Search

CN-121996464-A - Reverse reasoning method and system for fault root cause based on graph neural network

CN121996464ACN 121996464 ACN121996464 ACN 121996464ACN-121996464-A

Abstract

The invention provides a fault root reverse reasoning method and system based on a graph neural network, which are used for receiving a fault image set after fault occurrence, performing multidimensional deconstruction processing to generate an image feature set, constructing a dynamic fault conduction track based on the image feature set to form a structured fault conduction track containing node dependency relationships, generating a hierarchical root reverse tracing constraint condition according to the structured fault conduction track, calling a customized graph neural network, loading the structured fault conduction track and the constraint condition, performing feature aggregation and weight distribution through a double-track convolution layer, performing double-track tracing reasoning reversely along the conduction track, and generating a final fault root reasoning result by combining tracing path conflict regulation operation. The invention can effectively process complex fault appearance information, accurately construct fault conduction tracks and realize efficient reverse reasoning of fault root cause.

Inventors

  • LI FENG
  • WANG YUN

Assignees

  • 博德韦尔(成都)科技有限公司

Dates

Publication Date
20260508
Application Date
20260227

Claims (10)

  1. 1. The fault root cause reverse reasoning method based on the graph neural network is characterized by comprising the following steps of: Receiving a fault appearance set after a fault occurs, wherein the fault appearance set comprises various types of description information externally presented by the fault, and the description information covers state change, phenomenon parameters and associated influences when the fault occurs; performing multidimensional deconstructing on the fault image set, splitting description information in the fault image set into atomic level expression fragments, performing feature coding and cross fusion on the rest fragments after irrelevant fragments are removed through association filtering, and generating an image feature set; Constructing a dynamic fault conduction track based on the appearance feature set, constructing a conduction path, arranging and transmitting sequence and association type through modeling of interaction intensity among features, and forming a structured fault conduction track containing node dependency relationship; Generating a hierarchical root cause reverse tracing constraint condition according to node density, path branch coefficient and associated strength distribution of the structured fault conduction track, wherein the constraint condition comprises a path allowable range, an associated type limit and an strength threshold of a tracing level; and calling a customized graph neural network, loading the structured fault conduction track and the hierarchical root cause reverse tracing constraint condition, performing feature aggregation and weight distribution through a double-track convolution layer, reversely developing double-track tracing reasoning along the conduction track, and generating a final fault root cause reasoning result by combining tracing path conflict regulation operation.
  2. 2. The reverse reasoning method of fault root cause based on graphic neural network according to claim 1, wherein the performing multidimensional deconstructing on the fault image set, splitting description information in the fault image set into atomic level expression segments, performing feature coding and cross fusion on the remaining segments after removing irrelevant segments by association filtering, and generating an image feature set comprises: Carrying out multidimensional resolution on description information in the fault representation set according to state dimension, phenomenon dimension and influence dimension, wherein each dimension corresponds to a plurality of independent atomic level expression fragments, and the atomic level expression fragments only describe specific fault presentation in a single dimension; a preset fault associated feature dictionary is called, the fault associated feature dictionary comprises a feature expression set and expression associated weights related to fault occurrence, the expressions in the atomic level expression fragments and the fault associated feature dictionary are compared sentence by sentence, and the matching degree of the fragment expressions and the dictionary expressions is calculated; setting a matching degree screening threshold, reserving atomic-level expression fragments with matching degree reaching the threshold, and removing irrelevant atomic-level expression fragments with matching degree not reaching the threshold to form an effective expression set corresponding to the description information in the fault expression set; vector coding is carried out on the atomic-level expression fragments in the effective expression set, and a pre-trained feature encoder is adopted to convert the character-form fragment expression into a fixed-dimension expression feature vector, so that the dimension of the expression feature vector under different dimensions is kept the same; inputting all expression feature vectors corresponding to the description information in the fault representation set into a cross fusion module, realizing cross correlation of different dimension features through element-level multiplication operation, extracting key information components in the fusion vector, carrying out feature calibration on the key information components corresponding to all the description information in the fault representation set, and combining to form the representation feature set after eliminating dimension deviation.
