A device fault diagnosis method based on knowledge sub-graph uncertainty convergence

Abstract

The application discloses a kind of based on knowledge subgraph uncertainty convergence equipment fault diagnosis method, comprising the following steps: obtaining and pre-processing equipment fault natural language description text;Extract fault feature and map to fault knowledge graph;Secondary screening is carried out to fault feature node set;Upstream influencing factor node is backtracked in fault knowledge graph layer by layer, and intersection is obtained to obtain candidate root factor graph node set;Candidate root factor graph is constructed and root source influencing factor node set is determined;Determine uncertainty index and carry out threshold comparison;When uncertainty index exceeds threshold, generate directional question and update fault feature set, iterative calculation until threshold is satisfied;When uncertainty index meets threshold, obtain document input set and carry out semantic reasoning, output fault diagnosis result.The present application is based on knowledge subgraph convergence and semantic reasoning method, realizes equipment fault intelligent diagnosis, with the advantages of accurate positioning, efficient interaction and result interpretable.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and in particular to a method for equipment fault diagnosis based on the uncertainty convergence of knowledge subgraphs. Background Technology

[0002] As industrial equipment becomes increasingly large-scale and complex, it generates a large amount of alarm information and fault description text during operation. Maintenance personnel typically rely on experience or keyword-based knowledge base retrieval for fault location and troubleshooting. Some systems introduce knowledge graphs to structure the relationships between fault characteristics and influencing factors, and combine natural language processing techniques to extract features from fault text, thus aiding in diagnosis. However, existing solutions are mostly based on single-retrieval or static matching, lacking uncertainty measurement and dynamic convergence mechanisms for the scope of fault root causes. In complex causal chain scenarios, this can easily lead to problems such as an excessively large or narrow range of candidate root causes.

[0003] Meanwhile, existing technologies typically fail to deeply integrate knowledge graph structure information with semantic reasoning during dialogue interaction, and cannot constrain the diagnostic process based on candidate causal subgraphs, leading to discrepancies between diagnostic results and the actual root causes. The lack of a convergence control mechanism based on changes in the number of nodes makes it difficult for the system to proactively guide users to supplement key features when information is insufficient, and also makes it difficult to promptly enter the diagnostic stage when information is sufficient, affecting the efficiency and accuracy of fault diagnosis.

[0004] Therefore, how to provide a device fault diagnosis method based on knowledge subgraph uncertainty convergence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a device fault diagnosis method based on knowledge subgraph uncertainty convergence. This invention, based on knowledge subgraph convergence and semantic reasoning, achieves intelligent device fault diagnosis and has the advantages of accurate positioning, efficient interaction, and interpretable results.

[0006] A device fault diagnosis method based on knowledge subgraph uncertainty convergence according to an embodiment of the present invention includes the following steps:

[0007] Obtain the natural language description text of the equipment fault and perform semantic preprocessing to obtain standardized fault description text;

[0008] Based on standardized fault description text, fault features are extracted and mapped to a fault knowledge graph, and a set of fault feature nodes is output.

[0009] By combining node type consistency, relationship reachability, and relationship type consistency, a secondary filtering is performed on the set of fault feature nodes to obtain the target set of fault feature nodes.

[0010] Based on the target fault feature node set, the upstream influencing factor nodes are traced back layer by layer in the fault knowledge graph, and the intersection is obtained to obtain the candidate root factor graph node set.

[0011] A candidate root factor graph is constructed based on the node set of the candidate root factor graph and the edge relationship in the fault knowledge graph, and the node set of root cause influencing factors is determined in the candidate root factor graph.

[0012] The uncertainty index is determined based on the number of nodes in the root cause influencing factor node set, and the uncertainty index is compared with the preset uncertainty threshold to obtain the dialogue control judgment result.

[0013] When the uncertainty index indicated by the dialogue control judgment result is greater than the preset uncertainty threshold, a set of targeted questions is generated, and a new set of fault features is extracted based on the response text corresponding to the targeted questions. The new set of fault features is merged with the original set of fault features to obtain an updated set of fault features. The uncertainty index is iteratively calculated until the preset uncertainty threshold is met, and a new dialogue control judgment result is output.

[0014] If the uncertainty index indicated by the dialogue control judgment result is less than or equal to the preset uncertainty threshold, the document input set is obtained based on the joint retrieval conditions and semantic reasoning is performed to output the fault diagnosis result.

[0015] Optionally, obtaining the standardized fault description text includes:

[0016] Receive user-input natural language description text of device faults and obtain device identification information bound to the natural language description text of device faults;

[0017] Perform character normalization processing on the natural language description text of the equipment fault to obtain normalized fault description text;

[0018] Sentence segmentation is performed on the standardized fault description text to obtain a sequence of sentence-level text units;

[0019] Perform word segmentation on each sentence-level text unit in the sentence-level text unit sequence to obtain a word element sequence, and generate a word order identifier for each word element in the word element sequence;

[0020] Perform entity standardization on the word sequence to obtain a standardized entity set;

[0021] Standardized fault description text is obtained by assembling sentence-level text unit sequences, word sequences, standardized entity sets, sentence order identifiers, and word order identifiers.

[0022] Optionally, the output of the fault feature node set specifically includes:

[0023] Obtain the sentence-level text unit sequence, word sequence, standardized entity set, and the association index relationship between entity words and standardized entity sets from the standardized fault description text, and construct the input data structure for fault feature extraction;

[0024] Based on the fault feature extraction input data structure, fault feature candidate recognition processing is performed on the word sequence to obtain a fault feature candidate set;

[0025] Perform feature normalization processing on the candidate fault feature set to obtain the fault feature set;

[0026] For each fault feature in the fault feature set, a feature weight value is generated. The feature weight value is calculated by weighting the location weight, type weight, and consistency weight.

[0027] For each fault feature in the fault feature set, calculate its semantic matching value with each candidate node in the fault knowledge graph. Then, select the node with the largest semantic matching value from the candidate nodes as the corresponding fault feature node and output the fault feature node set.

