A large model root cause stable reasoning method based on consistency graph constraint

Abstract

The application discloses a large model root cause stable reasoning method based on consistency graph constraint. The method comprises the following steps: acquiring alarm data flow, device topology and knowledge base information; constructing an alarm-device structure diagram and generating an alarm candidate relationship diagram; determining a certain alarm propagation skeleton through rule construction; under the constraint of the propagation skeleton, calling a large language model multiple times to generate a candidate propagation tree and performing structure standardization; based on consistency statistics of the candidate propagation tree and skeleton structure weight, constructing a weighted consistency graph and decoding to obtain a stable propagation tree; and finally generating root cause propagation explanation information and a root cause device. Through the certain propagation skeleton and the consistency constraint mechanism, the application realizes stable reasoning of the large model root cause propagation structure, reduces the propagation path drift problem in a complex alarm scene, improves the stability, interpretability and auditability of the root cause analysis result, and is suitable for intelligent operation and maintenance scenes of communication networks, cloud computing platforms and complex systems.

Description

Technical Field

[0001] This invention belongs to the field of large language model-enabled AIOps fault diagnosis, specifically involving a large model root cause stable inference method based on consistency graph constraints. Background Technology

[0002] With the rapid development of 5G mobile communication technology, network architecture is evolving deeply towards cloud-native and microservice architectures, leading to an exponential increase in network scale and complexity. The future mobile internet will carry even more massive amounts of alarm data. To address the tens of millions of fault alarms caused by network expansion, and considering the advantages of large language models in logical reasoning, semantic understanding, and user interaction, as well as the future research trend of intelligent agent communication, AI-based intelligent operation and maintenance has developed rapidly, and the application of large models in fault root cause localization is maturing. Although current large models have demonstrated powerful reasoning capabilities, considering the stringent deterministic requirements of telecommunications networks, large models can exhibit randomness and "illusion" problems when handling complex and massive alarms. This can easily generate alarm relationship edges that do not conform to the network topology, resulting in random, unreproducible, or unreliable root cause propagation chains. This invention proposes constructing a propagation skeleton that includes the topological relationships between devices and the actual alarms issued by the devices. By utilizing graph structures combined with the reasoning capabilities of large models, a stable output of the fault propagation chain can be achieved. Summary of the Invention

[0003] To address the aforementioned issues, this invention discloses a large-model root cause stable reasoning method based on consistency graph constraints. Addressing the pain point of unknown correlations between alarms, it projects probabilistic thought chain reasoning onto a deterministic topological correlation space, utilizes multi-path sampling voting and structured reconstruction algorithms to eliminate model illusions, and outputs a fault propagation tree and its semantic interpretation that strictly adhere to physical architecture constraints. This ensures the determinism and logical consistency of the large-model output results in telecommunications operation and maintenance scenarios where accuracy is paramount.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A root cause-stable inference method for large models based on consistency graph constraints includes the following steps:

[0006] S1. Obtain the original alarm data stream, device topology information, and knowledge base information. Clean, compress, reconstruct, and enrich the knowledge of the original alarm data stream to obtain a standardized alarm set.

[0007] S2. Based on the standardized alarm set, device nodes and device topology, construct an alarm-device structure diagram containing alarm nodes and device nodes, and calculate the importance weight of device nodes.

[0008] S3. Based on the alarm-device structure diagram, construct a bidirectional mapping relationship between alarms and devices, calculate the correlation weight between alarms through device structure evidence, and obtain an alarm-alarm candidate relationship diagram.

[0009] S4. Perform pruning, connectivity analysis, and skeleton construction on the alarm-alarm candidate relationship graph to generate a global propagation skeleton graph, and transform the global propagation skeleton graph into a structured semantic representation that can be parsed by a large language model.

[0010] S5. Under the constraints of the global propagation skeleton graph, the large language model is called multiple times to generate candidate propagation trees, and the candidate propagation trees are subjected to legality normalization processing to obtain a set of normalized candidate propagation trees.

[0011] S6. Perform edge-level consistency statistics on the normalized candidate propagation tree set, and construct a weighted consistency graph by combining the edge weights in the global propagation skeleton graph.