  3. 3. The reverse reasoning method of fault root cause based on graph neural network according to claim 1, wherein the constructing a dynamic fault conduction track based on the appearance feature set, constructing a conduction path, arranging and transmitting sequence and association type through inter-feature interaction strength modeling, forming a structured fault conduction track including node dependency relation comprises: performing association relation detection on the appearance features in the appearance feature set, identifying interaction association among different appearance features by adopting a feature co-occurrence analysis algorithm, and recording all feature pairs and action parameters with interaction; Extracting interaction information of feature pairs, wherein the interaction information comprises action triggering conditions, action transmission media, action influence parameters and action duration parameters, screening source features initiating actions based on the action triggering conditions, screening target features receiving actions based on the action influence parameters, and establishing a feature action mapping relation; based on all feature action mapping relations, constructing a feature transmission chain by adopting a path mining algorithm, setting the transmission priority of the feature in the feature set by calculating the action initiation times and the receiving times of the feature in the chain, and arranging the feature transmission sequence according to the transmission priority; Taking the apparent features as track nodes, taking interaction association among the features as track connecting edges, arranging the nodes according to the sequence of feature transmission, and forming an initial fault conduction track by the connecting edges corresponding to the interaction information; Dynamically optimizing an initial fault conduction track, adjusting node connection strength based on continuous parameters of characteristic action, supplementing the association type identifier of a connection edge according to action transfer media, and adding a node dependency mark in combination with characteristic conduction priority to form a structured fault conduction track comprising nodes, connection edges, association type identifiers, node dependency marks and connection strength.
  4. 4. The method for reverse reasoning of fault root cause based on graph neural network according to claim 1, wherein the generating a hierarchical root cause reverse tracing constraint condition according to node density, path branching coefficient and associated intensity distribution of the structured fault conduction track, wherein the constraint condition includes path allowable range, associated type limitation and intensity threshold of tracing hierarchy, includes: analyzing the structured fault conduction track, counting the total number of nodes, the total number of connecting edges and the number of path branches in the track, calculating the node density in unit path length and the path branch coefficient of the nodes in the track, wherein the node density is obtained through the association operation of the total number of the nodes and the total length of the track, and the path branch coefficient is obtained through the association operation of the number of branch paths associated with the nodes and the total number of the connecting edges of the nodes; Extracting the association strength and association type identification of the connecting edge, drawing an association strength distribution curve, determining a peak value interval, a mean value interval and a valley value interval of the association strength, and counting the distribution proportion of different types of association based on the association type identification; Dividing the structured fault conduction track into a plurality of continuous tracing levels according to the distribution states of node density and path branch coefficients, wherein the divided tracing levels comprise a fixed number of nodes and corresponding connecting edges, and the level dividing boundaries are set at positions where the node density and the path branch coefficients are at the valley values; Distributing corresponding constraint levels for the divided tracing levels, wherein the constraint levels are lifted along with the lifting of average association strength of connecting edges in the corresponding levels, the constraint levels are lifted along with the tightening of constraint standards, association strength thresholds of the corresponding levels are set based on the constraint levels, and connecting edges with association strength not reaching the thresholds are not included in tracing paths; And defining the allowed tracing path direction, the limited association type and the corresponding intensity threshold value of each hierarchy by combining the path range, the association type distribution proportion and the constraint level of the divided tracing hierarchy, and integrating the tracing path direction, the limited association type and the corresponding intensity threshold value to form a hierarchical root cause reverse tracing constraint condition, wherein the tracing hierarchy corresponds to an independent constraint sub-condition.