[0028] Optionally, the output of the target fault feature node set specifically includes:

[0029] Obtain the set of fault feature nodes and the set of nodes and directed edges in the fault knowledge graph, and generate filtering and retrieval conditions for each fault feature node in the set of fault feature nodes.

[0030] Based on the filtering and retrieval conditions, a set of candidate matching nodes is retrieved in the fault knowledge graph. For each candidate matching node in the set, the node type consistency value, relation reachability value, and relation type consistency value are calculated and then weighted and summed to obtain the node consistency value.

[0031] Candidate matching nodes whose node consistency value is less than a preset consistency threshold are removed from the candidate matching node set, and a comprehensive matching score is calculated for the remaining candidate matching nodes;

[0032] Consistency is sorted by the node set based on the overall matching score. The node with the highest overall matching score is selected as the target fault feature node, and the set of target fault feature nodes is obtained by summarizing them.

[0033] Optionally, obtaining the candidate root factor graph node set specifically includes:

[0034] Obtain the set of target fault feature nodes and the set of directed relation edges in the fault knowledge graph. Determine the direction of the edges used for backtracking as the directed relation edges from upstream influencing factor nodes to target fault feature nodes.

[0035] For each target fault feature node in the target fault feature node set, the first-level upstream influencing factor node set with a pointing relationship to the target fault feature node is retrieved based on the directed relation edge set, and a node level identifier is generated for each upstream influencing factor node in the first-level upstream influencing factor node set.

[0036] Based on the node hierarchy identifier, perform a layer-by-layer backtracking process on the current layer's upstream influencing factor node set to obtain a multi-layer upstream influencing factor node set;

[0037] For each target fault feature node, a multi-layer upstream influencing factor node set is summarized to obtain the upstream influencing factor node set corresponding to the target fault feature node. The intersection operation of the upstream influencing factor node sets corresponding to each target fault feature node is performed to obtain the candidate root factor graph node set.

[0038] Optionally, determining the set of root cause influencing factor nodes specifically includes:

[0039] Obtain the candidate root factor graph node set and the directed relation edge set in the fault knowledge graph. Based on the candidate root factor graph node set, filter the directed relation edge set to obtain the candidate root factor graph edge set.

[0040] A candidate root factor graph is constructed based on the node set and edge set of the candidate root factor graph, and the in-degree count and out-degree count of each node in the candidate root factor graph are counted.

[0041] Nodes with an in-degree count of zero and an out-degree count of greater than zero in the candidate root factor graph are identified as root cause factor nodes, and all root cause factor nodes are summarized to obtain the root cause factor node set.

[0042] Optionally, obtaining the dialogue control determination result specifically includes:

[0043] Obtain the set of root cause influencing factor nodes, and perform deduplication on the root cause influencing factor nodes in the set of root cause influencing factor nodes to obtain the deduplicated set of root cause influencing factor nodes.

[0044] The number of root cause factor nodes in the deduplicated root cause factor node set is counted to obtain the root cause factor node count value, and the root cause factor node count value is determined as an uncertainty index.

[0045] Obtain a preset uncertainty threshold, and compare the uncertainty index with the preset uncertainty threshold to obtain the dialogue control judgment result.

[0046] Optionally, the output of the new dialogue control determination result specifically includes:

[0047] Obtain the dialogue control determination result, and if the dialogue control determination result indicates that the uncertainty index is greater than the preset uncertainty threshold, obtain the root cause influencing factor node set;

[0048] For each root cause influencing factor node in the root cause influencing factor node set, generate targeted question entries and summarize them to obtain a targeted question set;

[0049] The system sends a set of targeted questions to the user terminal and receives the user's response text to the set of targeted questions. It then establishes a correspondence between the response text and the targeted question item based on the question item identifier.

[0050] Based on the response text corresponding to the question item identifier, perform fault feature extraction processing to obtain a new fault feature set, and then merge the new fault feature set with the original fault feature set to update the fault feature set.

[0051] Based on the updated fault feature set, a new root cause influencing factor node set is generated and an updated uncertainty index is calculated. The updated uncertainty index is compared with a preset uncertainty threshold to generate a new dialogue control judgment result. If the new dialogue control judgment result continues to indicate that the uncertainty index is greater than the preset uncertainty threshold, the process of generating a targeted question set, receiving response text, and updating the fault feature set is repeated until the new dialogue control judgment result indicates that the uncertainty index is less than or equal to the preset uncertainty threshold, and then a new dialogue control judgment result is output.

[0052] Optionally, the output of the fault diagnosis results specifically includes:

[0053] When the uncertainty index indicated by the dialogue control determination result is less than or equal to the preset uncertainty threshold, obtain the fault feature set and the candidate root factor graph node set.

[0054] A set of fault feature retrieval terms is generated based on the fault feature set, and a set of root cause node retrieval terms is generated based on the candidate root factor graph node set.

[0055] The set of fault feature search terms and the set of root cause node search terms are assembled to obtain the joint search conditions. The joint search conditions include fault feature matching conditions, root cause node matching conditions, and joint constraint conditions of fault features and root cause nodes.

[0056] The joint search criteria are used to perform a search in the equipment operation and maintenance knowledge base to obtain a set of target documents, and a document relevance value is generated for each target document in the set of target documents.

[0057] Sort the target document set in descending order of document relevance value, select the top-ranked target documents by a predetermined number of relevance values, and assemble them into the document input set;

[0058] The document input set and the node identifiers of the candidate root factor graph node set are input into the large language model for semantic reasoning to generate fault diagnosis conclusions and fault troubleshooting and repair solutions.

[0059] The fault diagnosis conclusions, fault investigation and repair plans are associated and encapsulated with the document input set to output the fault diagnosis results.

[0060] The beneficial effects of this invention are:

[0061] This invention constructs a causal structure representation model centered on a fault knowledge graph. It structurally models the directed relationships between equipment fault feature nodes and upstream influencing factor nodes. Based on the target fault feature node set, it backtracks layer by layer and forms a candidate root factor graph node set through intersection. Then, it determines the root cause influencing factor node set through in-degree and out-degree constraints. This transforms the fault localization process from traditional experience-based judgment to a causal reasoning process based on knowledge subgraph structural constraints. This structured backtracking and intersection convergence mechanism effectively compresses the root cause search space, avoids interference from irrelevant influencing factors, improves the accuracy and interpretability of root cause localization, and provides a clear causal chain for the diagnostic path.