[0012] S7. Based on the weighted consistency graph, perform stable propagation tree decoding under the constraints of the global propagation skeleton graph to obtain the final stable propagation tree.

[0013] S8. Input the final stable propagation tree into the large language model to generate the corresponding root cause propagation explanation information, and output the root cause alarm, root cause device and propagation path results.

[0014] Furthermore, in step S2, constructing the alarm-device structure diagram includes the following steps:

[0015] S201. Construct a graph structure containing alarm nodes and device nodes, wherein CAUSES association is established between alarm nodes and corresponding device nodes, and PEER_DEVICE association is established between device nodes based on topological connection relationships.

[0016] S202. Different basic weights are assigned to different types of relationships, with the CAUSES relationship having a greater weight than the PEER_DEVICE relationship.

[0017] S203. Calculate the importance weight of the device nodes based on the alarm-device structure diagram, and use the importance weight as a measure of the importance of the device structure.

[0018] Furthermore, in step S3, constructing the alarm-alarm candidate relationship graph includes the following steps:

[0019] S301. Establish the mapping relationship between alarm nodes and device nodes. And the inverse mapping relationship from device nodes to alarm nodes. ;

[0020] S302, Set This represents the total number of alarm nodes. To connect with device nodes The number of connected alarms defines the device node. The inverse coverage penalty term is:

[0021] ;

[0022] in, This indicates the total number of alarm nodes. Indicates the relationship with device nodes The number of associated alarms.

[0023] S303, combined with equipment nodes Importance weight The structural evidence for the equipment is defined as follows:

[0024] ;

[0025] in, Represents device node Importance weights Represents device node The inverse coverage penalty term, e(D) represents the device node. The structural evidence value.

[0026] S304, For any pair of alarm nodes and If the two devices share at least one device node, then a candidate association edge is established between them, and its weight is defined as the sum of the evidence from the top m devices with the highest evidence values ​​among the shared devices:

[0027] ;

[0028] in, Indicates alarm node The corresponding set of device nodes, Indicates alarm node The corresponding set of device nodes, This indicates that the top-ranked items are selected after sorting by structural evidence value. Each device node. Represents device node Structural evidence value, This indicates the weight of the candidate association edges between alarm nodes. Select the number of shared device nodes.

[0029] Furthermore, in step S4, generating the global propagation skeleton graph includes the following steps:

[0030] S401. Perform Mutual Top-K pruning on the alarm-alarm candidate relationship graph. For each alarm node, only retain the top k neighbor nodes with the highest edge weight.

[0031] S402. The corresponding alarm association edge is retained only when two alarm nodes are each other's first k neighbor nodes;

[0032] S403. Perform connected component analysis on the pruned alarm relationship graph and construct the maximum spanning tree for each connected component to obtain the local propagation skeleton.

[0033] S404. Select the cross-component edge with the highest weight between different connected components, bridge multiple local propagation skeletons, and generate a global propagation skeleton graph.

[0034] S405. The global propagation skeleton diagram is represented as follows:

[0035] ;

[0036] in, Represents the set of alarm nodes. Describes the set of allowed edges on the skeleton. Representing an edge The structural weights.

[0037] Furthermore, in step S5, the normalization process for the candidate propagation tree includes the following steps:

[0038] S501. Under the constraint of the global propagation skeleton graph, the large language model is called multiple times to generate a set of candidate propagation trees:

[0039] ;

[0040] S502. Perform normalization processing on each candidate propagation tree:

[0041] ;

[0042] in, Indicates the first A candidate propagation tree, Represents the global propagation skeleton diagram. This represents the normalization function of the propagation tree based on the constraints of the global propagation skeleton graph. Represents the normalized first... A candidate propagation tree.

[0043] S503, the normalization process includes out-of-bounds edge deletion, multi-parent node conflict resolution, local loop repair, and connectivity repair based on skeleton path;

[0044] S504. The normalized candidate propagation tree satisfies the following constraints:

[0045] (1) All propagation edges belong to the set of allowed edges on the skeleton;

[0046] (2) Each non-root node retains at most one parent node;

[0047] (3) The propagation structure is entirely acyclic;

[0048] (4) Key alarm nodes maintain reachability and connectivity.

