Fault root cause reverse reasoning method and system based on graph neural network

Abstract

The invention provides a fault root cause reverse reasoning method and system based on a graph neural network, and the method comprises the steps: receiving a fault representation set after a fault occurs, carrying out the multi-dimensional deconstruction processing, and generating a representation feature set; constructing a dynamic fault conduction track based on the representation feature set, and forming a structured fault conduction track containing a node dependency relationship; generating a hierarchical root cause reverse traceability constraint condition according to the structured fault conduction trajectory; calling a customized graph neural network, loading a structured fault conduction track and constraint conditions, executing feature aggregation and weight distribution through a double-track convolutional layer, reversely carrying out double-track traceability reasoning along the conduction track, and generating a final fault root cause reasoning result in combination with a traceability path conflict mediation operation. According to the method, complex fault representation information can be effectively processed, a fault conduction track is accurately constructed, and efficient fault root cause reverse reasoning is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a method and system for reverse reasoning of fault root causes based on graph neural networks. Background Technology

[0002] In the field of fault diagnosis for complex systems, traditional root cause reasoning methods often face numerous challenges. On the one hand, the information presented when a fault occurs is rich and diverse, encompassing multiple aspects such as state changes, phenomenal parameters, and related influences. This information is intertwined and complex, making it difficult for traditional methods to comprehensively and effectively process and analyze it. For example, in large-scale industrial production systems, a single equipment failure can trigger a series of chain reactions, generating a large amount of fault manifestation information of different types and dimensions. Traditional methods often fail to accurately extract key features when faced with such massive and complex information, significantly reducing the accuracy and efficiency of root cause reasoning. On the other hand, the fault propagation process is dynamic and uncertain. Traditional methods struggle to construct accurate fault propagation models to describe the propagation path and mechanism of the fault from manifestation to root cause. Most existing fault reasoning methods are based on fixed rules or simple correlation analysis, which cannot adapt to the diversity and variability of fault propagation paths in complex systems, making it difficult to accurately trace the root cause of the fault. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for reverse reasoning of fault root causes based on graph neural networks, the method comprising: Receive a set of fault manifestations after a fault occurs. The set of fault manifestations includes various descriptive information about the external presentation of the fault. The descriptive information covers the state changes, phenomenon parameters and related effects when the fault occurs. The fault manifestation set is subjected to multi-dimensional deconstruction processing, the descriptive information in the fault manifestation set is split into atomic-level expression fragments, irrelevant fragments are removed by association filtering, and the remaining fragments are feature-encoded and cross-fused to generate a manifestation feature set. Based on the aforementioned set of appearance features, a dynamic fault propagation trajectory is constructed. By modeling the interaction strength between features, the propagation path is constructed, the order of propagation and the association type are arranged, and a structured fault propagation trajectory containing node dependencies is formed. Based on the node density, path branch coefficient, and association strength distribution of the structured fault propagation trajectory, hierarchical root cause reverse tracing constraints are generated. The constraints include the path allowable range, association type restrictions, and strength thresholds at the tracing level. A customized graph neural network is invoked, and the structured fault propagation trajectory and hierarchical root cause reverse tracing constraints are loaded. Feature aggregation and weight allocation are performed through dual-track convolutional layers. Dual-track tracing reasoning is carried out in reverse along the propagation trajectory. Finally, the root cause reasoning result is generated by combining the source tracing path conflict mediation operation.

[0004] In another aspect, embodiments of the present invention also provide a fault root cause reverse reasoning system based on graph neural networks, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0005] Based on the above, this embodiment of the invention, through multi-dimensional deconstruction of the fault manifestation set, can break down complex fault manifestation information into atomic-level representation fragments. After operations such as association filtering, feature encoding, and cross-fusion, a precise manifestation feature set is generated, effectively extracting key information from the fault manifestation. Secondly, the dynamic fault propagation trajectory constructed based on the manifestation feature set can accurately simulate the fault propagation process in the system. Through modeling the interaction strength between features, the propagation path, transmission sequence, and association type are clearly presented, forming a structured fault propagation trajectory containing node dependencies. Furthermore, the hierarchical root cause reverse tracing constraints generated based on the structured fault propagation trajectory can finely constrain the tracing process from multiple dimensions such as node density, path branch coefficient, and association strength distribution, ensuring the rationality and accuracy of the tracing path, avoiding invalid or erroneous tracing paths, and improving the efficiency of fault root cause reasoning. Finally, a customized graph neural network is invoked, loading structured fault propagation trajectories and hierarchical root cause reverse tracing constraints. Feature aggregation and weight allocation are performed through dual-track convolutional layers, and dual-track tracing reasoning is carried out in reverse along the propagation trajectory. Combined with the source path conflict mediation operation, the final fault root cause reasoning result can be generated accurately and efficiently, providing strong support for fault diagnosis and maintenance of complex systems and greatly improving the reliability and stability of the system. Attached Figure Description

[0006] Figure 1 This is a schematic diagram of the execution flow of the fault root cause reverse reasoning method based on graph neural network provided in the embodiments of the present invention.

[0007] Figure 2 This is a schematic diagram of exemplary hardware and software components of the fault root cause reverse reasoning system based on graph neural networks provided in an embodiment of the present invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating a graph neural network-based reverse reasoning method for fault root causes, provided in one embodiment of the present invention. The following is a detailed description of this graph neural network-based reverse reasoning method for fault root causes.

[0009] Step S110: Receive a set of fault manifestations after the fault occurs. The set of fault manifestations includes various descriptive information of the external presentation of the fault. The descriptive information covers the state changes, phenomenon parameters and related effects when the fault occurs.

[0010] In this embodiment, taking motor fault diagnosis on an industrial production line as an example, when a motor malfunctions, the system collects a set of fault symptoms through various methods such as sensors, monitoring equipment, and manual recording. The descriptive information in this set of fault symptoms specifically includes: Regarding state changes, the motor suddenly changes from normal operation to abnormal vibration, the speed fluctuates from a stable value, and the vibration frequency gradually increases; regarding phenomenological parameters, the motor casing temperature rises compared to normal operation, and an imbalance occurs in the three-phase current, with one phase current increasing and the other two phase currents decreasing, accompanied by abnormal noise; regarding related impacts, due to the motor malfunction, the speed of the connected conveyor belt decreases, leading to material accumulation in subsequent processes and a reduction in the overall production efficiency of the production line. These descriptive information collectively constitute the set of fault symptoms after a motor malfunction.

[0011] Step S120: Perform multi-dimensional deconstruction processing on the fault manifestation set, split the descriptive information in the fault manifestation set into atomic-level expression fragments, remove irrelevant fragments through association filtering, and perform feature encoding and cross-fusion on the remaining fragments to generate a manifestation feature set.

[0012] In this embodiment, the fault manifestation set of the aforementioned motor fault needs to be deconstructed in multiple dimensions. First, the descriptive information in the fault manifestation set is split into multiple dimensions according to the state dimension, phenomenon dimension, and impact dimension. Each dimension corresponds to several independent atomic-level description fragments, and each atomic-level description fragment only describes the specific fault presentation under a single dimension.