  5. 5. The method for reverse reasoning of fault root cause based on graph neural network according to claim 1, wherein the invoking the customized graph neural network, loading the structured fault conduction track and the hierarchical root cause reverse tracing constraint condition, performing feature aggregation and weight distribution through a double-track convolution layer, performing double-track tracing reasoning reversely along the conduction track, and generating a final fault root cause reasoning result by combining tracing path conflict regulation operation, comprises: Converting the structured fault conduction track into heterogeneous graph structure data compatible with the graph neural network, wherein nodes in the heterogeneous graph structure data correspond to track nodes, edges in the heterogeneous graph structure data correspond to track connecting edges, node attributes comprise characteristic attribute information and node dependency marks, and edge attributes comprise association type identifiers and connection strength; Embedding the hierarchical root into the iso-composition structural data according to the tracing hierarchy by using the reverse tracing constraint condition, and adding constraint marks for the graph structural parts corresponding to the divided tracing hierarchy, wherein the constraint marks comprise constraint levels, association strength thresholds and allowed association types; Inputting the heterogeneous graph structure data with the constraint marks into a double-track convolution layer of a customized graph neural network, wherein the first track convolution layer executes local neighborhood feature aggregation aiming at node attributes, and the second track convolution layer executes association strength feature aggregation aiming at edge attributes to generate node and edge fusion features; The node and edge fusion characteristics are subjected to weight distribution through a dynamic attention mechanism layer of the graph neural network, attention weights are dynamically adjusted based on constraint levels in constraint marks, characteristic weights conforming to constraint conditions are strengthened, characteristic weights not conforming to constraint conditions are weakened, and weighted fusion characteristics are generated; Carrying out double-track tracing traversal by taking a track end node as a starting point and combining a weighted fusion characteristic and a constraint mark along the reverse direction of the structured fault conduction track, tracing along the node dependency relationship to form a first tracing path, tracing along the association strength standard reaching path to form a second tracing path, and forming an initial tracing path together by the two tracing paths; Performing conflict detection on the initial tracing path, identifying conflict points of path nodes, association types and intensity thresholds, analyzing the logic fit degree of the conflict points by adopting a logic consistency analysis algorithm, reserving path fragments with logic fit, and integrating to form a unified tracing path; And extracting the apparent features corresponding to the initial nodes of the unified tracing paths as root candidate, repeating the tracing traversal process to obtain a plurality of root candidates, carrying out relevance ranking on all the obtained root candidates, integrating the feature information, tracing paths and association strength of the root candidates, and generating a final fault root reasoning result.
  6. 6. The method for reverse reasoning of fault cause based on graph neural network according to claim 5, wherein the converting the structured fault conduction trace into heterogeneous graph structure data compatible with the graph neural network, wherein nodes in the heterogeneous graph structure data correspond to trace nodes, edges in the heterogeneous graph structure data correspond to trace connection edges, node attributes comprise feature attribute information and node dependency marks, edge attributes comprise association type identifiers and connection strength, and the method comprises: Extracting all track nodes in the structured fault conduction track, corresponding characteristic attribute information and node dependency marks, and distributing exclusive node identifiers for the track nodes in the structured fault conduction track, wherein the node identifiers, the characteristic attribute information and the node dependency marks are bound to form a node attribute set; extracting all track connecting edges in the structured fault conduction track and corresponding association type identifiers and connection strength, and distributing exclusive edge identifiers for the track connecting edges in the structured fault conduction track, wherein the edge identifiers are bound with the association type identifiers and the connection strength to form an edge attribute set; Constructing an abnormal-pattern topological structure based on the arrangement sequence and the connection relation of the track nodes, wherein the connection relation of the nodes in the topological structure is consistent with the connection relation of the nodes in the structured fault conduction track; Embedding the information in the node attribute set into the node corresponding to the topological structure, so that the attribute information of the node in the topological structure is completely mapped, and the characteristic attribute information and the node dependency mark are omitted; and embedding the information in the edge attribute set into the edge corresponding to the topological structure, so that the attribute information of the edge in the topological structure is completely mapped, and the correlation type identifier and the connection strength are not missed, thereby forming heterogeneous graph structure data compatible with the graph neural network.