[0062] Building upon this foundation, this application further introduces an uncertainty index based on the number of nodes influencing root causes, and combines this with a preset uncertainty threshold to form a dialogue control and judgment mechanism. When the range of candidate root causes is large, the system guides the user to supplement key fault features through a set of targeted questions, and re-executes node matching and causal backtracking based on the updated fault feature set, thereby achieving gradual convergence of the knowledge subgraph. This uncertainty convergence mechanism enables the system to dynamically adjust its interaction strategy according to the diagnostic status, proactively collecting supplementary information when information is insufficient, and promptly entering the retrieval and semantic reasoning stages when information is sufficient, thus improving the efficiency of human-machine collaborative diagnosis.

[0063] Simultaneously, after meeting the uncertainty threshold condition, this application structurally combines the fault feature set and the candidate root factor graph node set through joint retrieval conditions. Highly relevant documents are then selected from the equipment operation and maintenance knowledge base. The document input set and candidate root factor graph node identifiers are input into a large language model for semantic reasoning, enabling the diagnostic results to output fault diagnosis conclusions and fault troubleshooting and repair solutions under causal structure constraints. By deeply integrating the causal subgraph of the knowledge graph with the semantic reasoning process, the scope of semantic reasoning can be limited to candidate root causes, improving the consistency and reliability of diagnostic conclusions, while also enhancing the interpretability and practicality of the results. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart of a device fault diagnosis method based on knowledge subgraph uncertainty convergence proposed in this invention;

[0066] Figure 2 This is a flowchart of a new dialogue control decision generation process for a device fault diagnosis method based on knowledge subgraph uncertainty convergence proposed in this invention.

[0067] Figure 3 This is a flowchart illustrating the process of generating fault diagnosis results for a device fault diagnosis method based on knowledge subgraph uncertainty convergence proposed in this invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figures 1-3 A device fault diagnosis method based on knowledge subgraph uncertainty convergence includes the following steps:

[0070] Obtain the natural language description text of the equipment fault and perform semantic preprocessing to obtain standardized fault description text;

[0071] Based on standardized fault description text, fault features are extracted and mapped to a fault knowledge graph, and a set of fault feature nodes is output.

[0072] By combining node type consistency, relationship reachability, and relationship type consistency, a secondary filtering is performed on the set of fault feature nodes to obtain the target set of fault feature nodes.

[0073] Based on the target fault feature node set, the upstream influencing factor nodes are traced back layer by layer in the fault knowledge graph, and the intersection is obtained to obtain the candidate root factor graph node set.

[0074] A candidate root factor graph is constructed based on the node set of the candidate root factor graph and the edge relationship in the fault knowledge graph, and the node set of root cause influencing factors is determined in the candidate root factor graph.

[0075] The uncertainty index is determined based on the number of nodes in the root cause influencing factor node set, and the uncertainty index is compared with the preset uncertainty threshold to obtain the dialogue control judgment result.

[0076] When the uncertainty index indicated by the dialogue control judgment result is greater than the preset uncertainty threshold, a set of targeted questions is generated, and a new set of fault features is extracted based on the response text corresponding to the targeted questions. The new set of fault features is merged with the original set of fault features to obtain an updated set of fault features. The uncertainty index is iteratively calculated until the preset uncertainty threshold is met, and a new dialogue control judgment result is output.

[0077] If the uncertainty index indicated by the dialogue control judgment result is less than or equal to the preset uncertainty threshold, the document input set is obtained based on the joint retrieval conditions and semantic reasoning is performed to output the fault diagnosis result.

[0078] In this embodiment, obtaining the standardized fault description text includes:

[0079] Receives user-inputted natural language description text of device faults and obtains device identification information bound to the natural language description text of device faults, wherein the device identification information includes device model identifier, device component identifier, and device operating scenario identifier;

[0080] The natural language description text of equipment faults is subjected to character normalization processing to obtain normalized fault description text. Character normalization processing includes unifying full-width and half-width characters, unifying uppercase and lowercase English letters, unifying synonyms, unifying the writing format of numbers, and unifying the writing format of units of measurement.

[0081] The standardized fault description text is processed by sentence segmentation to obtain a sentence-level text unit sequence. The sentence segmentation process includes segmentation based on period, question mark, exclamation mark, semicolon, comma, and newline character, and a sentence order identifier is generated for each sentence-level text unit in the sentence-level text unit sequence.

[0082] Each sentence-level text unit in the sentence-level text unit sequence is segmented to obtain a word unit sequence, and a word order identifier is generated for each word unit in the word unit sequence. The word unit sequence is divided into fault phenomenon word units, alarm information word units, equipment component word units, operating status word units, operation action word units, and environmental condition word units according to the semantic type of word units.

[0083] Entity standardization is performed on the lexical sequence to obtain a standardized entity set. Entity standardization includes: identifying device model entities, device component entities, alarm code entities, operating parameter entities, and status description entities in the lexical sequence; mapping device model entities to standard device model names, device component entities to standard component names, alarm code entities to standard alarm code codes, operating parameter entities to standard operating parameter field identifiers, and status description entities to status semantic tags; performing unified merging processing on entities corresponding to aliases, abbreviations, synonyms, and different writing styles; performing expression normalization processing on entities corresponding to unit differences, format differences, and mixed Chinese and English usage; and structurally encapsulating the merged and normalized entities according to entity type, standard name, standard code, standard field identifier, and status semantic tag to obtain the standardized entity set.

[0084] Standardized fault description text is obtained by assembling sentence-level text unit sequences, word sequences, standardized entity sets, sentence order identifiers, and word order identifiers. The standardized fault description text uses sentence order identifiers to index the sentence-level text unit sequences and establishes an association index relationship between entity words in the word sequence and the standardized entity set.