[0013] Step S121: The descriptive information in the fault manifestation set is split into multiple dimensions according to the state dimension, phenomenon dimension, and impact dimension. Each dimension corresponds to several independent atomic-level description fragments. The atomic-level description fragments only describe the specific fault presentation under a single dimension.

[0014] In this embodiment, regarding the state dimension, the statement "The motor suddenly changes from normal operation to abnormal vibration, the speed starts to fluctuate from a stable value, and the vibration frequency gradually increases" is broken down into atomic-level descriptions such as "The motor's operating state changes to abnormal vibration," "The motor speed fluctuates," and "The motor vibration frequency gradually increases." Regarding the phenomenon dimension, the statement "The motor casing temperature rises compared to normal operation, and the three-phase current becomes unbalanced, with one phase current increasing and the other two phase currents decreasing, accompanied by abnormal noise" is broken down into atomic-level descriptions such as "The motor casing temperature rises," "The three-phase current is unbalanced," "One phase current increases," "One phase current decreases," "One phase current decreases," and "The motor makes abnormal noise." Regarding the impact dimension, the statement "Due to the motor's abnormality, the speed of the conveyor belt connected to it decreases, leading to material accumulation in subsequent processes and a reduction in the overall production efficiency of the production line" is broken down into atomic-level descriptions such as "The conveyor belt speed decreases," "Material accumulation in subsequent processes," and "The production line's production efficiency decreases."

[0015] Step S122: Retrieve a preset fault association feature dictionary, which contains a set of feature expressions related to fault occurrence and expression association weights. Compare the atomic-level expression fragments with the expressions in the fault association feature dictionary sentence by sentence, and calculate the matching degree between the fragment expression and the dictionary expression.

[0016] In this embodiment, the preset fault association feature dictionary is constructed based on a large number of motor fault cases and domain knowledge. It contains various feature expressions related to motor faults, such as "abnormal vibration," "temperature rise," "current imbalance," and "speed fluctuation." Each expression corresponds to an expression association weight, which indicates the degree of association between the expression and the motor fault. The atomic-level expression fragment obtained in step S121, such as "motor casing temperature rises," is compared with the expression "temperature rise" in the fault association feature dictionary. The matching degree is determined by calculating the similarity between the two in terms of semantics, keywords, etc.

[0017] Step S123: Set a matching degree filtering threshold, retain atomic-level description fragments with matching degree reaching the threshold, and remove irrelevant atomic-level description fragments with matching degree not reaching the threshold to form a set of valid descriptions corresponding to the description information in the fault manifestation set.

[0018] In this embodiment, a matching degree filtering threshold is set. For the matching degree calculated in step S122, such as "motor casing temperature rises" matching "temperature rises" in the dictionary, if the matching degree reaches the threshold, it is retained. However, if a certain atomic-level description fragment, "the motor surface has slight scratches," does not match any description in the dictionary, it is discarded. After filtering, the retained atomic-level description fragments form a valid description set, such as "motor operating state changes to abnormal vibration state," "motor speed fluctuates," "motor vibration frequency gradually increases," "motor casing temperature rises," "three-phase current imbalance," "one phase current increases," "one phase current decreases," "one phase current decreases," "motor accompanied by abnormal noise," "conveyor belt speed decreases," "material accumulation in subsequent processes," "production line production efficiency decreases," etc.

[0019] Step S124: Perform vector encoding processing on the atomic-level expression fragments in the effective expression set. Use a pre-trained feature encoder to convert the text-based fragment expressions into fixed-dimensional expression feature vectors, so that the dimension of the expression feature vectors under different dimensions remains the same.

[0020] In this embodiment, a BERT-based pre-trained feature encoder is employed. This encoder can transform atomic-level textual fragments into fixed-dimensional feature vectors. For example, for the atomic-level fragment "motor casing temperature rises," the feature encoder produces a fixed-dimensional vector, where each element corresponds to a feature value. These feature values ​​comprehensively reflect the semantic information and feature attributes of the fragment. Similarly, other atomic-level fragments are also transformed into feature vectors of the same dimension, ensuring consistency in the dimensionality of feature vectors across different dimensions.

[0021] Step S125: Input all the descriptive feature vectors corresponding to the descriptive information in the fault manifestation set into the cross-fusion module, realize the cross-association of features of different dimensions through element-level multiplication operation, extract the key information components in the fusion vector, perform feature calibration on the key information components corresponding to all descriptive information in the fault manifestation set, and combine them to form the manifestation feature set after eliminating dimensional bias.

[0022] In this embodiment, all the description feature vectors obtained in step S124 are input into the cross-fusion module. For example, the description feature vector corresponding to "motor operating state changes to abnormal vibration state" in the state dimension is multiplied element-wise with the description feature vector corresponding to "three-phase current imbalance" in the phenomenon dimension to obtain a new fusion vector. Each element in the fusion vector is the product of the corresponding elements of the two original vectors, thereby realizing the cross-correlation of features of different dimensions. Then, by setting a threshold or using an attention mechanism, key information components are extracted from the fusion vector. These key information components can highlight the correlation between features of different dimensions. Next, feature calibration is performed on the key information components corresponding to all descriptive information. For example, through standardization, the numerical ranges of different key information components are adjusted to the same interval to eliminate dimensional bias. Finally, the calibrated key information components are combined to form a set of appearance features containing multiple feature vectors. Each feature vector corresponds to a key information component that has been cross-fused and calibrated.

[0023] Step S130: Construct a dynamic fault propagation trajectory based on the aforementioned feature set. By modeling the interaction strength between features, construct the propagation path, arrange the propagation order and association types, forming a structured fault propagation trajectory containing node dependencies. In this embodiment, based on the set of motor fault manifestation features obtained in step S125, a dynamic fault transmission trajectory is constructed. First, it is necessary to identify the interaction and correlation between manifestation features, then construct the transmission path, determine the transmission order and correlation type, and finally form a structured fault transmission trajectory.

[0024] Step S131: Probe the correlation between the representational features in the representational feature set, use the feature co-occurrence analysis algorithm to identify the interaction relationships between different representational features, and record all feature pairs with interaction and their interaction parameters.

[0025] In this embodiment, a feature co-occurrence analysis algorithm is used to analyze each apparent feature in the apparent feature set. For example, the analysis found that the apparent feature "three-phase current imbalance" and other apparent features such as "motor operating state changes to abnormal vibration state" and "motor casing temperature rise" occur simultaneously multiple times during the fault occurrence process, exhibiting a high co-occurrence frequency. Therefore, it is determined that there is an interaction relationship between them. At the same time, the interaction parameters of these feature pairs, such as the number of co-occurrences and the co-occurrence time interval, are recorded.

[0026] Step S132: Extract the interaction information of feature pairs, including the interaction triggering condition, the interaction transmission medium, the interaction influence parameter and the interaction duration parameter. Based on the interaction triggering condition, filter the source features that trigger the interaction, and based on the interaction influence parameter, filter the target features that receive the interaction, and establish the feature interaction mapping relationship.

