated through a feature perception mapping operation based on a multilayer perceptron. This enables the digital translation of physical dimension information and dynamic evolution information at the data source.

[0103] In the specific implementation process, this step first starts the data preprocessing engine to process the time-series running state matrix. Perform feature normalization and situational awareness operations. The specific execution logic of the feature normalization and situational awareness operations is as follows: For the time-series running state matrix... Each physical metric value in the dataset undergoes Z-Score normalization to eliminate interference from differences in physical units and their magnitudes, such as CPU utilization (percentage %), memory usage (gigabytes, GB), and network latency (milliseconds, ms), affecting subsequent calculations. This Z-Score normalization maps the physical metric values ​​to a standard normally distributed numerical space, generating dimensionless values ​​that characterize the degree to which the current metric deviates from the baseline. Meanwhile, in order to capture the dynamic evolution of fault occurrences, such as distinguishing between network jitter and persistent congestion, a sliding sampling time window is used. Within a defined continuous time series, the numerical values ​​of dimensionless quantities are processed. Performing a first-order difference operation involves calculating the difference between the dimensionless quantity at the current sampling time and the dimensionless quantity at the previous sampling time to extract the rate of change of the dimensionless quantity over time, thereby generating the time-series fluctuation rate characterizing the acceleration of system state deterioration or recovery. .

[0104] In obtaining the dimensionless quantity operation value and time series fluctuation rate Next, this step performs a feature perception mapping operation based on a multilayer perceptron. The specific execution logic is as follows: using the vector concatenation operator, the set of associated dependent components is... Dimensionless running values ​​corresponding to all IT component nodes in the system and time series fluctuation rate Channel-level aggregation is performed to construct a joint feature input vector. This joint feature input vector is then fed into a pre-built multilayer perceptron network. This network, acting as a nonlinear feature encoder, utilizes its internal feature weight matrix and bias parameters to perform weighted combination and nonlinear spatial transformation on the input joint feature vector, projecting the discrete numerical values ​​of the physical space into a latent feature space vector, i.e., the environmental state feature vector. .

[0105] Step S3000: Fuse subjective semantic feature vectors and environmental state feature vector To generate mixed query vectors Based on hybrid query vectors Traversing the spectral feature matrix To extract the initial candidate set And combined with the knowledge graph topology For the preliminary candidate set After filtering and sorting, the final knowledge recommendation list is obtained. .

[0106] Specifically, this step aims to transform the subjective semantic feature vector from step S2100... With the environmental state feature vector from step S2300 As query input, the spectral feature matrix from step S1300 is used. and the knowledge graph topology from step S1200 As a retrieval probe, a dynamic gated attention arbitration unit is introduced to generate a hybrid query vector that can express the requirements of the current operation and maintenance scenario. Subsequently, a complex similarity measurement operation is performed in the vector space to obtain an initial candidate set. and with the target device identifier received from step S2200 Using physical anchors, based on the knowledge graph topology For the preliminary candidate set Perform interpretability reordering to output the final knowledge recommendation list sorted by confidence. This enables an intelligent leap from fuzzy queries to explainable recommendations.

[0107] Further, step S3000 includes:

[0108] Step S3100, for the subjective semantic feature vector Execution probability distribution normalization to calculate semantic uncertainty index The semantic uncertainty index is determined by using a dynamic gating attention arbitration unit. The mapping is performed as semantic feature weight coefficients, and the subjective semantic feature vector is adjusted based on these semantic feature weight coefficients. and environmental state feature vector Perform linear weighted fusion to generate a hybrid query vector. .

[0109] Specifically, this step aims to transform the subjective semantic feature vector from step S2100... The environmental state feature vector from step S2300 is considered as a semantic signal carrying subjective uncertainty. Treating the objective physical state signal as an objective physical state signal, the intent certainty of the subjective semantic signal is quantified using the Shannon information entropy principle in information theory. A dynamic gated attention arbitration unit automatically adjusts the weight ratio of both signals in the final query representation, thereby generating a hybrid query vector that maps the essence of the current fault. This enables a digital synthesis that moves from fuzzy queries in a single dimension to holographic queries that complement both the subject and object.

[0110] In the specific implementation process, in order to quantitatively measure at the numerical level whether the natural language description input by the user is clear and unambiguous in the semantic space, that is, to distinguish between deterministic technical identifiers with feature convergence characteristics, such as specific error status codes, and unstructured fuzzy descriptions with feature divergence characteristics, such as discrete fault phenomenon descriptions, this step first initiates the semantic deterministic evaluation unit. The specific execution logic of the semantic deterministic evaluation unit is as follows: using subjective semantic feature vectors... As the computational object, its probability distribution is normalized for each dimensional component, for example, by using the Softmax function to transform it into a probability density distribution in the feature space, obtaining the normalized feature probability value corresponding to each dimensional component. Subsequently, using the Shannon information entropy principle, the semantic uncertainty index is calculated based on the normalized feature probability values. This transforms the feature distribution discreteness within a high-dimensional vector into a measurable scalar value. The semantic uncertainty index... The magnitude of the value is positively correlated with the ambiguity of the intent; the larger the value, the more evenly and dispersed the feature vector corresponding to the user description is distributed in the semantic space, that is, the more ambiguous the search intent is; conversely, the smaller the value, the more peaked the feature vector is in a specific dimension, that is, the more precise the search intent is.