[0085] In this embodiment, the output of the fault feature node set specifically includes:

[0086] Obtain the sentence-level text unit sequence, word sequence, standardized entity set, and the association index relationship between entity words and standardized entity sets from the standardized fault description text, and construct the input data structure for fault feature extraction;

[0087] Based on the fault feature extraction input data structure, fault feature candidate recognition processing is performed on the lexical sequence to obtain a fault feature candidate set. The fault feature candidate recognition processing includes: identifying candidate lexical combinations representing fault states based on syntactic collocation relationships in sentence-level text unit sequences, adjacent combination relationships in lexical sequences, and association index relationships between entity lexicals and standardized entity sets; extracting candidate lexical combinations containing descriptions of fault phenomena, alarm information, equipment components, operating status, operation actions, and environmental conditions as fault feature candidates; and classifying each fault feature candidate according to semantic type to obtain a fault feature candidate set composed of fault phenomenon features, alarm information features, equipment component features, operating status features, operation action features, and environmental condition features.

[0088] The fault feature candidate set is subjected to feature normalization processing to obtain the fault feature set. Feature normalization processing includes: performing synonym merging processing on candidate features with the same semantics; performing hierarchical merging processing on candidate features containing a relationship between higher-level and lower-level descriptions; performing splitting processing on composite candidate features that contain multiple semantic components; performing field completion processing on candidate features that lack equipment component information, status information, or alarm information; and generating the fault feature set based on the normalized candidate features.

[0089] For each fault feature in the fault feature set, a feature weight value is generated. The feature weight value is calculated by weighting the position weight, type weight and consistency weight. The position weight is determined according to the sentence order position of the fault feature in the sentence-level text unit sequence, the type weight is determined according to the semantic type of the fault feature, and the consistency weight is determined according to the consistency matching result between the fault feature and the equipment model standard name and component standard name in the standardized entity set.

[0090] For each fault feature in the fault feature set, calculate its semantic matching value with each candidate node in the fault knowledge graph. Then, select the node with the largest semantic matching value from the candidate nodes as the corresponding fault feature node based on the semantic matching value. Output the fault feature node set, which includes the target node identifier corresponding to each fault feature, the corresponding semantic matching value, and the association index relationship with the fault feature.

[0091] The semantic matching value is obtained by multiplying the weighted sum of the text representation matching value and the graph structure matching value by the feature weight value of the corresponding fault feature. The text representation matching value is determined based on the vector similarity between the fault feature text and the candidate node name text. The graph structure matching value is determined based on the overlap between the set of adjacent relationship types of the candidate node in the fault knowledge graph and the set of relationship types corresponding to the standardized entity set. The weighted summation method is to multiply the text representation matching value by the first matching weight, multiply the graph structure matching value by the second matching weight, and then sum the two results. The sum of the values ​​of the first matching weight and the second matching weight is one.

[0092] The fault knowledge graph is a directed graph structure data model used to represent the causal relationship between equipment fault feature nodes and influencing factor nodes. It includes a set of nodes and a set of directed relation edges, where the directed relation edges are used to represent the direction of influence of influencing factor nodes on fault feature nodes. The fault knowledge graph is constructed by performing entity extraction and relation extraction on equipment operation and maintenance manual text, historical fault case text, and expert experience text to generate knowledge triples.

[0093] This invention performs a coarse screening of candidate nodes in the fault knowledge graph through semantic matching. This allows for the retention of nodes that are semantically related to the fault description text, narrowing the scope of subsequent node comparisons and reducing the computational overhead caused by irrelevant nodes participating in subsequent processing. By adopting a coarse screening followed by fine ranking approach, it provides a candidate basis for the structural consistency verification of subsequent target fault feature nodes, thereby balancing node screening efficiency and subsequent matching accuracy.

[0094] In this embodiment, the output of the target fault feature node set specifically includes:

[0095] Obtain the set of fault feature nodes and the set of nodes and directed edges in the fault knowledge graph, and generate filtering and retrieval conditions for each fault feature node in the set of fault feature nodes.

[0096] The generation of filtering and retrieval conditions includes, for each fault feature node, extracting the node index information of the fault feature node in the fault feature node set, and binding the node index information with the fault feature node identifier, fault feature node type, and semantic matching value to form a filtering and retrieval condition item corresponding to a single fault feature node; summarizing the filtering and retrieval condition items corresponding to each fault feature node to obtain a filtering and retrieval condition set, wherein the semantic matching value is used to characterize the degree of semantic association between the corresponding fault feature node and the node to be retrieved in the fault knowledge graph, the fault feature node type is used to limit the type retrieval range of candidate matching nodes, and the fault feature node identifier is used to mark the source of the fault feature node corresponding to the current filtering and retrieval condition.

[0097] Based on the filtering and retrieval conditions, candidate matching nodes are retrieved in the fault knowledge graph. The candidate matching nodes retrieved for each filtering and retrieval condition are then aggregated according to the node index information of the fault feature nodes to form a corresponding subset in the candidate matching node set. For each candidate matching node, the node identifier, node type, associated directed relation edges, and initial semantic matching results with the current fault feature node are first obtained. Then, the node type consistency value, relation reachability value, and relation type consistency value are calculated respectively. The node type consistency value, relation reachability value, and relation type consistency value are weighted and fused to obtain the node consistency value corresponding to each candidate matching node. The node consistency value is then associated with the corresponding candidate matching node to form a candidate matching node set with consistency evaluation results.

[0098] The node type consistency value is obtained by performing a type correspondence check on the node type of the candidate matching node and the node type of the fault feature node. It is used to characterize the matching status of the candidate matching node and the fault feature node at the node category level. The type correspondence check includes reading the node type of the candidate matching node, reading the node type of the fault feature node, performing a consistency comparison on the two node types, and writing the type consistency comparison result into the node type consistency value.

[0099] The relation reachability value is obtained by performing path retrieval on the directed relation paths of candidate matching nodes in the fault knowledge graph. It is used to characterize the structural connectivity between candidate matching nodes and device identification information, standardized entity set, and other fault feature nodes in the fault knowledge graph. The path retrieval includes starting from the candidate matching node, searching along the directed relation edges in the fault knowledge graph to the corresponding node of device identification information, the corresponding node of standardized entity set, and other fault feature nodes, recording the path existence result, path direction verification result, and path connection result of each directed connected path, and summarizing the retrieval results of each directed connected path and writing them into the relation reachability value.