[0027] In this embodiment, for the feature pairs determined in step S131, such as "three-phase current imbalance" and "motor casing temperature rise", their interaction information is extracted. The triggering condition is that the degree of three-phase current imbalance exceeds a preset threshold; the transmission medium is electromagnetic induction and heat conduction inside the motor; the influence parameters are the rate and magnitude of temperature rise; and the duration parameter is the duration of the current imbalance state. Based on the triggering condition, "three-phase current imbalance" is determined as the source feature that triggers the action; based on the influence parameters, "motor casing temperature rise" is determined as the target feature that receives the action. In this way, a feature-action mapping relationship is established for each feature pair, such as "three-phase current imbalance → motor casing temperature rise".

[0028] Step S133: Based on all feature effect mapping relationships, a feature transmission chain is constructed using a path mining algorithm. The transmission priority of the appearance features in the appearance feature set is set by calculating the number of times the feature initiates and receives its effect in the chain, and the sequence of feature transmission is arranged according to the transmission priority.

[0029] In this embodiment, based on all feature role mapping relationships established in step S132, a path mining algorithm is used to construct feature transmission chains.

[0030] Step S1331: Classify all feature interaction mapping relationships according to the source features to form a mapping subset with the source features as the core. The classified mapping subset contains all target features and interaction information corresponding to the source features.

[0031] In this embodiment, the feature-action mapping relationship is classified according to the source feature. For example, the mapping subset with "three-phase current imbalance" as the source feature includes mapping relationships such as "three-phase current imbalance → motor casing temperature rise" and "three-phase current imbalance → motor operating state changes to abnormal vibration state", as well as corresponding interaction information, such as action triggering conditions and transmission medium.

[0032] Step S1332: Perform path extension analysis on the classified mapping subset, starting from the source feature and using the target feature as the intermediate node or end point, to gradually extend and form multiple feature transfer sub-chains.

[0033] In this embodiment, for the mapping subset with "three-phase current imbalance" as the source feature, starting from "three-phase current imbalance," the target feature "increased motor casing temperature" can be used as an intermediate node to continue extending the search for mapping relationships with "increased motor casing temperature" as the source feature, such as "increased motor casing temperature → aging motor insulation," thus forming a feature transfer sub-chain: "three-phase current imbalance → increased motor casing temperature → aging motor insulation." By performing such path extension analysis on all mapping subsets, multiple feature transfer sub-chains are formed.

[0034] Step S1333: Merge duplicate path segments in all feature transfer sub-chains, remove redundant chains, and form a complete feature transfer chain that includes all representational features.

[0035] In this embodiment, after obtaining multiple feature transmission sub-chains, these sub-chains are inspected, and duplicate path segments are found, such as "three-phase current imbalance → motor operating state changes to abnormal vibration state" appearing in multiple sub-chains. These duplicate path segments are merged, and redundant sub-chains are removed, ultimately forming a complete feature transmission chain containing all apparent features, such as "three-phase current imbalance → motor operating state changes to abnormal vibration state → motor accompanied by abnormal noise → conveyor belt speed decreases → material accumulation in subsequent processes → production line efficiency decreases", "three-phase current imbalance → motor casing temperature rises →...", etc., ensuring that all apparent features are included in this chain.

[0036] Step S1334: Statistically count the number of times a representation feature is initiated and received in the feature transmission chain. The number of times an action is initiated is the number of times the corresponding feature is used as a source feature, and the number of times it is received is the number of times the corresponding feature is used as a target feature.

[0037] In this embodiment, statistics are performed on each apparent feature in the complete feature transmission chain. For example, "three-phase current imbalance" appears multiple times as a source feature in the chain, meaning it is initiated multiple times; it appears zero times as a target feature, meaning it is received zero times. "Motor casing temperature rise" appears multiple times as a source feature, meaning it is initiated multiple times; it appears once as a target feature, meaning it is received once. And so on, the number of times each apparent feature is initiated and received is counted.

[0038] Step S1335: Calculate the transmission priority coefficient of the appearance features in the appearance feature set. The transmission priority coefficient is obtained by the correlation calculation of the number of actions initiated and the number of actions received. Arrange all appearance features in descending order of transmission priority coefficient to determine the order of feature transmission.

[0039] In this embodiment, the conduction priority coefficient is obtained through correlation calculation between the number of action initiations and the number of actions received. For example, the conduction priority coefficient is positively correlated with the number of action initiations and the number of actions received. All apparent features are arranged in descending order of conduction priority coefficient to determine the order of feature transmission. For example, "three-phase current imbalance" is ranked first, followed by "motor casing temperature rise", etc.

[0040] Step S134: Using the apparent features as trajectory nodes and the interaction relationships between features as trajectory connecting edges, arrange the nodes according to the order of feature transmission, and the connecting edges correspond to the interaction information to form the initial fault propagation trajectory.

[0041] In this embodiment, each apparent feature is treated as a trajectory node. For example, "three-phase current imbalance," "motor casing temperature rise," and "motor operating state changes to abnormal vibration state" are each treated as a node. The interaction relationships between features are used as trajectory connecting edges. For example, the interaction relationship between "three-phase current imbalance" and "motor casing temperature rise" is used as the edge connecting the two. Then, these nodes are arranged according to the feature transmission order determined in step S1335, and the connecting edges correspond to interaction information, such as the triggering conditions and transmission medium, thereby forming the initial fault propagation trajectory.

[0042] Step S135: Dynamically optimize the initial fault propagation trajectory, adjust the node connection strength based on the continuous parameters of the feature action, supplement the association type identifier of the connection edge according to the action propagation medium, and add node dependency markers in combination with the feature propagation priority to form a structured fault propagation trajectory containing nodes, connection edges, association type identifiers, node dependency markers and connection strengths.

[0043] In this embodiment, for the initial fault propagation trajectory, the node connection strength is adjusted based on the duration parameter of the characteristic action. For example, if the duration parameter of "three-phase current imbalance" acting on "motor casing temperature rise" is relatively long, the strength of the connection edge between the two is strengthened. Based on the action transmission medium, such as electromagnetic induction or heat conduction, association type identifiers are added to the connection edges, such as "electromagnetic induction association" or "heat conduction association." Combined with the characteristic propagation priority, dependency markers are added to the nodes. For example, "three-phase current imbalance" has a high priority, and its subsequent node "motor casing temperature rise" has a strong dependency on it, so it is marked as "strong dependency." After such dynamic optimization, a structured fault propagation trajectory is formed, which includes information such as nodes, connection edges, association type identifiers, node dependency markers, and connection strength.

[0044] Step S140: Based on the node density, path branch coefficient, and association strength distribution of the structured fault propagation trajectory, generate hierarchical root cause reverse tracing constraints. These constraints include the allowable path range, association type restrictions, and strength thresholds for each tracing level. In this embodiment, based on the structured fault propagation trajectory of the motor fault obtained in step S135, hierarchical root cause reverse tracing constraints are generated by analyzing its node density, path branch coefficient and correlation strength distribution.

[0045] Step S141: Analyze the structured fault propagation trajectory, count the total number of nodes, the total number of connecting edges and the number of path branches in the trajectory, calculate the node density within a unit path length and the path branch coefficient of the nodes in the trajectory. The node density is obtained by the correlation calculation between the total number of nodes and the total length of the trajectory, and the path branch coefficient is obtained by the correlation calculation between the number of branch paths associated with the node and the total number of connecting edges of the node.