[0111] The semantic uncertainty index was calculated. Next, this step constructs and runs a dynamic gated attention arbitration unit based on a deep neural network. This unit acts as a confidence controller for multimodal feature fusion, and its specific execution logic is as follows: Using a tensor concatenation operator, the subjective semantic feature vector is... Its corresponding semantic uncertainty index A concatenation operation is performed along the feature channel dimension to construct a joint input feature tensor. This joint input feature tensor is then fed into a pre-defined gated perception layer. This gated perception layer uses an internally pre-trained, converged gated perception weight matrix and gated bias term vector to perform a linear weighted transformation on the input joint input feature tensor, and maps the transformation result to a sigmoid non-linear activation function. The numerical range is determined, thus outputting a dimensionless semantic feature weight coefficient. This semantic feature weight coefficient is a dimensionless scalar floating-point value whose value range is constrained to the open interval (0,1). The gated perception weight matrix and the gated bias term serve as the current parameter set. The components are updated via end-to-end backpropagation within the policy gradient framework of step S4200. A negative correlation mapping mechanism is used to establish the relationship between the ambiguity of the natural language description and the retrieval system's dependence on text features: when ambiguity in the user description is detected, i.e., semantic uncertainty index... When the value is large, the semantic feature weight coefficient is automatically reduced to attenuate the influence of the text vector on the final mixed query vector. The projection amplitude in the text vector is adjusted; conversely, when the user description is precise, the semantic feature weight coefficient is automatically increased to enhance the dominant role of the text vector.

[0112] After obtaining the semantic feature weight coefficients, this step performs a multimodal feature fusion operation. Based on the generated semantic feature weight coefficients, the subjective semantic feature vector is... and environmental state feature vector Complementary linear weighted summaries are applied to generate a hybrid query vector that incorporates the physical context. The specific execution logic is as follows: First, the semantic feature weight coefficients are used to adjust the subjective semantic feature vector. Perform scalar multiplication to obtain the confidence-weighted semantic component vector; simultaneously, calculate the difference between 1 and the semantic feature weight coefficients to obtain the environment state weight coefficients, and use these environment state weight coefficients to weight the environment state feature vector. Scalar multiplication is performed to obtain a complementary weighted state component vector. Subsequently, a vector addition operator is used to element-wise superimpose the semantic component vector and the state component vector along their corresponding feature dimensions. Through this operation, when the semantic feature weight coefficients are relatively small, i.e., in scenarios with ambiguous user descriptions, a hybrid query vector is synthesized. The geometric orientation in the vector space will adaptively point towards the environmental state feature vector. By aligning with objective physical indicators, such as database deadlock characteristics, the search direction can be guided when users cannot express themselves accurately, thus automatically locking in the solution to the problem.

[0113] Step S3200, from the spectral feature matrix Extract candidate solution entity vectors Perform a complex similarity metric operation to compute the mixed query vector. and candidate solution entity vectors The semantic cosine similarity values ​​are used to rank and filter the results to construct a preliminary candidate set. .

[0114] Specifically, this step aims to combine the mixed query vector from step S3100. As a retrieval probe in high-dimensional space, it is projected onto the map feature matrix from step S1300. In the defined multidimensional semantic vector space, quantized hybrid query vector The degree of overlap between the projections of entity vectors of all standard solutions in the knowledge base and their semantic connotation magnitudes and logical relation phases is used to quickly screen out a preliminary candidate set with high confidence from a massive knowledge base. This provides a convergent search space for subsequent steps.

[0115] In the specific implementation process, this step first calls the graph feature matrix stored in the system memory unit. And traverse the feature matrix of the graph The entity type labels of knowledge graph nodes are used to filter and extract the semantic feature state vectors corresponding to all knowledge graph nodes identified as "solution" or "knowledge entry" to construct a target semantic vector set for retrieval. Subsequently, the mixed query vectors are... The vector is projected into the multidimensional semantic vector space initialized in step S1300. During this projection process, the hybrid query vector... It is instantiated as a semantic retrieval probe in the semantic anchor coordinate system, enabling it to have the same coordinates as the graph feature matrix. The vector is mathematically isomorphic to the magnitude of the fault semantic intensity and the phase angle attribute, which characterizes the direction of fault logical evolution, thereby ensuring the hybrid query vector. It possesses geometric projection consistency and measurability with the pre-stored semantic feature state vectors in the target semantic vector set to be retrieved within a continuous geometric space.

[0116] Iterate through the set of target semantic vectors to be retrieved, and define each independent semantic feature state vector indexed in the set as a candidate solution entity vector. For candidate solution entity vectors , to use mixed query vectors As a semantic retrieval probe, a complex similarity measurement operation is performed. The specific execution logic of the complex similarity measurement operation is as follows: First, the conjugate transpose operator is called to perform a complex similarity measurement on the entity vectors of candidate solutions. Perform transpose and conjugate transformations to generate the corresponding conjugate transpose vector. Then, use the vector dot product operator to calculate the mixed query vector. The product with the conjugate transpose vector is performed as a Hermitian inner product operation, and the real part extraction operator is invoked to extract the real component of the inner product result. This real component physically represents the effective projection magnitude of the query intent in the logical direction of the solution. Finally, the hybrid query vector is calculated separately. With candidate solution entity vector The vector norm, i.e., the magnitude, is used as the normalization factor, and the effective projection magnitude is divided to obtain the normalized semantic cosine similarity value. This calculation logic aims to quantify the comprehensive geometric consistency of two complex vectors in terms of semantic strength and logical evolution direction at the physical level.