[0049] Furthermore, in step S6, constructing the weighted consistency graph includes the following steps:

[0050] S601, Counting Arbitrary Directed Edges Frequency of occurrence in the normalized candidate propagation tree set:

[0051] ;

[0052] in, For indicator functions, This represents the total number of normalized candidate propagation trees. Indicates the first The set of edges in a normalized candidate propagation tree. Representing an edge The frequency of occurrence This indicates the candidate propagation tree number.

[0053] S602, Combining the structural weights in the global propagation skeleton diagram Define edge The consistency score is:

[0054] ;

[0055] in, This represents a monotonic transformation function that performs normalization or compression on the structural weights. Representing an edge The frequency of occurrence Representing an edge structural weights, Representing an edge Consistency score.

[0056] S603. Constructing a weighted consistency graph based on consistency scores:

[0057] ;

[0058] in, Represents the set of alarm nodes. Denotes the set of candidate propagation edges. This represents the consistency score of the corresponding edge.

[0059] Furthermore, in step S7, generating the final stable propagation tree includes the following steps:

[0060] S701. Based on the consistency score, select the candidate parent edge with the highest score for each node;

[0061] S702. Perform global loop cancellation and connectivity repair on the obtained propagation structure;

[0062] S703. Under the constraints of single root, single parent, acyclic, and reachable critical nodes, the final stable propagation tree is obtained through decoding. .

[0063] Furthermore, in step S8, the large language model is only used to generate natural language explanation information based on the final stable propagation tree, and does not participate in the generation of propagation structure and the decision of propagation path.

[0064] The beneficial effects of this invention are as follows:

[0065] The method proposed in this invention employs a deterministic global propagation skeleton diagram and a consistent reasoning approach, which can significantly improve the stability and reproducibility of root cause propagation path analysis results; effectively suppress noise propagation relationships and reduce erroneous associations in complex alarm scenarios; restrict the use of large models within deterministic structural constraints, thereby improving the credibility of inference; the output results have a clear propagation structure and natural language explanation, making them easy for operation and maintenance personnel to understand and audit; the method has good versatility and can be adapted to system environments of different sizes and topologies. Attached Figure Description

[0066] Figure 1 The flowchart of the large model root cause stable inference method based on consistency graph constraints of this invention.

[0067] Figure 2 This invention presents a large-scale model root cause stable inference system architecture diagram based on consistency graph constraints.

[0068] Figure 3 This invention presents an application scenario diagram for a large-scale model root cause stable inference system based on consistency graph constraints.

[0069] Figure 4 The present invention provides a schematic diagram of the propagation structure stability of the large model root cause stable inference method based on consistency graph constraints in different independent inference instances, where (a) and (b) are the final stable propagation trees generated by two independent executions, respectively. Detailed Implementation

[0070] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0071] like Figure 1 As shown, the specific steps of the large model root cause stable inference method based on consistency graph constraints described in this invention are as follows:

[0072] S1. Obtain the original alarm data stream, device topology information, and knowledge base information. Clean, compress, reconstruct, and enrich the knowledge of the original alarm data stream to obtain a standardized alarm set.

[0073] S2. Construct an alarm-device structure diagram and extract device structure context information. The specific content is as follows:

[0074] S201. Construct a graph structure containing alarm nodes and device nodes, wherein a CAUSES association is established between alarm nodes and their corresponding device nodes, and a PEER_DEVICE association is established between device nodes based on the device topology connection relationship.

[0075] S202. Assign different base weights to the CAUSES relationship and the PEER_DEVICE relationship, with the CAUSES relationship having a greater weight than the PEER_DEVICE relationship.

[0076] S203. Calculate the importance weight PR(D) of the device node based on the alarm-device structure diagram, which is used to characterize the propagation importance of the device node in the overall structure.

[0077] S3. Construct an alarm-alarm candidate relationship graph, the details of which are as follows:

[0078] S301. Establish the mapping relationship between alarm nodes and device nodes. And the inverse mapping relationship from device nodes to alarm nodes. ;

[0079] S302. Let N be the total number of alarm nodes. This represents the number of alarms associated with device node D. To reduce spurious propagation caused by highly connected devices, the inverse coverage penalty term for device node D is defined as follows:

[0080] ;

[0081] in, Indicates the relationship with device nodes The number of associated alarms.