[0046] In this embodiment, the structured fault propagation trajectory is analyzed to obtain the total number of nodes, the total number of connecting edges, and the number of path branches. The total trajectory length is obtained by summing the lengths of all connecting edges. The node density is derived by a correlation calculation between the total number of nodes and the total trajectory length; for example, the node density is the ratio of the total number of nodes to the total trajectory length. For each node, the number of its associated branch paths and the total number of connecting edges are counted. The path branch coefficient is derived by a correlation calculation between the number of branch paths associated with the node and the total number of connecting edges; for example, the path branch coefficient is the ratio of the number of branch paths to the total number of connecting edges.

[0047] Step S142: Extract the association strength and association type identifier of the connecting edges, draw the association strength distribution curve, determine the peak interval, mean interval and valley interval of the association strength, and statistically analyze the distribution ratio of different types of associations based on the association type identifier.

[0048] In this embodiment, the association strength and association type identifiers of all connecting edges in the structured fault propagation trajectory are extracted. Distribution curves are plotted based on these association strength values, and the peak, mean, and trough intervals of the association strength are determined through curve analysis. Simultaneously, based on the association type identifiers, the distribution proportions of different association types are statistically analyzed, such as a certain proportion for "electromagnetic induction association," a certain proportion for "thermal conduction association," and a certain proportion for "mechanical vibration association," etc.

[0049] Step S143: Based on the distribution of node density and path branch coefficient, the structured fault propagation trajectory is divided into multiple continuous tracing levels. Each tracing level contains a fixed number of nodes and corresponding connecting edges. The level division boundary is set at a position where both node density and path branch coefficient are at their lowest values.

[0050] In this embodiment, based on the distribution of node density and path branch coefficients calculated in step S141, the location where both node density and path branch coefficients are at their lowest values ​​is identified as the level division boundary. For example, if both node density and path branch coefficients are in a low-value range between two nodes on the trajectory, this location is considered a level division boundary. In this way, the structured fault propagation trajectory is divided into multiple continuous tracing levels, each level containing a certain number of nodes and corresponding connecting edges.

[0051] Step S144: Assign corresponding constraint levels to the divided traceability levels. The constraint level increases as the average association strength of the connecting edges within the corresponding level increases. The constraint level also increases as the constraint criteria tighten. Set the association strength threshold for the corresponding level based on the constraint level. Connecting edges whose association strength does not reach the threshold are not included in the traceability path.

[0052] Step S1441: Calculate the average association strength of all connecting edges within the divided source tracing level, and use it as the average association strength of the corresponding level.

[0053] In this embodiment, for each traceability level after division, the average value of the association strength of all the connection edges contained therein is calculated as the average association strength of that traceability level.

[0054] Step S1442: Sort the average association strength of all source tracing levels, divide the constraint level intervals according to the sorting results, and the average association strength of the higher sorted levels corresponds to higher constraint levels.

[0055] In this embodiment, the average association strength of all source tracing levels is sorted, and different constraint level intervals are divided according to the sorting results, such as high constraint level interval, medium constraint level interval, low constraint level interval, etc. The constraint level interval corresponding to the average association strength of the higher sorting is higher.

[0056] Step S1443: Assign a unique constraint level identifier to each of the divided constraint level intervals. The constraint level identifiers correspond one-to-one with the interval ranges, forming a constraint level mapping table.

[0057] In this embodiment, a unique constraint level identifier is assigned to each constraint level interval, such as "H" for high constraint level intervals, "M" for medium constraint level intervals, and "L" for low constraint level intervals, forming a constraint level mapping table to clarify the correspondence between the average correlation strength intervals and the constraint level identifiers.

[0058] Step S1444: Based on the constraint level range to which the average correlation strength of the divided traceability levels belongs, set the constraint level identifier for the corresponding traceability level.

[0059] In this embodiment, the constraint level range to which each traceability level belongs is determined based on the average association strength of each traceability level. Then, the corresponding constraint level identifier is found from the constraint level mapping table, and a constraint level identifier is set for that traceability level.

[0060] Step S1445: Set the association strength threshold based on the constraint level identifier. The constraint level increases with the association strength threshold. The association strength threshold is a preset proportion of the average association strength of the corresponding level. Connection edges whose association strength does not reach the threshold are not included in the tracing path.

[0061] In this embodiment, an association strength threshold is set based on the constraint level identifier; the higher the constraint level, the higher the association strength threshold. The association strength threshold is a preset proportion of the average association strength of the corresponding level. For example, the association strength threshold for a high constraint level is a relatively high proportion of the average association strength, for a medium constraint level it is a medium proportion, and for a low constraint level it is a relatively low proportion. Connection edges whose association strength does not reach the corresponding threshold will not be included in the tracing path.

[0062] Step S145: Combining the path range, association type distribution ratio and constraint level of the divided tracing levels, define the allowed tracing path direction, limited association type and corresponding strength threshold of each level, and integrate them to form hierarchical root cause reverse tracing constraints, with each tracing level corresponding to an independent constraint sub-condition.

[0063] In this embodiment, for each tracing level, the allowed tracing path direction (such as reverse), the limited association type (such as primary association type), and the corresponding intensity threshold are defined based on its path range, the distribution ratio of association types, and the constraint level. This information is integrated to form hierarchical root cause reverse tracing constraints, with each tracing level corresponding to an independent constraint sub-condition.

[0064] Step S150: Invoke the customized graph neural network, load the structured fault propagation trajectory and hierarchical root cause reverse tracing constraints, perform feature aggregation and weight allocation through dual-track convolutional layers, conduct dual-track tracing reasoning in reverse along the propagation trajectory, and generate the final fault root cause reasoning result by combining the tracing path conflict mediation operation. In this embodiment, a customized graph neural network specifically designed for motor fault root cause reasoning is invoked. The structured fault propagation trajectory obtained in step S135 and the hierarchical root cause reverse tracing constraints generated in step S145 are loaded into the network for subsequent reasoning.

[0065] Step S151: Convert the structured fault propagation trajectory into heterogeneous graph structure data compatible with graph neural networks. The nodes in the heterogeneous graph structure data correspond to the trajectory nodes, and the edges in the heterogeneous graph structure data correspond to the trajectory connection edges. The node attributes include feature attribute information and node dependency markers, and the edge attributes include association type identifiers and connection strength.

[0066] Step S1511: Extract all trajectory nodes and their corresponding feature attribute information and node dependency tags from the structured fault propagation trajectory. Assign exclusive node identifiers to the trajectory nodes in the structured fault propagation trajectory. Bind the node identifiers with the feature attribute information and node dependency tags to form a node attribute set.

[0067] In this embodiment, all trajectory nodes are extracted from the structured fault propagation trajectory. Each node has its corresponding characteristic attribute information, such as the characteristic attribute information of "three-phase current imbalance" including the degree of current imbalance, the time of occurrence, etc., as well as node dependency markers, such as "strong dependency" and "weak dependency". A unique node identifier is assigned to each trajectory node, and the node identifier is bound together with the characteristic attribute information and the node dependency marker to form a node attribute set.

[0068] Step S1512: Extract all trajectory connection edges and their corresponding association type identifiers and connection strengths from the structured fault propagation trajectory. Assign exclusive edge identifiers to the trajectory connection edges in the structured fault propagation trajectory. Bind the edge identifiers with the association type identifiers and connection strengths to form an edge attribute set.