[0100] The relation type consistency value is obtained by performing a relation type comparison between the directed relation edge types associated with candidate matching nodes and the relation types corresponding to the standardized entity set. It is used to characterize the semantic consistency between the relation associated with candidate matching nodes and the relation corresponding to the standardized entity set. The relation type comparison includes extracting the types of each directed relation edge associated with candidate matching nodes, reading the relation types corresponding to the standardized entity set, performing an item-by-item comparison between each directed relation edge type and the corresponding relation type, and writing the relation type comparison results into the relation type consistency value.

[0101] Candidate matching nodes whose node consistency value is less than the preset consistency threshold are removed from the corresponding candidate matching node subset, and candidate matching nodes whose node consistency value is greater than or equal to the preset consistency threshold are retained as nodes that pass the consistency test. For each node that passes the consistency test, the semantic matching value and the node consistency value are weighted and summed to output the corresponding comprehensive matching score. The comprehensive matching score is used to uniformly represent the matching status of candidate matching nodes in two dimensions: semantic association result and graph structure consistency result.

[0102] Based on the node index information of the fault feature nodes, the candidate matching nodes corresponding to each consistent node are grouped and sorted. Within each group of consistent nodes corresponding to a fault feature node, they are sorted from high to low according to the overall matching score, and the first consistent node in the sorting result is determined as the target fault feature node corresponding to that fault feature node. The target fault feature nodes corresponding to each fault feature node are summarized and a set of target fault feature nodes is output. Each target fault feature node in the set of target fault feature nodes maintains an association index relationship with the original fault feature node and carries the corresponding overall matching score.

[0103] Based on the fault feature nodes obtained from coarse screening, this invention further combines node type consistency, relation reachability, and relation type consistency to perform fine ranking, which can filter out nodes with abnormal structural connectivity or inconsistent relation types from semantically related nodes, avoiding the deviation of target nodes caused by relying solely on semantic matching; the fine ranking process improves the consistency between the target fault feature nodes and the structural semantics of the fault knowledge graph, thereby improving the reliability of subsequent root factor graph construction and diagnosis results.

[0104] In this embodiment, obtaining the candidate root factor graph node set specifically includes:

[0105] Obtain the set of target fault feature nodes and the set of directed relation edges in the fault knowledge graph. Determine the direction of the edges used for backtracking as the directed relation edges from upstream influencing factor nodes to target fault feature nodes.

[0106] For each target fault feature node in the target fault feature node set, the first-level upstream influencing factor node set with a pointing relationship to the target fault feature node is retrieved based on the directed relation edge set, and a node level identifier is generated for each upstream influencing factor node in the first-level upstream influencing factor node set.

[0107] Based on the node hierarchy identifier, perform a layer-by-layer backtracking process on the current layer upstream influencing factor node set to obtain a multi-layer upstream influencing factor node set. The layer-by-layer backtracking process includes taking the current layer upstream influencing factor node as the backtracking starting node set, retrieving the next layer upstream influencing factor node set pointing to the backtracking starting node set, and generating an updated node hierarchy identifier for the next layer upstream influencing factor node set.

[0108] For each target fault feature node, a multi-layer upstream influencing factor node set is summarized to obtain the corresponding upstream influencing factor node set. The intersection operation of the upstream influencing factor node sets corresponding to each target fault feature node is performed to obtain the candidate root factor graph node set. The candidate root factor graph node set includes the candidate root factor graph node identifier and the association index relationship between the candidate root factor graph node and the target fault feature node.

[0109] In this embodiment, the determination of the root cause influencing factor node set specifically includes:

[0110] Obtain the candidate root factor graph node set and the directed relation edge set in the fault knowledge graph. Based on the candidate root factor graph node set, filter the directed relation edge set to obtain the candidate root factor graph edge set. The candidate root factor graph edge set includes directed relation edges where both the starting node and the ending node belong to the candidate root factor graph node set.

[0111] A candidate root factor graph is constructed based on the candidate root factor graph node set and the candidate root factor graph edge set. The in-degree count and out-degree count of each node in the candidate root factor graph are counted. The candidate root factor graph includes the candidate root factor graph node set and the candidate root factor graph edge set. The in-degree count is the number of directed edges pointing to the node, and the out-degree count is the number of directed edges pointing from the node to other nodes.

[0112] Nodes in the candidate root factor graph with an in-degree count of zero and an out-degree count of greater than zero are identified as root factor nodes. All root factor nodes are then aggregated to form a root factor node set, which includes the root factor node identifier, the corresponding in-degree count, the corresponding out-degree count, and the association index between the root factor node and the candidate root factor graph node set.

[0113] In this embodiment, obtaining the dialogue control determination result specifically includes:

[0114] Obtain the set of root cause influencing factor nodes, and perform deduplication on the root cause influencing factor nodes in the set of root cause influencing factor nodes to obtain the deduplicated set of root cause influencing factor nodes.

[0115] The deduplication process includes: extracting the root cause factor node identifier for each root cause factor node in the root cause factor node set; establishing a node identifier mapping table based on the root cause factor node identifiers; checking whether there are duplicate root cause factor node identifiers in the node identifier mapping table; if duplicate root cause factor node identifiers exist, retaining only one root cause factor node with the same identifier and deleting the remaining duplicate nodes with the same identifier; if no duplicate root cause factor node identifiers exist, retaining the original root cause factor node.

[0116] The number of root cause factor nodes in the deduplicated root cause factor node set is counted to obtain the root cause factor node count value, and the root cause factor node count value is determined as an uncertainty index.

[0117] A preset uncertainty threshold is obtained, and the uncertainty index is compared with the preset uncertainty threshold to obtain the dialogue control judgment result. When the uncertainty index is greater than the preset uncertainty threshold, the dialogue control judgment result is the first judgment state, and when the uncertainty index is less than or equal to the preset uncertainty threshold, the dialogue control judgment result is the second judgment state. The dialogue control judgment result includes the dialogue control identifier, the corresponding uncertainty index value, the corresponding preset uncertainty threshold, and the association index relationship with the root cause influencing factor node set.