[0117] Based on the calculated semantic cosine similarity scores, this step performs result sorting and filtering logic. The specific execution logic is as follows: using the semantic cosine similarity scores as the sorting key, all candidate solution entity vectors contained in the target semantic vector set to be retrieved are sorted. The system performs a descending numerical sort to generate an ordered entity index sequence. Then, using a pre-set candidate retention threshold as the filtering boundary, it sequentially extracts the top-ranking candidate solution entity identifiers from the entity index sequence, thereby constructing a preliminary candidate set. This set physically converges with hybrid query vectors in a multidimensional semantic vector space. The target solution with the highest geometric projection consistency is the one that filters out the mixed query vectors that are consistent with both the fault semantic strength and the logical evolution direction. A highly compatible set of optimal solutions.

[0118] Step S3300, using the target device identifier Retrieve the topology of the knowledge graph for physical starting anchor points Preliminary candidate set of the middle link The effective connected paths are used to perform topological association metric calculations to generate topological logical association scores, and then combined with semantic cosine similarity scores to perform linear weighted fusion to generate a comprehensive confidence score. And based on the overall confidence score For the preliminary candidate set Reorder the results to output the final knowledge recommendation list. .

[0119] Specifically, this step aims to transform the knowledge graph topology structure containing knowledge graph nodes and directed edges from step S1200. As a logical constraint network in the physical world, it uses the target device identifier received from step S2200. As physical anchor points, for the preliminary candidate set from step S3200 Each candidate solution entity vector in Perform multi-hop path reachability verification, quantify the topological logical relevance from the source of the fault to the potential solution using multi-hop path retrieval operations, and then linearly weight and fuse this topological logical relevance value with the semantic cosine similarity value from step S3200 to generate a comprehensive confidence score. This process eliminates semantically relevant but physically unreachable illusory noise, resulting in a final knowledge recommendation list that conforms to the logical facts of the current operational environment. .

[0120] In the specific implementation process, firstly, this step performs topology path search and reachability verification operations. The specific execution logic is as follows: First, the knowledge graph topology structure is loaded into the computer memory space. and with the target device identifier As the physical starting anchor point for path search, with an initial candidate set Each candidate solution entity vector traversed in the middle This serves as the logical termination node for path searching. Subsequently, a breadth-first search algorithm is invoked within the knowledge graph topology. A multi-hop path retrieval operation is performed to search for the existence of a set of valid connected paths linking the physical start anchor point and the logical end node. If, after traversal and retrieval, it is confirmed that no valid connected path exists, then the candidate solution entity vector is determined. Although semantically similar to hybrid query vectors Having numerical similarity, but lacking physical reachability in the IT component dependency topology, and unrelated to the currently faulty device, a logical blocking operation is performed and the entity vector of the candidate solution is automatically blocked. Implement set removal or confidence level reduction processing.

[0121] Further, please refer to Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the generation logic of the set of valid connected paths in the knowledge graph of this invention. For example... Figure 3 As shown in the diagram, the icon on the left represents the physical starting anchor point, i.e., the target device identifier, which serves as the starting point for path retrieval. The dashed box on the right side of the diagram represents the preliminary candidate set. This includes entity vectors of multiple candidate solutions that serve as logical termination nodes. A breadth-first search algorithm is invoked within the knowledge graph topology. Multi-hop path retrieval is performed. For paths that can be successfully connected, solid and dashed lines originating from the physical starting anchor point and passing through several intermediate nodes to reach logical termination nodes 1, 2, and 3 are used in the diagram to represent valid connected paths with physical reachability in the topology, forming a set of valid connected paths. Conversely, for paths that cannot be connected, a dashed line pointing to logical termination node 4 and a red cross mark on it are used to represent that logical termination node 4 lacks physical reachability in the topology, i.e., no valid connected path exists.