  7. 7. The reverse reasoning method of fault cause based on graph neural network according to claim 5, wherein the step of performing weight distribution on the node and edge fusion feature by the dynamic attention mechanism layer of the graph neural network, dynamically adjusting attention weight based on constraint level in constraint mark, strengthening feature weight conforming to constraint condition, weakening feature weight not conforming to constraint condition, and generating weighted fusion feature comprises: extracting feature dimension information in the node and edge fusion features and constraint grade information in constraint marks, and establishing a mapping relation between feature dimensions and constraint grades; setting an attention weight reference value based on a constraint level, wherein the constraint level is lifted along with the lifting of the attention weight reference value to form a weight reference table; Analyzing the degree of agreement between the characteristic dimension and the association type limit and the strength threshold value of the corresponding constraint level through a constraint matching algorithm; The method comprises the steps of distributing strengthening weights based on weight reference values for feature dimensions conforming to constraint conditions, obtaining the strengthening weights through association operation of the weight reference values and feature association strengths, distributing weakening weights based on the weight reference values for feature dimensions not conforming to the constraint conditions, and obtaining the weakening weights through association operation of the weight reference values and constraint deviation coefficients; and carrying out weighted calculation on the strengthening weights or weakening weights of all the feature dimensions and the corresponding feature dimension values, integrating all the weighted feature dimension values, generating weighted fusion features, and retaining key feature information conforming to constraint conditions by the weighted fusion features.
  8. 8. The reverse reasoning method of fault root cause based on graphic neural network according to claim 3, wherein the constructing a feature transmission chain based on all feature action mapping relations by adopting a path mining algorithm, setting the transmission priority of the feature in the feature set by calculating the action initiation times and the receiving times of the feature in the chain, and arranging the feature transmission sequence according to the transmission priority, comprises: classifying all feature action mapping relations according to source features to form a mapping subset taking the source features as cores, wherein the classified mapping subset contains all target features and interaction information corresponding to the source features; Carrying out path extension analysis on the classified mapping subsets, and gradually extending to form a plurality of feature transmission sub-chains by taking a source feature as a starting point and a target feature as an intermediate node or a terminal point; combining repeated path segments in all characteristic transmission sub-chains, removing redundant chains to form a complete characteristic transmission chain, and enabling the chain to contain all the appearance characteristics; counting the number of times of initiation of action and the number of times of reception of the apparent features in the feature transfer chain, wherein the number of times of initiation of action is the number of times of taking the corresponding feature as a source feature, and the number of times of reception is the number of times of taking the corresponding feature as a target feature; And calculating the conduction priority coefficient of the appearance feature in the appearance feature set, wherein the conduction priority coefficient is obtained through the association operation of the action initiation times and the receiving times, all the appearance features are arranged according to the order of the conduction priority coefficient from large to small, and the sequence of feature transmission is determined.
  9. 9. The method for reverse reasoning of fault root cause based on graph neural network according to claim 4, wherein the assigning the corresponding constraint level to the classified trace-source hierarchy, the constraint level being raised along with the increase of the average association strength of the connected edges in the corresponding hierarchy, the constraint level being raised along with the tightening of constraint criteria, the setting of the association strength threshold of the corresponding hierarchy based on the constraint level, comprises: calculating the average value of the association strengths of all the connecting edges in the divided tracing levels, and taking the average value as the average association strength of the corresponding level; sorting the average association strengths of all the tracing layers, dividing constraint level intervals according to sorting results, wherein the average association strength of the front sorting corresponds to a higher constraint level; Distributing exclusive constraint grade identifiers for the divided constraint grade intervals, wherein the constraint grade identifiers are in one-to-one correspondence with the interval ranges to form a constraint grade mapping table; setting constraint grade identification corresponding to the tracing level according to constraint grade interval of average association strength of the divided tracing level; and setting a correlation strength threshold based on the constraint grade identification, wherein the constraint grade is lifted along with the lifting of the correlation strength threshold, the correlation strength threshold is a preset proportion of the average correlation strength of the corresponding level, and the connecting edge with the correlation strength reaching below the threshold does not contain the tracing path.
  10. 10. The system for reverse reasoning of the fault root cause based on the graph neural network is characterized by comprising a processor and a memory, wherein the memory is connected with the processor, the memory is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the memory so as to realize the reverse reasoning method of the fault root cause based on the graph neural network according to any one of claims 1-9.