[0118] In this embodiment, the output of the new dialogue control determination result specifically includes:

[0119] Obtain the dialogue control determination result, and if the dialogue control determination result indicates that the uncertainty index is greater than the preset uncertainty threshold, obtain the root cause influencing factor node set;

[0120] For each root cause influencing factor node in the root cause influencing factor node set, a targeted question item is generated and aggregated to obtain a targeted question set. The generation of targeted question items includes: based on each root cause influencing factor node in the root cause influencing factor node set, the corresponding leaf observation node is retrieved in reverse along the directed relation edges in the fault knowledge graph, and the union of the retrieved leaf observation nodes is taken; each leaf observation node is used as a supplementary observation condition in turn, and the decrease in the uncertainty index corresponding to each leaf observation node is evaluated; the leaf observation node with the largest decrease in uncertainty index is selected as the priority question node; a query statement is constructed around the observation content corresponding to the priority question node, and the constructed query statement is determined as the targeted question item.

[0121] The system sends a set of targeted questions to the user terminal and receives the user's response text to the set of targeted questions. It then establishes a correspondence between the response text and the targeted question item based on the question item identifier.

[0122] Based on the response text corresponding to the question item identifier, perform fault feature extraction processing to obtain a new fault feature set, and then merge the new fault feature set with the original fault feature set to update the fault feature set.

[0123] Based on the updated fault feature set, a new root cause influencing factor node set is generated and an updated uncertainty index is calculated. The updated uncertainty index is compared with a preset uncertainty threshold to generate a new dialogue control judgment result. If the new dialogue control judgment result continues to indicate that the uncertainty index is greater than the preset uncertainty threshold, the process of generating a targeted question set, receiving response text, and updating the fault feature set is repeated until the new dialogue control judgment result indicates that the uncertainty index is less than or equal to the preset uncertainty threshold, and then a new dialogue control judgment result is output.

[0124] In this embodiment, the output of the fault diagnosis results specifically includes:

[0125] When the uncertainty index indicated by the dialogue control determination result is less than or equal to the preset uncertainty threshold, obtain the fault feature set and the candidate root factor graph node set.

[0126] A set of fault feature retrieval items is generated based on the fault feature set, and a set of root cause node retrieval items is generated based on the candidate root factor graph node set. The fault feature retrieval item set includes fault phenomenon retrieval items, alarm information retrieval items, equipment component retrieval items, operating status retrieval items, operation action retrieval items, and environmental condition retrieval items. The root cause node retrieval item set includes candidate root factor graph node identifier retrieval items, candidate root factor graph node name retrieval items, and candidate root factor graph node association relationship retrieval items.

[0127] The joint search conditions are obtained by assembling the set of fault feature search terms and the set of root cause node search terms. The joint search conditions include fault feature matching conditions, root cause node matching conditions, and joint constraint conditions of fault features and root cause nodes.

[0128] A search is performed in the equipment operation and maintenance knowledge base using joint search criteria to obtain a set of target documents. A document relevance value is generated for each target document in the set. The document relevance value is obtained by weighted summation of fault feature matching score and root cause node matching score. The fault feature matching score is determined by the ratio of the number of fault feature search terms contained in the target document to the number of fault features in the fault feature set. The root cause node matching score is determined by the ratio of the number of root cause node search terms contained in the target document to the number of root cause node search terms in the root cause node search term set.

[0129] Sort the target document set in descending order of document relevance value, select the top-ranked target documents by a predetermined number of relevance values, and assemble them into the document input set;

[0130] The document input set and the node identifiers of the candidate root factor graph node set are input into the large language model for semantic reasoning to generate fault diagnosis conclusions and fault investigation and repair solutions.

[0131] The large language model is an operational domain-adaptive generative model built on a pre-trained language representation model. The large language model is used to perform contextual semantic encoding on the document input set to obtain the document semantic representation, and then fuse the document semantic representation with the root cause node semantic representation corresponding to the candidate root factor graph node set to form a joint semantic representation. Under the constraint of the candidate root factor graph node set, the model outputs fault diagnosis conclusions and fault troubleshooting and repair solutions based on the joint semantic representation. The node identifiers of the candidate root factor graph node set serve as conditional prompts in the semantic encoding and semantic fusion process to restrict the semantic reasoning content to revolve around the candidate root factor graph nodes.

[0132] The fault diagnosis conclusion, fault investigation and repair plan and document input set are associated and encapsulated to output the fault diagnosis result. The fault diagnosis conclusion includes the fault type identifier and fault cause description, and the fault investigation and repair plan includes the fault investigation step set and the repair operation set.

[0133] Example 1: To verify the feasibility of this invention in practice, it was applied to an equipment operation and maintenance (O&M) scenario at a large substation. This substation is responsible for the core power supply of the region, and the equipment on site has complex models and varying service lives. Historical O&M data is stored in a scattered text record format in an equipment O&M knowledge base. During a night inspection, the on-duty personnel discovered an abnormally high temperature in the main transformer, accompanied by intermittent alarms. However, the alarm descriptions were vague, making it impossible to directly pinpoint the cause of the fault. Previous methods relied on manually consulting the O&M manual and historical case studies, which was not only time-consuming but also prone to errors due to differing understandings of the fault symptoms among different personnel, leading to repeated adjustments in the troubleshooting direction and delays in response.

[0134] In this scenario, on-duty personnel input a natural language description of the equipment fault via the maintenance terminal. The system first performs semantic preprocessing on the text to form a standardized fault description text, and automatically extracts fault features and maps them to a fault knowledge graph, obtaining a set of fault feature nodes. Subsequently, the system performs a secondary screening of the fault feature node set based on node type consistency, relationship reachability, and relationship type consistency to obtain a target fault feature node set. Based on the target fault feature node set, the system traces back upstream influencing factor nodes layer by layer in the fault knowledge graph to construct a candidate root factor graph and determine the root cause influencing factor node set. If the number of root cause influencing factor nodes is large, a dialogue control mechanism generates a set of targeted questions to guide the on-duty personnel to supplement key operating status and operation information. The supplemented information re-enters the fault feature extraction and root factor graph update process, and the uncertainty index converges accordingly. When the uncertainty index drops below a preset uncertainty threshold, the system retrieves relevant documents in the equipment maintenance knowledge base based on joint retrieval conditions, forming a document input set. It then performs semantic reasoning based on the node identifiers of the candidate root factor graph node set, outputting a fault diagnosis conclusion and a fault troubleshooting and repair plan.