[0122] For each valid connected path that has been verified by previous reachability, this step determines the path's connectivity based on the edge type attribute field of the directed edges distributed along the path and the path's position within the knowledge graph topology. The path traversal sequence is used to perform topological association metric calculation. The specific execution logic of the topological association metric calculation is as follows: traverse each directed edge in the currently processed valid connected path, identify and read its preset edge weight coefficient. To quantify the decisive role of dynamic numerical monitoring of violations relative to static ordinary dependencies in root cause diagnosis at the computational level, this step establishes a nonlinear weight mapping strategy for the edge weight coefficients. The following conditional amplitude is applied: If the edge type attribute field of the currently traversed directed edge is marked as violating the association edge "Violation_of", that is, the knowledge graph node has an abnormal state where the dynamic numerical index violates the preset benchmark, such as the database connection number index overflowing, then the edge weight coefficient of the directed edge is adjusted. The algorithm assigns a preset first-priority anomaly key weight coefficient to mathematically enhance its sensitivity in capturing the root cause propagation path of faults. Conversely, if the currently traversed directed edge is a normal associative edge, for example, if the edge type attribute field of the directed edge identifies it as belonging to a static topological relationship such as "Belongs_to", then the edge weight coefficient of that directed edge is... A preset static association weight coefficient with second priority is assigned. Then, a path weighted summation operation is performed, assigning the edge weight coefficients of all directed edges in the currently valid connected path. The values ​​are accumulated to obtain the path-weighted cumulative value. Simultaneously, to conform to the physical law that the impact of a fault decreases with propagation distance, this step introduces a path length attenuation mechanism. This mechanism establishes an inverse mapping relationship between the number of path propagation levels and the directness of causal correlation; that is, the more path propagation levels an effective connected path contains, the weaker its explanatory power for the target fault. The path-weighted cumulative value is then divided by the number of path propagation levels. The power of the power is used to define the quotient as the single-path logical contribution value of the effective connected path, wherein... This is a preset path decay factor, a positive real constant stored in the system configuration file, with a value range of (0,2], typically 1. The first priority anomaly key weight coefficient and the second priority static association weight coefficient are both preset positive real constants, with the former being greater than the latter (e.g., 2.0 and 1.0 respectively). Further, a normalized accumulation operation is performed on the single-path logical contribution values ​​corresponding to all valid connected paths in the valid connected path set to generate an entity vector representing the candidate solution. The topological logical correlation degree is a numerical value representing the degree of closeness between the logical topology and the current fault scenario.

[0123] Finally, this step performs a hybrid scoring calculation, the specific execution logic of which is as follows: A preset modality balance hyperparameter is introduced, and the semantic cosine similarity value from step S3200, used to characterize semantic matching degree, is multiplied and weighted to obtain the contribution score of the semantic dimension; simultaneously, the difference between 1 and the modality balance hyperparameter is calculated, and this difference is used as the weight coefficient of the topology dimension, and the topology logical relevance value is multiplied and weighted to obtain the contribution score of the topology dimension; finally, the contribution scores of the semantic dimension and the topology dimension are linearly superimposed to generate the final comprehensive confidence score. This achieves the fusion of unstructured semantic similarity and structured logical connections. Ultimately, based on this comprehensive confidence score... For the preliminary candidate set All candidate solution entity vectors Reorder the data, generate and output the final knowledge recommendation list. The modal balance hyperparameter is a dimensionless constant pre-defined and stored in the system configuration file. Its value range is strictly constrained within the open interval (0,1). The modal balance hyperparameter is a dimensionless floating-point constant stored in the system configuration file. It is set to 0.5 by default to achieve equal weight fusion. Alternatively, it can be determined by grid search on the historical operation and maintenance work order dataset with the recommendation accuracy as the target. It is used to adjust the dependency bias weights of the two retrieval modes: unstructured text semantic matching and structured graph logical reasoning.

[0124] Step S4000: Based on the final knowledge recommendation list Retrieve pre-built automated tool library index Generate smart push cards by encapsulation Collect data for smart push cards User behavior data Real-time state stream of the environment To calculate the overall reward value, the policy gradient algorithm is used on the current parameter set. Perform backpropagation update.

[0125] Specifically, this step aims to transform the final knowledge recommendation list from step S3300, which serves as the product of background computational inference, into a list of recommended knowledge. As an object to be encapsulated, indexing is performed using front-end rendering techniques and automated tool libraries. Its association mapping capability transforms abstract solution entities into intelligent push cards that are visible to the user, contain structured information, and have executable action controls. This enables a leap from simple information retrieval to physical operation and maintenance; simultaneously, it integrates user behavior data captured by the front end. Real-time environmental state stream re-acquired after operation As a feedback incentive signal, a delayed reward function is constructed using the policy gradient algorithm. The model weights are then iteratively corrected through backpropagation to generate an updated set of model update parameters. This allows for the establishment of an adaptive evolutionary closed loop of human-machine collaboration in real-world operation and maintenance scenarios.

[0126] Further, step S4000 includes:

[0127] Step S4100: Traverse the final knowledge recommendation list Extract the visualization metadata set, and index it in the pre-built automated tool library based on the visualization metadata set. The automated execution scripts retrieved from the target device identifier will be based on the target device identifier. Extracted static configuration parameters and environmental state-based feature vectors The extracted dynamic performance metrics are populated into the automated execution script to generate a one-click repair control, which is then combined with a visual metadata set and a comprehensive confidence score. Perform structured encapsulation and rendering to generate smart push cards. .

[0128] Specifically, this step aims to transform the final knowledge recommendation list, which is discretely distributed and in a static text state from step S3300, into a more coherent and meaningful list. As an abstract logical object to be processed, index it using an automation tool library. The provided function mapping interface logically binds and encapsulates static text descriptions with dynamic automated execution scripts, thereby transforming abstract operational knowledge into intelligent push cards that are visible, interactive, and executable on the user side. This enables a physical transformation from passive reading-based knowledge acquisition to proactive operational fault handling, providing real-time interactive data input that integrates diagnosis, decision-making, and execution for subsequent steps.