[0082] S303, Combining the importance weight of device node D Evidence regarding the structure of the computing device:

[0083] ;

[0084] in, Represents device node Inverse coverage penalty term,

[0085] S304, For any pair of alarm nodes and If the two devices share at least one device node, then a candidate association edge is established between them, and its weight is defined as the sum of the evidence from the top m devices with the highest evidence values ​​among the shared devices:

[0086] ;

[0087] in, Indicates alarm node The corresponding set of device nodes, Indicates alarm node The corresponding set of device nodes, This indicates that the top-ranked items are selected after sorting by structural evidence value. Each device node. Represents device node Structural evidence value, This indicates the weight of the candidate association edges between alarm nodes. Select the number of shared device nodes. In this embodiment, the default value is [number to be filled in]. =2.

[0088] S4. Construct the global propagation skeleton diagram, the details of which are as follows:

[0089] S401. Perform Mutual Top-K pruning on the alarm-alarm candidate relationship graph. For each alarm node, only retain the top k neighbor nodes with the highest edge weight.

[0090] S402. The corresponding alarm association edge is retained only when two alarm nodes are each other's first k neighbor nodes. In this embodiment, k=3 is the default value.

[0091] S403. Perform connected component analysis on the pruned alarm graph and construct the maximum spanning tree for each connected component to obtain the local propagation skeleton.

[0092] S404. Select the cross-component edge with the highest weight between different connected components, bridge multiple local propagation skeletons, and generate a global propagation skeleton graph.

[0093] S405, The final global propagation skeleton diagram is represented as follows:

[0094] ;

[0095] in, Represents the set of alarm nodes. Describes the set of allowed edges on the skeleton. Representing an edge The structural weights.

[0096] S5. Perform consistency constraint candidate propagation tree sampling under the constraints of the global propagation skeleton graph. The specific details are as follows:

[0097] S501. For the same alarm stream, call the large language model multiple times to generate a set of candidate propagation trees:

[0098] ;

[0099] S502. Perform normalization processing on each candidate propagation tree:

[0100] ;

[0101] in, Indicates the first A candidate propagation tree, Represents the global propagation skeleton diagram. This represents the normalization function of the propagation tree based on the constraints of the global propagation skeleton graph. Represents the normalized first... A candidate propagation tree.

[0102] S503, the normalization process includes out-of-bounds edge deletion, multi-parent node conflict resolution, local loop repair, and connectivity repair based on skeleton path.

[0103] S504. The normalized candidate propagation tree satisfies the following constraints:

[0104] (1) All propagation edges belong to the set of allowed edges on the skeleton;

[0105] (2) Each non-root node retains at most one parent node;

[0106] (3) The propagation structure is entirely acyclic;

[0107] (4) Key alarm nodes maintain reachability and connectivity.

[0108] S6. Construct a weighted consistency graph, the details of which are as follows:

[0109] S601, Counting Arbitrary Directed Edges Frequency of occurrence in the normalized candidate propagation tree:

[0110] ;

[0111] in, For indicator functions; This represents the total number of normalized candidate propagation trees. Indicates the first The set of edges in a normalized candidate propagation tree. Representing an edge The frequency of occurrence This indicates the candidate propagation tree number.

[0112] S602, Combining the structural weights in the global propagation skeleton diagram Define edge The consistency score is:

[0113] ;

[0114] in, This represents a monotonic transformation function that performs normalization or compression on the structural weights. Representing an edge The frequency of occurrence Representing an edge structural weights, Representing an edge Consistency score.

[0115] S603. Constructing a weighted consistency graph based on consistency scores:

[0116] ;

[0117] in, Denotes the set of candidate propagation edges. This represents the consistency score of the corresponding edge.

[0118] S7. Perform stable propagation tree decoding based on the weighted consistency graph to obtain the final stable propagation tree. The system first selects the candidate parent edge with the highest score for each node based on the consistency score. Then, it performs global loop elimination and connectivity repair to ensure that the final propagation structure satisfies the constraints of single root, single parent, no loops, and reachability of critical nodes.