Description

Reverse reasoning method and system for fault root cause based on graph neural network Technical Field The invention relates to the field of artificial intelligence, in particular to a fault root cause reverse reasoning method and system based on a graph neural network. Background In the field of fault diagnosis of complex systems, conventional fault root cause reasoning methods often face a plurality of challenges. On the one hand, the appearance information presented when the fault occurs is rich and various, and covers the aspects of state change, phenomenon parameters, association influence and the like when the fault occurs, and the information is mutually interwoven and intricate, so that the traditional method is difficult to comprehensively and effectively process and analyze the information. For example, in a large industrial production system, a device failure may cause a series of cascading reactions, and generate a large amount of fault representation information with different types and dimensions, and when the conventional method faces such massive and complex information, key features cannot be accurately extracted, so that the accuracy and efficiency of fault root cause are greatly reduced. On the other hand, the fault conduction process has dynamics and uncertainty, and the traditional method has difficulty in constructing an accurate fault conduction model to describe the propagation path and action mechanism of the fault from the appearance to the root cause. Most of the existing fault reasoning methods are based on fixed rules or simple association analysis, cannot adapt to the diversity and variability of fault conduction paths in a complex system, and cannot accurately trace the root cause of faults. Disclosure of Invention In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for reverse reasoning of a root cause of a fault based on a graph neural network, where the method includes: Receiving a fault appearance set after a fault occurs, wherein the fault appearance set comprises various types of description information externally presented by the fault, and the description information covers state change, phenomenon parameters and associated influences when the fault occurs; performing multidimensional deconstructing on the fault image set, splitting description information in the fault image set into atomic level expression fragments, performing feature coding and cross fusion on the rest fragments after irrelevant fragments are removed through association filtering, and generating an image feature set; Constructing a dynamic fault conduction track based on the appearance feature set, constructing a conduction path, arranging and transmitting sequence and association type through modeling of interaction intensity among features, and forming a structured fault conduction track containing node dependency relationship; Generating a hierarchical root cause reverse tracing constraint condition according to node density, path branch coefficient and associated strength distribution of the structured fault conduction track, wherein the constraint condition comprises a path allowable range, an associated type limit and an strength threshold of a tracing level; and calling a customized graph neural network, loading the structured fault conduction track and the hierarchical root cause reverse tracing constraint condition, performing feature aggregation and weight distribution through a double-track convolution layer, reversely developing double-track tracing reasoning along the conduction track, and generating a final fault root cause reasoning result by combining tracing path conflict regulation operation. In still another aspect, an embodiment of the present invention further provides a system for reverse reasoning of a fault root cause based on a graph neural network, including a processor, and a machine-readable storage medium, where the machine-readable storage medium is connected to the processor, and the machine-readable storage medium is used to store a program, an instruction, or a code, and the processor is used to execute the program, the instruction, or the code in the machine-readable storage medium, so as to implement the method described above. Based on the above aspects, the embodiment of the invention can split complex fault image information into atomic level expression fragments by performing multidimensional deconstructing processing on the fault image set, generate an accurate image feature set through operations such as association filtering, feature encoding, cross fusion and the like, effectively extract key information in the fault image, accurately simulate the propagation process of faults in a system based on a dynamic fault conduction track constructed by the image feature set, clearly present conduction paths, transmission sequence and association types throug