[0129] In the specific implementation process, this step first iterates through the final knowledge recommendation list. The sorting position is within the preset quantity threshold Previous candidate solutions entity vector The preset quantity threshold This is a positive integer constant used to control the number of cards displayed on the front end. For each candidate solution entity, the metadata parser is invoked to read its storage structure in the knowledge base foundation and extract a set of visual metadata for front-end display. This set of visual metadata includes, but is not limited to: a summary of the root cause of the failure, a description of standard operating procedures, and the candidate solution entity's position in the knowledge graph topology. The attribute information of knowledge graph nodes, such as entity type labels and edge type attribute fields.

[0130] Subsequently, this step executes the automated script association activation logic, the specific execution logic of which is as follows: using the currently processed candidate solution entity as the unique retrieval key, the pre-set automated tool library index is searched. A matching query operation is initiated to detect whether there is an automated execution script that has a predefined functional mapping relationship with the solution entity. The automated tool library index is mentioned here. This is an index lookup table pre-stored in an external operations automation platform or local database, containing automated execution scripts such as Python scripts for restarting services, shell commands for cleaning logs, and API call instructions for scaling, mapping their relationships to knowledge graph nodes. If a matching automated execution script is found, the script call interface definition and parameter configuration template of that script are read. To achieve context awareness and automatic alignment of fault repair operations, this step is based on the target device identifier from step S2200. Retrieve and extract static configuration parameters of the target device from the associated configuration management database attributes, such as the target server IP address and service listening port number; simultaneously, based on the environmental state feature vector from step S2300... The system parses and extracts dynamic operating indicators under the current fault scenario, such as the PID of the abnormal process and the overflowed memory address segment. It automatically fills the static configuration parameters and dynamic operating indicators into the corresponding fields of the parameter configuration template, and encapsulates the filled parameter configuration template with the script call interface definition to generate a one-click repair control pre-bound with specific execution logic.

[0131] Finally, the extracted visual metadata set and the comprehensive confidence score from step S3300 will be used to... Visualized rendering data, such as percentage values ​​or confidence progress bars, and generated one-click repair controls, are used as input parameters for dynamically encapsulated functions and then structured and assembled to generate smart push cards. Finally, the front-end rendering engine calls the preset card template to display the smart push card. The message is sent to the operation terminal interface of the maintenance personnel through the message push channel. The smart push card... It is a composite encapsulated object that contains visual presentation layer data for defining the page layout structure and business logic layer for defining event responses and API command issuance.

[0132] Step S4200, in response to smart push card Collected user operation behavior data Obtain real-time environmental state stream The comprehensive reward value is calculated by combining the physical verification closed-loop mechanism, and the policy gradient algorithm is used to apply the comprehensive reward value to the current parameter set. Perform backpropagation to update and generate the model update parameter set. .

[0133] Specifically, this step aims to target the smart push card from step S4100 to the front-end user. The generated user action data Treated as a subjective explicit feedback signal, the real-time environmental state stream is collected after the operation is executed with a delay. As an objective implicit verification data object, a physical verification closed-loop mechanism is used to orthogonally align the user's subjective interaction intent with the objective state of the physical environment in the time dimension. A comprehensive reward function containing both action execution and effect verification dimensions is constructed, and the quantified reward value is mapped to the updated model update parameter set through a policy gradient algorithm. This enables closed-loop coupling between human-computer interaction intent and physical environment feedback at the underlying computing architecture, driving the system to adaptively evolve towards maximizing fault resolution rate.

[0134] In the specific implementation process, this step first deploys an event listening interface in the graphical user interface of the operation and maintenance terminal to monitor and capture the actions of operation and maintenance personnel on the smart push cards in real time. The executed interaction command log generates user operation behavior data. To address the deficiency of relying solely on subjective click data for objective verification of the effectiveness of physical repairs, this step introduces a physical verification closed-loop mechanism. When the event listening interface detects that a user has executed an interactive command representing confirmation of adoption and implementation of the solution, such as clicking a one-click repair control or executing an automated script assignment operation, the system kernel automatically triggers a hardware timer, initiating a state verification time window with a preset duration. At the end of this state verification time window, the environment awareness module in step S2300 is invoked to perform a secondary scan of the current IT infrastructure, re-collect physical environment operating indicators, and generate a time-series-adapted real-time environment state stream. Next, the differential verification logic is executed, and its specific execution logic is as follows: The real-time state stream of the post-sequence environment is... With the environmental state feature vector from step S2300 when the fault occurred By comparing the features, the environmental state feature vector was located. The system identifies dimensional components characterizing fault features in numerical anomalies, such as CPU utilization and specific error code alarms, and detects these dimensional components within the real-time environmental state stream. The comparison will be conducted to determine whether the corresponding values ​​have returned to the preset normal operating range. For example, it will check whether a specific alarm signal flag has been reset or whether the CPU load value has dropped below the baseline value. The comparison results will serve as an objective basis for evaluating the actual physical repair effectiveness of the recommended solution.