[0119] S8, Based on the final stable propagation tree Generate root cause propagation explanation information and output root cause alarms, root cause devices, and key propagation path results. The final stable propagation tree and its propagation relationships can be further visualized using a graph structure, such as... Figure 4 As shown. In this stage, the large language model is only used to generate natural language interpretations and does not participate in the generation of propagation structure or the decision-making of propagation path.

[0120] This invention transforms the open-ended large model generation process into a structure-constrained propagation reasoning process by introducing a deterministic propagation skeleton and a consistency-constrained reasoning mechanism. This enables stable recovery and consistent convergence of the root cause propagation structure in complex alarm scenarios. At the same time, through multi-round consistent aggregation and global stable decoding, it effectively reduces the randomness of propagation paths and the drift of results, and improves the stability, reproducibility, and engineering auditability of the root cause propagation structure.

[0121] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that when the processor executes the computer program, it corresponds to the specific steps of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0122] This invention also proposes a computer-readable storage medium storing a computer program. It should be noted that when the computer program is executed by a processor, it corresponds to the specific steps of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0123] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A root cause stable inference method for large models based on consistency graph constraints, characterized in that, include: S1. Obtain the original alarm data stream, device topology information, and knowledge base information. Clean, compress, reconstruct, and enrich the knowledge of the original alarm data stream to obtain a standardized alarm set. S2. Based on the standardized alarm set, device nodes and device topology, construct an alarm-device structure diagram containing alarm nodes and device nodes, and calculate the importance weight of device nodes; S3. Based on the alarm-device structure diagram, construct a bidirectional mapping relationship between alarms and devices, calculate the correlation weight between alarms through device structure evidence, and obtain an alarm-alarm candidate relationship diagram; S4. Perform pruning, connectivity analysis, and skeleton construction on the alarm-alarm candidate relationship graph to generate a global propagation skeleton graph, and transform the global propagation skeleton graph into a structured semantic representation that can be parsed by a large language model; S5. Under the constraints of the global propagation skeleton graph, the large language model is called multiple times to generate candidate propagation trees, and the legality normalization processing of the candidate propagation trees is performed to obtain a set of normalized candidate propagation trees. S6. Perform edge-level consistency statistics on the normalized candidate propagation tree set, and construct a weighted consistency graph by combining the edge weights in the global propagation skeleton graph; S7. Based on the weighted consistency graph, perform stable propagation tree decoding under the constraints of the global propagation skeleton graph to obtain the final stable propagation tree; S8. Input the final stable propagation tree into the large language model to generate the corresponding root cause propagation explanation information, and output the root cause alarm, root cause device and propagation path results.

2. The large-model root cause stable inference method based on consistency graph constraints according to claim 1, characterized in that, In step S2, constructing the alarm-device structure diagram includes the following steps: S201. Construct a graph structure containing alarm nodes and device nodes, wherein CAUSES association is established between alarm nodes and corresponding device nodes, and PEER_DEVICE association is established between device nodes based on topological connection relationships. S202. Different basic weights are assigned to different types of relationships, with the CAUSES relationship having a greater weight than the PEER_DEVICE relationship. S203. Calculate the importance weight of the device nodes based on the alarm-device structure diagram, and use the importance weight as a measure of the importance of the device structure.

3. The large-scale model root cause stable inference method based on consistency graph constraints according to claim 1, characterized in that, In step S3, constructing the alarm-alarm candidate relationship graph includes the following steps: S301. Establish the mapping relationship between alarm nodes and device nodes. And the inverse mapping relationship from device nodes to alarm nodes. ; S302, Set This represents the total number of alarm nodes. To connect with device nodes The number of connected alarms defines the device node. The inverse coverage penalty term is: ； in, This indicates the total number of alarm nodes. Indicates the relationship with device nodes The number of associated alarms; S303, combined with equipment nodes Importance weight The structural evidence for the equipment is defined as follows: ； in, Represents device node Importance weights Represents device node The inverse coverage penalty term, e(D) represents the device node. Structural evidence value; S304, For any pair of alarm nodes and If the two devices share at least one device node, then a candidate association edge is established between them, and its weight is defined as the sum of the evidence from the top m devices with the highest evidence values ​​among the shared devices: ； in, Indicates alarm node The corresponding set of device nodes, Indicates alarm node The corresponding set of device nodes, This indicates that the top-ranked items are selected after sorting by structural evidence value. Each device node. Represents device node Structural evidence value, This indicates the weight of the candidate association edges between alarm nodes. Select the number of shared device nodes.