[0069] In this embodiment, all trajectory connection edges are extracted, and each connection edge corresponds to the interaction between two trajectory nodes. A unique edge identifier is assigned to each connection edge, and the edge identifier is bound to the association type identifier (such as "electromagnetic induction association" or "thermal conduction association") and the connection strength to form an edge attribute set.

[0070] Step S1513: Based on the arrangement order and connection relationship of the trajectory nodes, construct the topology of the heterogeneous graph. The connection relationship of the nodes in the topology is consistent with the connection relationship of the nodes in the structured fault propagation trajectory.

[0071] In this embodiment, a heterogeneous graph topology is constructed in the graph neural network based on the arrangement order and connection relationship of the trajectory nodes in the structured fault propagation trajectory. For example, if trajectory node N1 (three-phase current imbalance) is connected to N2 (motor casing temperature rise), a directed edge from N1 to N2 is established in the topology. Its connection relationship is completely consistent with the connection relationship in the structured fault propagation trajectory, ensuring that the heterogeneous graph can accurately reflect the path and direction of fault propagation.

[0072] Step S1514: Embed the information in the node attribute set into the corresponding node of the topology structure, so that the attribute information of the node in the topology structure is completely mapped without missing feature attribute information and node dependency tags.

[0073] In this embodiment, the information from the node attribute set formed in step S1511 is embedded into the corresponding nodes in the heterogeneous graph topology. This operation ensures that each node in the topology has complete attribute information, without any omissions of feature attribute information or node dependency markers.

[0074] Step S1515: Embed the information in the edge attribute set into the corresponding edge of the topology, so that the attribute information of the edge in the topology is fully mapped, without missing the association type identifier and connection strength, forming heterogeneous graph structure data compatible with graph neural networks.

[0075] In this embodiment, the information from the edge attribute set formed in step S1512 is embedded into the corresponding edges in the heterogeneous graph topology. This ensures that each edge in the topology has complete attribute information, with no omissions of association type identifiers or connection strengths, ultimately forming heterogeneous graph structure data that can be directly processed by the graph neural network.

[0076] Step S152: Embed the hierarchical root cause reverse tracing constraints into the heterogeneous graph structure data according to the tracing hierarchy, and add constraint markers to the graph structure parts corresponding to the divided tracing hierarchy. The constraint markers include constraint level, association strength threshold and allowed association types.

[0077] In this embodiment, based on the hierarchical root cause reverse tracing constraints generated in step S145, the constraint information of each tracing level is embedded into the heterogeneous graph structure data. For example, for a certain tracing level (constraint level H, association strength threshold, allowed association types are electromagnetic induction association and thermal conduction association), the graph structure part corresponding to this level is found in the heterogeneous graph structure data, and constraint labels are added to its nodes and edges, such as marking "constraint level: H" on nodes and "association strength threshold:, allowed association types: electromagnetic induction association, thermal conduction association" on edges. In this way, the heterogeneous graph structure data carries constraint information.

[0078] Step S153: Input the heterogeneous graph structure data with constraint labels into the dual-track convolutional layer of the customized graph neural network. The first track convolutional layer performs local neighborhood feature aggregation on node attributes, and the second track convolutional layer performs association strength feature aggregation on edge attributes to generate node and edge fusion features.

[0079] In this embodiment, the customized graph neural network's dual-track convolutional layer includes a first-track convolutional layer and a second-track convolutional layer. When heterogeneous graph structure data with constraint labels is input, the first-track convolutional layer processes the attribute information of each node, extracting local neighborhood features by aggregating the feature attribute information of the node and its local neighbors. For example, for node N2 (motor housing temperature rises), the first-track convolutional layer aggregates the feature attribute information of N2 and its neighboring nodes to generate local neighborhood aggregated features of node N2. The second-track convolutional layer processes edge attributes, aggregating the association strength information of each edge and its associated edges to extract association strength features. For example, for edge E1 (connecting N1 and N2), the second-track convolutional layer aggregates the connection strength of E1 and its associated edges to generate association strength aggregated features of edge E1. Then, the local neighborhood aggregated features of the nodes and the association strength aggregated features of the edges are fused to generate node-edge fused features.

[0080] Step S154: The dynamic attention mechanism layer of the graph neural network is used to assign weights to the fusion features of nodes and edges. The attention weights are dynamically adjusted based on the constraint level in the constraint label, strengthening the feature weights that meet the constraints and weakening the feature weights that do not meet the constraints, thus generating weighted fusion features.

[0081] Step S1541: Extract the feature dimension information from the node and edge fusion features and the constraint level information from the constraint labels, and establish a mapping relationship between feature dimensions and constraint levels.

[0082] In this embodiment, the node and edge fusion features include multiple feature dimensions, such as the node's temperature feature dimension, current feature dimension, vibration feature dimension, etc., and the edge's association strength feature dimension, etc. The constraint level information in the constraint label includes H, M, and L. A mapping relationship between feature dimensions and constraint levels is established; for example, the temperature feature dimension has a specific weight adjustment rule under constraint level H, and another rule under constraint level M, etc.

[0083] Step S1542: Set the attention weight benchmark value based on the constraint level. The constraint level increases as the attention weight benchmark value increases, forming a weight benchmark table.

[0084] In this embodiment, attention weight benchmark values ​​are set for different constraint levels. The higher the constraint level, the higher the attention weight benchmark value, forming a weight benchmark table to clarify the benchmark values ​​corresponding to different constraint levels.

[0085] Step S1543: Analyze the degree of fit between the feature dimension and the corresponding constraint level type restriction and strength threshold through the constraint matching algorithm.

[0086] In this embodiment, the constraint matching algorithm checks whether each feature dimension meets the association type restrictions and strength thresholds under its corresponding constraint level. For example, under constraint level H, if the association type corresponding to a certain feature dimension is "electromagnetic induction association" and its strength value is greater than the strength threshold, then the degree of fit is high; if the association type does not meet the requirements or the strength value is lower than the threshold, then the degree of fit is low.

[0087] Step S1544: Assign reinforcement weights based on the weight benchmark value to feature dimensions that meet the constraints. The reinforcement weights are obtained by the correlation calculation between the weight benchmark value and the feature correlation strength. Assign weakening weights based on the weight benchmark value to feature dimensions that do not meet the constraints. The weakening weights are obtained by the correlation calculation between the weight benchmark value and the constraint deviation coefficient.

[0088] In this embodiment, for feature dimensions that meet the constraints, the strengthening weight is obtained through a correlation calculation between the weight baseline value and the feature association strength. For example, the strengthening weight is the result of a certain combination calculation of the weight baseline value and the feature association strength. For feature dimensions that do not meet the constraints, the weakening weight is obtained through a correlation calculation between the weight baseline value and the constraint deviation coefficient. The constraint deviation coefficient is determined according to the degree of fit; the lower the degree of fit, the larger the deviation coefficient.

[0089] Step S1545: Calculate the weighting of the enhancement or weakening weights of all feature dimensions with the corresponding feature dimension values, integrate all weighted feature dimension values, and generate a weighted fusion feature. The weighted fusion feature retains key feature information that meets the constraints.