[0135] In actual operation, this method is deployed on the server in the substation's control center, covering the high-load power supply phase throughout the year. Compared to traditional manual troubleshooting processes, this invention significantly shortens the overall response time from fault description input to fault diagnosis results in multiple nighttime alarm scenarios, reducing redundant document reviews and ineffective troubleshooting steps, allowing maintenance personnel to handle more alarm events within the same time window. The system maintains stable convergence performance across multiple different equipment types, with uncertainty indicators rapidly decreasing after dialogue supplementation, a significant reduction in the size of the candidate root factor graph, and a significant improvement in the relevance of retrieved documents. Therefore, this invention effectively solves the problems of fault location relying on experience, divergent troubleshooting paths, and inconsistent diagnostic conclusions in existing technologies, achieving structured convergence and knowledge-driven reasoning in the fault diagnosis process, thus improving maintenance efficiency and diagnostic accuracy.

[0136] Table 1. Performance Comparison of the Invention and Traditional Equipment Fault Diagnosis Methods

[0137] Indicator Categories Traditional methods Method of the present invention Average diagnosis time (min) 42.6 33.8 Average number of documents retrieved (articles) 18 13 Percentage of valid documents (%) 63.4 74.9 Diagnostic accuracy (%) 85.7 91.6 Number of times a second supplementary question and answer session was held. 3.2 2.5 Uncertainty indicator convergence rounds (rounds) 3.5 2.4 Percentage of human intervention (%) 28.4 21.9

[0138] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0139] In terms of average diagnostic time, the traditional method takes 42.6 minutes, while the method of this invention takes 33.8 minutes. This improvement stems from the fact that this invention uses targeted questioning driven by uncertainty indicators, enabling the fault feature set to converge effectively in the early stages, avoiding blind manual searching in a large knowledge base, and thus reducing ineffective troubleshooting time.

[0140] Regarding the average number of documents retrieved, the traditional method retrieves 18 documents, while the method of this invention retrieves 13. The decrease in the search size did not reduce diagnostic accuracy; on the contrary, it increased the proportion of valid documents from 63.4% to 74.9%. This is because the joint search conditions simultaneously constrain both the fault feature set and the candidate root factor graph node set, allowing the search scope to be guided by the structured root factor graph, resulting in more accurate calculation of document relevance values ​​and improving document matching quality from the source.

[0141] The diagnostic accuracy improved from 85.7% to 91.6%. This improvement is closely related to the participation of the candidate root factor graph node set in semantic reasoning. The large language model integrates the semantic representation of root factor nodes during the semantic encoding stage, and the reasoning process is constrained by the candidate root factor graph node set, reducing the diffusion of reasoning unrelated to the current equipment failure, thereby improving the consistency of conclusions.

[0142] In terms of interaction efficiency, the number of secondary supplementary questions and answers decreased from 3.2 to 2.5, and the number of convergence rounds for the uncertainty index decreased from 3.5 to 2.4. This improvement reflects that the size of the root cause influencing factor node set shrinks more quickly under the action of the dialogue control mechanism, indicating that the uncertainty index can effectively quantify and diagnose the degree of divergence and converge quickly through targeted questioning.

[0143] The proportion of manual intervention decreased from 28.4% to 21.9%. This change indicates that the fault diagnosis results output by the system are more stable in terms of structural integrity and logical consistency, and maintenance personnel no longer need to frequently manually correct the troubleshooting path. Overall, the data shows that this invention achieves synergistic optimization of diagnostic efficiency, accuracy, and interaction convergence while ensuring a reasonable improvement.

[0144] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for diagnosing equipment faults based on the uncertainty convergence of knowledge subgraphs, characterized in that, Includes the following steps: Obtain the natural language description text of the equipment fault and perform semantic preprocessing to obtain standardized fault description text; Based on standardized fault description text, fault features are extracted and mapped to a fault knowledge graph, and a set of fault feature nodes is output. By combining node type consistency, relationship reachability, and relationship type consistency, a secondary filtering is performed on the set of fault feature nodes to obtain the target set of fault feature nodes. Based on the target fault feature node set, the upstream influencing factor nodes are traced back layer by layer in the fault knowledge graph, and the intersection is obtained to obtain the candidate root factor graph node set. A candidate root factor graph is constructed based on the node set of the candidate root factor graph and the edge relationship in the fault knowledge graph, and the node set of root cause influencing factors is determined in the candidate root factor graph. The uncertainty index is determined based on the number of nodes in the root cause influencing factor node set, and the uncertainty index is compared with the preset uncertainty threshold to obtain the dialogue control judgment result. When the uncertainty index indicated by the dialogue control judgment result is greater than the preset uncertainty threshold, a set of targeted questions is generated, and a new set of fault features is extracted based on the response text corresponding to the targeted questions. The new set of fault features is merged with the original set of fault features to obtain an updated set of fault features. The uncertainty index is iteratively calculated until the preset uncertainty threshold is met, and a new dialogue control judgment result is output. If the uncertainty index indicated by the dialogue control judgment result is less than or equal to the preset uncertainty threshold, the document input set is obtained based on the joint retrieval conditions and semantic reasoning is performed to output the fault diagnosis result.