[0135] Subsequently, this step instantiates the comprehensive reward function, which serves as a mathematical calculation model mapping multidimensional input signals to a single feedback indicator. This function aims to linearly weight and fuse the interaction signal representing the immediate subjective intent of the user with the physical state signal representing the lagging objective state of the system. The logic of this comprehensive reward function is as follows: For effective recommendation strategies that meet the dual conditions of "user subjective adoption and elimination of physical fault characteristics through the state verification time window," a positive incentive scalar value is assigned; while for invalid recommendation strategies that are "user subjectively adopted but the physical fault indicator is not reset" or "the user directly executes a rejection operation," a negative incentive scalar value is assigned. This is based on the collected multidimensional feedback signals, specifically including user operation behavior data from front-end event monitoring. And the real-time state stream of the environment that is aware of the backend environment and is time-delayed. The following calculation logic generates a comprehensive reward value to characterize the actual contribution of the recommended action in the current inference cycle to the final operation and maintenance goal: First, the event parser is invoked to perform feature recognition and Boolean state determination on the multi-dimensional feedback signal: If the user operation behavior data If the event includes confirmation commands such as "one-click repair", it is considered a positive interactive behavior event. True; if the environment's real-time state stream is true. With the initial state, i.e., the feature vector of the environment state If the comparison shows that the abnormal indicators have been reset, that is, returned to the preset normal operating value range, then the system state is judged to have returned to steady state. If the user's action data is true; If the content includes a "close" or "negative review" instruction, it is considered a negative interaction event. The conditional probability is true. This is determined by scoring the overall confidence level. The resulting rating vector is subjected to Softmax normalization to obtain the probability distribution of the recommended actions, and then the log-likelihood is calculated based on this distribution; the current parameter set The learning rate includes learnable node vectors and relation vectors from the graph feature matrix, weights of the dynamically gated attention arbitration unit, and weights of the multilayer perceptron network; the learning rate is a preset positive real constant, ranging from 0.0001 to 0.01. Subsequently, logical indicator operators are invoked to calculate positive interaction events. System state returns to steady state event and negative interactive behavior events The truth mapping result. Then, the positive interaction events... The mapping result is multiplied by a preset first reward weight coefficient to obtain the subjective feedback gain component; the system state is then returned to the steady-state event. The mapping result is multiplied by a preset second reward weight coefficient to obtain the objective therapeutic effect gain component; negative interactive behavior events are then... The mapping result is multiplied by a preset third penalty weight coefficient to obtain the interaction loss component. Finally, an algebraic summation operation is performed on the above three components to output the comprehensive reward value.

[0136] Finally, this step performs a policy gradient-based backpropagation update operation for the model parameters. This operation aims to utilize the comprehensive reward value to update the current model parameter set at the system's underlying level. The directional correction is performed, and its specific execution logic is as follows: First, the comprehensive reward value is defined as the gain coefficient of the gradient update. Then, the log-likelihood gradient operator is invoked, based on the current unstructured subjective semantic feature vector from step S2100. With environmental state feature vector Calculate and generate the current smart push card The logarithm of the conditional probability relative to the current model parameter set The gradient vector. Mathematically, this gradient vector indicates the gradient that maximizes the current smart push card. The selected probability of occurrence parameter is adjusted in the following direction. Next, the preset learning rate is read and backpropagation is performed, which involves multiplying the learning rate, the overall reward value, and the gradient vector to obtain the parameter correction amount for this iteration. This parameter correction amount is then added to the current model parameter set. This generates the updated model update parameter set. The backpropagation update operation of the model parameters follows the direction of gradient ascent, which automatically increases the weight values ​​of feature dimensions that can bring high fault resolution rates in the dynamic gating attention arbitration unit in step S3100, and automatically decreases the weight values ​​of feature dimensions that "lead to invalid recommendations", thereby realizing online incremental learning of the operation and maintenance knowledge base reasoning logic.

[0137] Example 2:

[0138] This embodiment, based on Embodiment 1, provides an AI-based intelligent push system for IT service knowledge bases, such as... Figure 4 As shown, the system includes a knowledge graph construction module, a feature extraction and encoding module, an intelligent knowledge retrieval module, and an intelligent push feedback module;

[0139] The knowledge graph construction module is based on the collected raw data set. Parse and generate a set of operation and maintenance entities and initial association triples Based on the set of operation and maintenance entities and initial association triples Constructing a knowledge graph topology and the knowledge graph topology Mapped to a graph feature matrix .

[0140] The feature extraction and encoding module is used to process the real-time captured real-time dialogue text stream. Discretization mapping to subjective semantic feature vectors And based on the knowledge graph topology Target device identifier pushed to the alarm interface of the external monitoring system Perform a traversal to generate environment state feature vectors .

[0141] The intelligent knowledge retrieval module is used to fuse subjective semantic feature vectors. and environmental state feature vector To generate mixed query vectors Based on hybrid query vectors Traversing the spectral feature matrix To extract the initial candidate set And combined with the knowledge graph topology For the preliminary candidate set After filtering and sorting, the final knowledge recommendation list is obtained. .

[0142] The intelligent push feedback module is used to recommend knowledge based on the final knowledge recommendation list. Retrieve pre-built automated tool library index Generate smart push cards by encapsulation Collect data for smart push cards User behavior data Real-time state stream of the environment To calculate the overall reward value, the policy gradient algorithm is used on the current parameter set. Perform backpropagation update.