[0090] In this embodiment, for each feature dimension, its enhancement or weakening weight is multiplied by its feature dimension value to obtain a weighted feature dimension value. All weighted feature dimension values ​​are then integrated to form a weighted fusion feature. In this feature, key feature information that meets the constraints is highlighted due to its higher weight, while information that does not meet the constraints is weakened.

[0091] Step S155: Along the reverse direction of the structured fault propagation trajectory, starting from the end node of the trajectory, a dual-track tracing traversal is carried out by combining weighted fusion features and constraint labels. The first tracing path is formed by tracing along the node dependency relationship, and the second tracing path is formed by tracing along the path of reaching the correlation strength standard. The two tracing paths together constitute the initial tracing path.

[0092] In this embodiment, the terminal node of the structured fault propagation trajectory is "reduced production line efficiency." Tracing the source begins from this node and proceeds in reverse. The first tracing path follows node dependencies; that is, based on the node's dependency marker, it traces from "reduced production line efficiency" to its dependent node "material accumulation in subsequent processes," then to "conveyor belt speed decrease," and so on. The second tracing path follows paths where the association strength meets the threshold; that is, it only selects paths where the association strength of the connecting edges reaches the corresponding level's strength threshold. For example, starting from "reduced production line efficiency," it checks whether the association strength of the edges connected to it meets the threshold. If it does, it traces back to the previous node, thus forming the second path. These two paths together constitute the initial tracing path.

[0093] Step S156: Perform conflict detection on the initial tracing path, identify conflict points in path nodes, association types and strength thresholds, use a logical consistency analysis algorithm to analyze the logical fit of conflict points, retain logically fit path segments, and integrate them to form a unified tracing path.

[0094] For example, step S1561: Extract all path nodes, association types and corresponding strength thresholds in the initial tracing path, and construct a path information lookup table.

[0095] In this embodiment, path nodes in the initial traceability path, such as "reduced production efficiency of the production line", "material accumulation in subsequent processes", "decrease in conveyor belt speed", etc., as well as the association type and strength threshold of the connecting edges between each node, are extracted to construct a path information reference table, which clearly displays each element in the path and its attributes.

[0096] Step S1562: Compare the corresponding positions of the traceability paths that constitute the initial traceability path one by one, identify the positions where the path nodes are inconsistent, the association types are mismatched, and the intensity threshold does not meet the constraint requirements, and mark them as conflict points.

[0097] In this embodiment, the corresponding positions of the two initial tracing paths are compared. For example, if a node at a certain position in the first path is "increased motor vibration frequency," while the corresponding node at the same position in the second path is "fluctuation in motor speed," then the nodes are inconsistent and marked as conflict points. If the association type is different in the two paths, or if the intensity threshold does not meet the constraint requirements, it is also marked as a conflict point.

[0098] Step S1563: Analyze the logical fit of the identified conflict points using the logical consistency analysis algorithm, and combine the constraint sub-conditions in the hierarchical root cause reverse tracing constraint conditions to screen out logically fit conflict points.

[0099] In this embodiment, the logical consistency analysis algorithm starts from the logical relationship of fault propagation and analyzes whether the nodes and association types of the two paths at the conflict point are logically consistent. For example, for conflict points with inconsistent nodes, it analyzes whether "increased motor vibration frequency" and "fluctuation of motor speed" are related in the motor fault propagation logic, and whether they may be different phenomena caused by the same reason. Combining the association type and strength threshold in the constraint sub-conditions, it determines which node is more logically consistent and filters out the logically consistent conflict points.

[0100] Step S1564: Retain the path segments corresponding to logically consistent conflict points, remove the logically inconsistent path segments, and connect the removed path segments at breakpoints to maintain the continuity of the path.

[0101] In this embodiment, for logically consistent conflict points, the corresponding path segments are retained, while logically inconsistent path segments are discarded. For example, if "increased motor vibration frequency" is judged as logically consistent, the path segment containing this node is retained, while the path segment containing "motor speed fluctuation" is discarded. Then, the discarded path segments are connected at breakpoints to ensure that the path remains continuous from the start point to the end point.

[0102] Step S1565: Supplement the necessary path information at the breakpoint connection point, generate transitional path segments based on the association type and strength threshold of adjacent path segments, and integrate all retained path segments and transitional path segments to form a unified tracing path.

[0103] In this embodiment, at the breakpoint connection, a transitional path segment is generated based on the association type and intensity threshold of adjacent path segments. For example, if the last node of the path segment before the breakpoint is A, the association type is electromagnetic induction association, and the intensity threshold is a certain value; and the first node of the path segment after the breakpoint is C, then a transitional path segment from A to C is generated, with an association type of electromagnetic induction association and an intensity value within the threshold range. The retained path segments and the transitional path segments are then integrated to form a unified traceability path.

[0104] Step S157: Extract the appearance features corresponding to the starting node of the unified tracing path as root cause candidates, repeat the tracing process to obtain multiple root cause candidates, sort all the obtained root cause candidates by correlation, integrate the feature information, tracing path and correlation strength of the root cause candidates, and generate the final fault root cause inference result.

[0105] In this embodiment, the starting node of the unified tracing path is "three-phase current imbalance," and its corresponding symptom features are the root cause candidates. Then, by changing the starting point of the initial tracing or adjusting the constraints, the tracing process of steps S155-S156 is repeated to obtain multiple root cause candidates, such as "motor bearing wear" and "unstable power supply voltage." These root cause candidates are ranked according to their correlation strength and frequency of occurrence in the tracing path. For example, "three-phase current imbalance" has the highest correlation strength and is ranked first. Finally, the feature information of the root cause candidates (such as the degree of current imbalance and the time of occurrence), the tracing path, and the correlation strength are integrated to generate the final fault root cause inference result, indicating that the root cause of the motor fault is most likely three-phase current imbalance, and providing relevant reasoning basis and path information.

[0106] Figure 2 Schematic diagrams are shown of exemplary hardware and software components of a graph neural network-based fault root cause inverse reasoning system 100 that can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the graph neural network-based fault root cause inverse reasoning system 100 and to perform the functions in this application.

[0107] The graph neural network-based fault root cause inverse reasoning system 100 can be a general-purpose server or a special-purpose server; both can be used to implement the graph neural network-based fault root cause inverse reasoning method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0108] For example, a graph neural network-based root cause inversion system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the graph neural network-based root cause inversion system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The graph neural network-based root cause inversion system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0109] For ease of explanation, only one processor is described in the graph neural network-based fault root cause inverse reasoning system 100. However, it should be noted that the graph neural network-based fault root cause inverse reasoning system 100 of this application may also include multiple processors, and therefore the steps executed by one processor described in this application may also be executed jointly by multiple processors or individually. For example, if the processor of the graph neural network-based fault root cause inverse reasoning system 100 executes steps A and B, it should be understood that steps A and B may also be executed jointly by two different processors or individually by one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0110] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned fault root cause reverse reasoning method based on graph neural network is implemented.