4. The large model root cause stable inference method based on consistency graph constraints according to claim 1, characterized in that, In step S4, generating the global propagation skeleton diagram includes the following steps: S401. Perform Mutual Top-K pruning on the alarm-alarm candidate relationship graph. For each alarm node, only retain the top k neighbor nodes with the highest edge weight. S402. The corresponding alarm association edge is retained only when two alarm nodes are each other's first k neighbor nodes; S403. Perform connected component analysis on the pruned alarm relationship graph and construct the maximum spanning tree for each connected component to obtain the local propagation skeleton. S404. Select the cross-component edge with the highest weight between different connected components, bridge multiple local propagation skeletons, and generate a global propagation skeleton graph. S405. The global propagation skeleton diagram is represented as follows: ； in, Represents the set of alarm nodes. Describes the set of allowed edges on the skeleton. Representing an edge The structural weights.

5. The large model root cause stable inference method based on consistency graph constraints according to claim 1, characterized in that, Step S5, which involves normalizing the candidate propagation tree, includes the following steps: S501. Under the constraint of the global propagation skeleton graph, the large language model is called multiple times to generate a set of candidate propagation trees: ； S502. Perform normalization processing on each candidate propagation tree: ； in, Indicates the first A candidate propagation tree, Represents the global propagation skeleton diagram. This represents the normalization function of the propagation tree based on the constraints of the global propagation skeleton graph. Represents the normalized first... One candidate propagation tree; S503, the normalization process includes out-of-bounds edge deletion, multi-parent node conflict resolution, local loop repair, and connectivity repair based on skeleton path; S504. The normalized candidate propagation tree satisfies the following constraints: (1) All propagation edges belong to the set of allowed edges on the skeleton; (2) Each non-root node retains at most one parent node; (3) The propagation structure is entirely acyclic; (4) Key alarm nodes maintain reachability and connectivity.

6. The large model root cause stable inference method based on consistency graph constraints according to claim 1, characterized in that, In step S6, constructing the weighted consistency graph includes the following steps: S601, Counting Arbitrary Directed Edges Frequency of occurrence in the normalized candidate propagation tree set: ； in, For indicator functions, This represents the total number of normalized candidate propagation trees. Indicates the first The set of edges in a normalized candidate propagation tree. Representing an edge The frequency of occurrence Indicates the candidate propagation tree number; S602, Combining the structural weights in the global propagation skeleton diagram Define edge The consistency score is: ； in, This represents a monotonic transformation function that performs normalization or compression on the structural weights. Representing an edge The frequency of occurrence Representing an edge structural weights, Representing an edge Consistency score; S603. Constructing a weighted consistency graph based on consistency scores: ； in, Represents the set of alarm nodes. Denotes the set of candidate propagation edges. This represents the consistency score of the corresponding edge.

7. The large model root cause stable inference method based on consistency graph constraints according to claim 1, characterized in that, Step S7, generating the final stable propagation tree, includes the following steps: S701. Based on the consistency score, select the candidate parent edge with the highest score for each node; S702. Perform global loop cancellation and connectivity repair on the obtained propagation structure; S703. Under the constraints of single root, single parent, no cycles, and key nodes meeting the requirements, the final stable propagation tree is obtained through decoding. .

8. The large model root cause stable inference method based on consistency graph constraints according to claim 1, characterized in that, In step S8, the large language model is only used to generate natural language explanation information based on the final stable propagation tree, and does not participate in the generation of propagation structure and the decision of propagation path.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the large model root cause stable inference method based on consistency graph constraints as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the large model root cause stable inference method based on consistency graph constraints as described in any one of claims 1 to 8.

Images