[0143] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0144] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-based intelligent push method for IT service knowledge bases, characterized in that: include: Based on the collected raw data set, an operation and maintenance entity set and an initial association triplet are generated. A knowledge graph topology structure is constructed based on the operation and maintenance entity set and the initial association triplet, and the knowledge graph topology structure is mapped to a graph feature matrix. The real-time captured dialogue text stream is discretized and mapped into a subjective semantic feature vector, and the target device identifier pushed by the alarm interface of the external monitoring system is traversed based on the knowledge graph topology to generate an environmental state feature vector. The subjective semantic feature vector and the environmental state feature vector are integrated to generate a hybrid query vector. Based on the hybrid query vector, the graph feature matrix is ​​traversed to extract a preliminary candidate set. The preliminary candidate set is then filtered and sorted in combination with the knowledge graph topology to obtain the final knowledge recommendation list. Based on the final knowledge recommendation list, a pre-set automated tool library index is retrieved to encapsulate and generate a smart push card. User operation behavior data and real-time environmental state stream for the smart push card are collected to calculate the comprehensive reward value. The policy gradient algorithm is used to perform backpropagation update on the current parameter set.

2. The AI-based intelligent push method for IT service knowledge bases according to claim 1, characterized in that, The method for generating the spectral feature matrix includes: A standard corpus sequence is generated based on the collected raw data set, and the standard corpus sequence is mapped into a state feature matrix through a sequence labeling calculation unit. Global path constraint optimization is performed on the state feature matrix to lock the global optimal label path. The standard corpus sequence is then cut according to the global optimal label path to generate a set of operation and maintenance entities and an initial association triplet. Based on the set of operation and maintenance entities and the initial association triplet, a directed graph data structure is instantiated. Cross-modal logical association verification is performed on the directed graph data structure to establish the violating association edges. The dependent association edges are derived by passing closure operation to complete the derivation. The knowledge graph topology structure is then reconstructed. Based on the knowledge graph topology, an initial state graph feature matrix is ​​instantiated. Then, parameter iterative optimization logic is executed to backpropagate and update the initial state graph feature matrix to obtain the graph feature matrix.

3. The AI-based intelligent push method for IT service knowledge bases according to claim 2, characterized in that, The method for locking the globally optimal tag path includes: A state selection topology graph is constructed based on the state feature matrix, which is defined by the temporal dimension and the state dimension. The temporal dimension corresponds to the discrete time step index of the standard corpus sequence, and the state dimension corresponds to the preset set of operation and maintenance entity labels. For each discrete time step, select each candidate state node in the topology graph, calculate the linear sum of the local observation potential of the candidate state node, the pre-optimal recursive potential pointing to the candidate state node, and the state transition potential of the candidate state node, and retain the local optimal sub-path pointer pointing to the current candidate state node that maximizes the linear sum. The local observation potential is the non-normalized emission probability intensity determined based on the state feature matrix at the corresponding temporal dimension index and state dimension index; the state transition potential is the logical dependency weight determined based on the pre-stored global transition probability matrix. In response to the recursive traversal to the last discrete time step, the candidate state node with the maximum accumulated potential energy value is selected. The reverse pointer backtracking is performed along the local optimal sub-path pointer in the state selection topology graph to lock the globally optimal label path.

4. The AI-based intelligent push method for IT service knowledge bases according to claim 2, characterized in that, The methods for identifying violations of associated edges include: Traverse the fault feature entity subset and the design baseline index entity subset in the operation and maintenance entity set to construct the entity pair to be verified, and perform semantic attribute consistency verification and numerical constraint violation verification on the entity pair to be verified. The semantic attribute consistency verification includes calculating the text similarity value between the fault object name field parsed from the fault feature entity and the benchmark indicator name field parsed from the design benchmark indicator entity. If the text similarity value is greater than the preset semantic similarity threshold, the semantic attribute consistency verification is passed. The numerical constraint violation verification includes extracting the runtime observation value from the fault feature entity and the design limit value from the design benchmark indicator entity, and performing a binary logical operation on the runtime observation value and the design limit value. If the entity pair to be verified passes the semantic attribute consistency check and numerical constraint violation check, and the result of the binary logic operation is true, then a violation association edge pointing from the fault feature entity to the design benchmark index entity is instantiated in the knowledge graph topology.

5. The AI-based intelligent push method for IT service knowledge bases according to claim 1, characterized in that, The method for generating the environmental state feature vector includes: The real-time captured dialogue text stream is discretized into a valid text sequence within a sliding sampling time window. Based on the mapping correspondence in the pre-set IT operation and maintenance semantic parsing unit, the valid text sequence is mapped into a feature state matrix sequence. The feature state matrix sequence is then compressed with full sequence information to generate a subjective semantic feature vector. In response to the target device identifier pushed by the alarm interface of the external monitoring system, a breadth-first search is performed in the knowledge graph topology to lock the set of related dependent components. The running index data of the set of related dependent components is collected within the sliding sampling time window. A time-series running status matrix is ​​generated based on the sliding sampling time window, the set of related dependent components, and the running index data. Z-score normalization and first-order difference operation are performed on the time-series running state matrix to generate dimensionless running values ​​and time-series fluctuation rates, respectively. A joint feature input vector is constructed based on the dimensionless running values ​​and time-series fluctuation rates. The joint feature input vector is then input into a multilayer perceptron network to generate an environmental state feature vector.