[0111] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for reverse reasoning of fault root causes based on graph neural networks, characterized in that, The method includes: Receive a set of fault manifestations after a fault occurs. The set of fault manifestations includes various descriptive information about the external presentation of the fault. The descriptive information covers the state changes, phenomenon parameters and related effects when the fault occurs. The fault manifestation set is subjected to multi-dimensional deconstruction processing, the descriptive information in the fault manifestation set is split into atomic-level expression fragments, irrelevant fragments are removed by association filtering, and the remaining fragments are feature encoded and cross-fused to generate a manifestation feature set. Based on the aforementioned set of appearance features, a dynamic fault propagation trajectory is constructed. By modeling the interaction strength between features, the propagation path is constructed, the order of propagation and the association type are arranged, and a structured fault propagation trajectory containing node dependencies is formed. Based on the node density, path branch coefficient, and association strength distribution of the structured fault propagation trajectory, hierarchical root cause reverse tracing constraints are generated. The constraints include the path allowable range, association type restrictions, and strength thresholds at the tracing level. A customized graph neural network is invoked, and the structured fault propagation trajectory and hierarchical root cause reverse tracing constraints are loaded. Feature aggregation and weight allocation are performed through dual-track convolutional layers. Dual-track tracing reasoning is carried out in reverse along the propagation trajectory. Finally, the root cause reasoning result is generated by combining the source tracing path conflict mediation operation.

2. The fault root cause reverse reasoning method based on graph neural networks according to claim 1, characterized in that, The process involves performing multi-dimensional deconstruction on the fault manifestation set, breaking down the descriptive information in the fault manifestation set into atomic-level representation fragments. After removing irrelevant fragments through correlation filtering, the remaining fragments are subjected to feature encoding and cross-fusion to generate a manifestation feature set, including: The descriptive information in the fault manifestation set is split into multiple dimensions according to the state dimension, phenomenon dimension, and impact dimension. Each dimension corresponds to several independent atomic-level description fragments, and the atomic-level description fragments only describe the specific fault presentation under a single dimension. A preset fault association feature dictionary is retrieved. The fault association feature dictionary contains a set of feature expressions related to the occurrence of faults and expression association weights. Atomic-level expression fragments are compared sentence by sentence with the expressions in the fault association feature dictionary, and the matching degree between the fragment expression and the dictionary expression is calculated. Set a matching degree filtering threshold, retain atomic-level description fragments that match the threshold, and remove irrelevant atomic-level description fragments that do not match the threshold, forming a set of valid descriptions corresponding to the descriptive information in the fault manifestation set; Atomic-level expression fragments in the effective expression set are vector encoded. A pre-trained feature encoder is used to convert textual fragments into fixed-dimensional expression feature vectors, so that the dimension of the expression feature vectors under different dimensions remains the same. All descriptive feature vectors corresponding to the descriptive information in the fault manifestation set are input into the cross-fusion module. Cross-association of features of different dimensions is achieved through element-level multiplication. Key information components in the fusion vector are extracted. Feature calibration is performed on the key information components corresponding to all descriptive information in the fault manifestation set. After eliminating dimensional bias, they are combined to form a manifestation feature set.

3. The fault root cause reverse reasoning method based on graph neural networks according to claim 1, characterized in that, The process of constructing a dynamic fault propagation trajectory based on the aforementioned feature set involves modeling the propagation path, arranging the order of propagation, and defining the association types through feature interaction strength modeling, thus forming a structured fault propagation trajectory containing node dependencies. This includes: The association relationship of the appearance features in the appearance feature set is detected. The feature co-occurrence analysis algorithm is used to identify the interaction relationship between different appearance features and record all feature pairs with interaction and their interaction parameters. Extract interaction information of feature pairs, including interaction triggering conditions, interaction transmission medium, interaction influence parameters and interaction duration parameters; filter source features that trigger the interaction based on the interaction triggering conditions; filter target features that receive the interaction based on the interaction influence parameters; and establish feature-interaction mapping relationship. Based on the mapping relationship of all features, a path mining algorithm is used to construct a feature transmission chain. The transmission priority of the apparent features in the apparent feature set is set by calculating the number of times the feature initiates and receives in the chain, and the transmission order of features is arranged according to the transmission priority. Using apparent features as trajectory nodes and the interaction relationships between features as trajectory connecting edges, the nodes are arranged according to the order of feature transmission, and the connecting edges correspond to the interaction information to form the initial fault propagation trajectory. The initial fault propagation trajectory is dynamically optimized. The node connection strength is adjusted based on the continuous parameters of the feature effect. The association type identifier of the connection edge is supplemented according to the action transmission medium. The node dependency mark is added in combination with the feature propagation priority, forming a structured fault propagation trajectory containing nodes, connection edges, association type identifiers, node dependency marks and connection strength.

4. The fault root cause reverse reasoning method based on graph neural networks according to claim 1, characterized in that, The hierarchical root cause tracing constraints are generated based on the node density, path branch coefficient, and association strength distribution of the structured fault propagation trajectory. These constraints include the allowed path range, association type restrictions, and strength thresholds at each tracing level, including: The structured fault propagation trajectory is analyzed, and the total number of nodes, the total number of connecting edges, and the number of path branches in the trajectory are counted. The node density within a unit path length and the path branch coefficient of the nodes in the trajectory are calculated. The node density is obtained by the correlation calculation between the total number of nodes and the total length of the trajectory, and the path branch coefficient is obtained by the correlation calculation between the number of branch paths associated with the node and the total number of connecting edges of the node. Extract the association strength and association type identifiers of the connecting edges, plot the association strength distribution curve, determine the peak interval, mean interval and valley interval of the association strength, and statistically analyze the distribution ratio of different types of associations based on the association type identifiers; Based on the distribution of node density and path branch coefficient, the structured fault propagation trajectory is divided into multiple continuous tracing levels. Each tracing level contains a fixed number of nodes and corresponding connecting edges. The level division boundary is set at a position where both node density and path branch coefficient are at their lowest values. The corresponding constraint level is assigned to the divided traceability level. The constraint level increases with the increase of the average association strength of the connecting edges in the corresponding level. The constraint level increases with the tightening of the constraint standard. Based on the constraint level, the association strength threshold of the corresponding level is set. Connecting edges whose association strength does not reach the threshold are not included in the traceability path. By combining the path range, distribution ratio of association types, and constraint levels of the divided tracing levels, the allowed tracing path directions, limited association types, and corresponding strength thresholds of each level are defined, and integrated to form hierarchical root cause reverse tracing constraints, with each tracing level corresponding to an independent constraint sub-condition.