6. The AI-based intelligent push method for IT service knowledge bases according to claim 1, characterized in that, The methods for obtaining the final knowledge recommendation list include: The probability distribution of the subjective semantic feature vector is normalized to calculate the semantic uncertainty index. The semantic uncertainty index is mapped to the semantic feature weight coefficient using a dynamic gated attention arbitration unit. Based on the semantic feature weight coefficient, the subjective semantic feature vector and the environmental state feature vector are linearly weighted and fused to generate a hybrid query vector. Extract candidate solution entity vectors from the graph feature matrix, perform complex similarity measurement operation to calculate the semantic cosine similarity value between the mixed query vector and the candidate solution entity vector, and perform result sorting and filtering logic based on the semantic cosine similarity value to construct a preliminary candidate set; Using the target device identifier as the physical starting anchor, the system retrieves valid connected paths in the knowledge graph topology that connect the initial candidate set. Based on the valid connected paths, it performs topological association quantification to generate a topological logical association value, and combines it with semantic cosine similarity values ​​to perform linear weighted fusion to generate a comprehensive confidence score. Finally, it reorders the initial candidate set based on the comprehensive confidence score to output the final knowledge recommendation list.

7. The AI-based intelligent push method for IT service knowledge bases according to claim 6, characterized in that, The method for performing the topological correlation metric calculation includes: Traverse the directed edges contained in the valid connected path. If the edge type attribute field of the directed edge is marked as a violation of association edge, then the preset edge weight coefficient is assigned to the first priority abnormal critical weight coefficient; if the edge type attribute field of the directed edge is marked as a normal association edge, then the edge weight coefficient is assigned to the second priority static association weight coefficient. The edge weight coefficients of all directed edges in the effective connected path are accumulated to generate the path weighted cumulative value. The power of the path transmission level in the knowledge graph topology based on the effective connected path is used as the denominator, and the path weighted cumulative value is used as the numerator to perform a division operation to obtain the single path logical contribution value. Perform a normalized summation operation on the single-path logical contribution values ​​of all valid connected paths in the preliminary candidate set to generate a topological logical correlation value.

8. The AI-based intelligent push method for IT service knowledge bases according to claim 1, characterized in that, The step of performing backpropagation update on the current parameter set includes: The system iterates through the final knowledge recommendation list to extract a set of visual metadata. Based on the set of visual metadata, it retrieves the associated automated execution scripts from the pre-set automated tool library index. It then fills the automated execution scripts with static configuration parameters extracted based on the target device identifier and dynamic operation indicators extracted based on the environmental state feature vector to generate a one-click repair control. Finally, it combines the set of visual metadata and the comprehensive confidence score to perform structured encapsulation and rendering to generate a smart push card. In response to user operation behavior data collected from smart push cards to obtain real-time environmental state stream, a comprehensive reward value is calculated by combining a physical verification closed-loop mechanism. Based on the comprehensive reward value, the current parameter set is updated by backpropagation using a policy gradient algorithm to generate a model update parameter set.

9. The AI-based intelligent push method for IT service knowledge bases according to claim 8, characterized in that, The method for calculating the comprehensive reward value includes: Boolean state determination is performed on user operation behavior data and real-time environmental state stream. If the user operation behavior data contains a confirmation execution command, the positive interaction behavior event is determined to be true. If the comparison between the real-time environmental state stream and the environmental state feature vector shows that the abnormal indicators have returned to the preset normal operating value range, the system state has returned to steady state event, which is determined to be true. If the user operation behavior data contains a "close" or "negative evaluation" command, the negative interaction behavior event is determined to be true. The truth mapping result of positive interactive behavior events is multiplied by a preset first reward weight coefficient to generate a subjective feedback gain component; the truth mapping result of system state return to steady state events is multiplied by a preset second reward weight coefficient to generate an objective therapeutic effect gain component; the truth mapping result of negative interactive behavior events is multiplied by a preset third penalty weight coefficient to generate an interactive loss component. Perform algebraic summation on the subjective feedback gain component, the objective therapeutic effect gain component, and the interaction loss component, and output the comprehensive reward value.

10. An AI-based intelligent push system for IT service knowledge bases, used to implement the AI-based intelligent push method for IT service knowledge bases as described in any one of claims 1-9, characterized in that, The system includes a knowledge graph construction module, a feature extraction and encoding module, an intelligent knowledge retrieval module, and an intelligent push feedback module; The knowledge graph construction module is used to parse and generate a set of operation and maintenance entities and an initial association triplet based on the collected raw data set, construct a knowledge graph topology structure based on the set of operation and maintenance entities and the initial association triplet, and map the knowledge graph topology structure into a graph feature matrix. The feature extraction and encoding module is used to discretize and map the real-time captured real-time dialogue text stream into a subjective semantic feature vector, and to traverse the target device identifier pushed by the alarm interface of the external monitoring system based on the knowledge graph topology to generate an environmental state feature vector. The intelligent knowledge retrieval module is used to fuse subjective semantic feature vectors and environmental state feature vectors to generate a hybrid query vector, traverse the graph feature matrix based on the hybrid query vector to extract a preliminary candidate set, and filter and sort the preliminary candidate set in combination with the knowledge graph topology to obtain the final knowledge recommendation list. The intelligent push feedback module is used to retrieve the pre-set automated tool library index based on the final knowledge recommendation list to encapsulate and generate intelligent push cards, collect user operation behavior data and real-time environmental state stream for intelligent push cards to calculate the comprehensive reward value, and use the policy gradient algorithm to perform backpropagation update on the current parameter set.

Images