5. The fault root cause reverse reasoning method based on graph neural networks according to claim 1, characterized in that, The process involves calling a customized graph neural network, loading the structured fault propagation trajectory and hierarchical root cause reverse tracing constraints, performing feature aggregation and weight allocation through dual-track convolutional layers, conducting dual-track tracing inference along the propagation trajectory, and generating the final fault root cause inference result by combining the tracing path conflict resolution operation. This includes: The structured fault propagation trajectory is transformed into heterogeneous graph structure data compatible with graph neural networks. The nodes in the heterogeneous graph structure data correspond to the trajectory nodes, and the edges in the heterogeneous graph structure data correspond to the trajectory connection edges. The node attributes include feature attribute information and node dependency markers, and the edge attributes include association type identifiers and connection strength. The hierarchical root cause reverse tracing constraints are embedded into the heterogeneous graph structure data according to the tracing hierarchy. Constraint tags are added to the graph structure part corresponding to the divided tracing hierarchy. The constraint tags include the constraint level, the association strength threshold, and the allowed association types. Heterogeneous graph structure data with constraint labels is input into a dual-track convolutional layer of a customized graph neural network. The first track convolutional layer performs local neighborhood feature aggregation on node attributes, and the second track convolutional layer performs association strength feature aggregation on edge attributes to generate node and edge fusion features. The dynamic attention mechanism layer of the graph neural network assigns weights to the fusion features of nodes and edges. The attention weights are dynamically adjusted based on the constraint level in the constraint label, strengthening the weights of features that meet the constraints and weakening the weights of features that do not meet the constraints, thus generating weighted fusion features. Along the reverse direction of the structured fault propagation trajectory, starting from the end node of the trajectory, a dual-track tracing traversal is carried out by combining weighted fusion features and constraint labels. The first tracing path is formed by tracing along the node dependency relationship, and the second tracing path is formed by tracing along the path of reaching the correlation strength standard. The two tracing paths together constitute the initial tracing path. Conflict detection is performed on the initial tracing path to identify conflict points in path nodes, association types, and strength thresholds. A logical consistency analysis algorithm is used to analyze the logical fit of the conflict points, retain logically fit path segments, and integrate them to form a unified tracing path. Extract the apparent features corresponding to the starting node of the unified tracing path as root cause candidates, repeat the tracing process to obtain multiple root cause candidates, sort all the obtained root cause candidates by correlation, integrate the feature information, tracing path and correlation strength of the root cause candidates, and generate the final fault root cause inference result.

6. The fault root cause reverse reasoning method based on graph neural networks according to claim 5, characterized in that, The process of converting the structured fault propagation trajectory into heterogeneous graph structure data compatible with graph neural networks is described. Nodes in the heterogeneous graph structure data correspond to trajectory nodes, and edges correspond to trajectory connection edges. Node attributes include feature attribute information and node dependency markers, while edge attributes include association type identifiers and connection strength. Extract all trajectory nodes and their corresponding feature attribute information and node dependency tags from the structured fault propagation trajectory. Assign unique node identifiers to the trajectory nodes in the structured fault propagation trajectory. Bind the node identifiers with the feature attribute information and node dependency tags to form a node attribute set. Extract all trajectory connection edges and their corresponding association type identifiers and connection strengths from the structured fault propagation trajectory. Assign unique edge identifiers to the trajectory connection edges in the structured fault propagation trajectory. Bind the edge identifiers with the association type identifiers and connection strengths to form an edge attribute set. Based on the arrangement order and connection relationship of trajectory nodes, a topology of heterogeneous graph is constructed. The connection relationship of nodes in the topology is consistent with the connection relationship of nodes in the structured fault propagation trajectory. The information in the node attribute set is embedded into the corresponding node of the topology structure, so that the attribute information of the node in the topology structure is completely mapped without missing feature attribute information and node dependency tags. By embedding information from the edge attribute set into the corresponding edges of the topology, the attribute information of the edges in the topology is fully mapped, without omitting association type identifiers and connection strengths, thus forming heterogeneous graph structure data compatible with graph neural networks.

7. The fault root cause reverse reasoning method based on graph neural networks according to claim 5, characterized in that, The dynamic attention mechanism layer of the graph neural network assigns weights to the fused features of nodes and edges, dynamically adjusting the attention weights based on the constraint level in the constraint label, strengthening the weights of features that meet the constraints, and weakening the weights of features that do not meet the constraints, generating weighted fused features, including: Extract the feature dimension information from the node and edge fusion features and the constraint level information from the constraint labels, and establish a mapping relationship between feature dimensions and constraint levels; Attention weight benchmark values ​​are set based on constraint levels. The constraint level increases as the attention weight benchmark value increases, forming a weight benchmark table. The constraint matching algorithm is used to analyze the degree of fit between the feature dimensions and the corresponding constraint levels, as well as the type of restrictions and strength thresholds. For feature dimensions that meet the constraints, strengthen weights are assigned based on the weight benchmark value. The strengthen weights are obtained by the correlation calculation between the weight benchmark value and the feature correlation strength. For feature dimensions that do not meet the constraints, weaken weights are assigned based on the weight benchmark value. The weaken weights are obtained by the correlation calculation between the weight benchmark value and the constraint deviation coefficient. The enhancement or weakening weights of all feature dimensions are weighted and calculated together with the corresponding feature dimension values. All weighted feature dimension values ​​are then integrated to generate a weighted fusion feature. The weighted fusion feature retains key feature information that meets the constraints.

8. The fault root cause reverse reasoning method based on graph neural networks according to claim 3, characterized in that, Based on the mapping relationship of all features, a path mining algorithm is used to construct a feature transmission chain. The transmission priority of the apparent features in the apparent feature set is determined by calculating the number of times a feature initiates and receives its role in the chain. The sequence of feature transmission is then arranged according to this priority, including: All feature interaction mapping relationships are classified according to the source features, forming a mapping subset with the source features as the core. The classified mapping subset contains all target features and interaction information corresponding to the source features. Path extension analysis is performed on the classified mapping subset, starting from the source feature and using the target feature as the intermediate node or end point, to gradually extend and form multiple feature transfer sub-chains; Merge duplicate path segments in all feature transfer sub-chains, remove redundant chains, and form a complete feature transfer chain that includes all apparent features; The number of times a representation feature is initiated and received in the feature transmission chain within the statistical representation feature set. The number of times an action is initiated is the number of times the corresponding feature is used as a source feature, and the number of times it is received is the number of times the corresponding feature is used as a target feature. Calculate the transmission priority coefficient of the appearance features in the appearance feature set. The transmission priority coefficient is obtained by the correlation calculation of the number of actions initiated and received. Arrange all appearance features in descending order of transmission priority coefficient to determine the order of feature transmission.

9. The fault root cause reverse reasoning method based on graph neural networks according to claim 4, characterized in that, The first step involves assigning corresponding constraint levels to the divided source tracing levels. The constraint level increases with the average association strength of the connected edges within the corresponding level, and also increases with tightening of the constraint criteria. Based on the constraint level, a threshold for the association strength of the corresponding level is set, including: Calculate the average association strength of all connecting edges within the divided source tracing hierarchy, and use it as the average association strength of the corresponding hierarchy; Sort the average association strength of all source tracing levels, divide the constraint level intervals according to the sorting results, and the average association strength of the higher sorting level corresponds to the higher constraint level. Each constraining level interval is assigned a unique constraining level identifier, and the constraining level identifier corresponds one-to-one with the interval range, forming a constraining level mapping table. Based on the constraint level range to which the average correlation strength of the divided traceability levels belongs, set the constraint level identifier for the corresponding traceability level. The association strength threshold is set based on the constraint level identifier. The constraint level increases as the association strength threshold increases. The association strength threshold is a preset proportion of the average association strength of the corresponding level. Connection edges whose association strength does not reach the threshold are not included in the tracing path.

10. A fault root cause reverse reasoning system based on graph neural networks, characterized in that, The fault root cause reverse reasoning system based on graph neural networks includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the fault root cause reverse reasoning method based on graph neural networks as described in any one of claims 1-9.

Images