2. The equipment fault diagnosis method based on knowledge subgraph uncertainty convergence according to claim 1, characterized in that, The standardized fault description text is obtained by: Receive user-input natural language description text of device faults and obtain device identification information bound to the natural language description text of device faults; Perform character normalization processing on the natural language description text of the equipment fault to obtain normalized fault description text; Sentence segmentation is performed on the standardized fault description text to obtain a sequence of sentence-level text units; Perform word segmentation on each sentence-level text unit in the sentence-level text unit sequence to obtain a word element sequence, and generate a word order identifier for each word element in the word element sequence; Perform entity standardization on the word sequence to obtain a standardized entity set; Standardized fault description text is obtained by assembling sentence-level text unit sequences, word sequences, standardized entity sets, sentence order identifiers, and word order identifiers.

3. The equipment fault diagnosis method based on knowledge subgraph uncertainty convergence according to claim 1, characterized in that, The output of the fault feature node set specifically includes: Obtain the sentence-level text unit sequence, word sequence, standardized entity set, and the association index relationship between entity words and standardized entity sets from the standardized fault description text, and construct the input data structure for fault feature extraction; Based on the fault feature extraction input data structure, fault feature candidate recognition processing is performed on the word sequence to obtain a fault feature candidate set; Perform feature normalization processing on the candidate fault feature set to obtain the fault feature set; For each fault feature in the fault feature set, a feature weight value is generated. The feature weight value is calculated by weighting the location weight, type weight, and consistency weight. For each fault feature in the fault feature set, calculate its semantic matching value with each candidate node in the fault knowledge graph. Then, select the node with the largest semantic matching value from the candidate nodes as the corresponding fault feature node and output the fault feature node set.

4. The equipment fault diagnosis method based on knowledge subgraph uncertainty convergence according to claim 1, characterized in that, The output of the target fault feature node set specifically includes: Obtain the set of fault feature nodes and the set of nodes and directed edges in the fault knowledge graph, and generate filtering and retrieval conditions for each fault feature node in the set of fault feature nodes. Based on the filtering and retrieval conditions, a set of candidate matching nodes is retrieved in the fault knowledge graph. For each candidate matching node in the set, the node type consistency value, relation reachability value, and relation type consistency value are calculated and then weighted and summed to obtain the node consistency value. Candidate matching nodes whose node consistency value is less than a preset consistency threshold are removed from the candidate matching node set, and a comprehensive matching score is calculated for the remaining candidate matching nodes; Consistency is sorted by the node set based on the overall matching score. The node with the highest overall matching score is selected as the target fault feature node, and the set of target fault feature nodes is obtained by summarizing them.

5. The equipment fault diagnosis method based on knowledge subgraph uncertainty convergence according to claim 1, characterized in that, The specific steps involved in obtaining the candidate root factor graph node set are as follows: Obtain the set of target fault feature nodes and the set of directed relation edges in the fault knowledge graph. Determine the direction of the edges used for backtracking as the directed relation edges from upstream influencing factor nodes to target fault feature nodes. For each target fault feature node in the target fault feature node set, the first-level upstream influencing factor node set with a pointing relationship to the target fault feature node is retrieved based on the directed relation edge set, and a node level identifier is generated for each upstream influencing factor node in the first-level upstream influencing factor node set. Based on the node hierarchy identifier, perform a layer-by-layer backtracking process on the current layer's upstream influencing factor node set to obtain a multi-layer upstream influencing factor node set; For each target fault feature node, a multi-layer upstream influencing factor node set is summarized to obtain the upstream influencing factor node set corresponding to the target fault feature node. The intersection operation of the upstream influencing factor node sets corresponding to each target fault feature node is performed to obtain the candidate root factor graph node set.

6. The equipment fault diagnosis method based on knowledge subgraph uncertainty convergence according to claim 1, characterized in that, The determination of the root cause influencing factor node set specifically includes: Obtain the candidate root factor graph node set and the directed relation edge set in the fault knowledge graph. Based on the candidate root factor graph node set, filter the directed relation edge set to obtain the candidate root factor graph edge set. A candidate root factor graph is constructed based on the node set and edge set of the candidate root factor graph, and the in-degree count and out-degree count of each node in the candidate root factor graph are counted. Nodes with an in-degree count of zero and an out-degree count of greater than zero in the candidate root factor graph are identified as root cause factor nodes, and all root cause factor nodes are summarized to obtain the root cause factor node set.

7. The equipment fault diagnosis method based on knowledge subgraph uncertainty convergence according to claim 1, characterized in that, The specific steps to obtain the dialogue control determination result include: Obtain the set of root cause influencing factor nodes, and perform deduplication on the root cause influencing factor nodes in the set of root cause influencing factor nodes to obtain the deduplicated set of root cause influencing factor nodes. The number of root cause factor nodes in the deduplicated root cause factor node set is counted to obtain the root cause factor node count value, and the root cause factor node count value is determined as an uncertainty index. Obtain a preset uncertainty threshold, and compare the uncertainty index with the preset uncertainty threshold to obtain the dialogue control judgment result.

8. The equipment fault diagnosis method based on knowledge subgraph uncertainty convergence according to claim 1, characterized in that, The output of the new dialogue control determination result specifically includes: Obtain the dialogue control determination result, and if the dialogue control determination result indicates that the uncertainty index is greater than the preset uncertainty threshold, obtain the root cause influencing factor node set; For each root cause influencing factor node in the root cause influencing factor node set, generate targeted question entries and summarize them to obtain a targeted question set; The system sends a set of targeted questions to the user terminal and receives the user's response text to the set of targeted questions. It then establishes a correspondence between the response text and the targeted question item based on the question item identifier. Based on the response text corresponding to the question item identifier, perform fault feature extraction processing to obtain a new fault feature set, and then merge the new fault feature set with the original fault feature set to update the fault feature set. Based on the updated fault feature set, a new root cause influencing factor node set is generated and an updated uncertainty index is calculated. The updated uncertainty index is compared with a preset uncertainty threshold to generate a new dialogue control judgment result. If the new dialogue control judgment result continues to indicate that the uncertainty index is greater than the preset uncertainty threshold, the process of generating a targeted question set, receiving response text, and updating the fault feature set is repeated until the new dialogue control judgment result indicates that the uncertainty index is less than or equal to the preset uncertainty threshold, and then a new dialogue control judgment result is output.

9. The equipment fault diagnosis method based on knowledge subgraph uncertainty convergence according to claim 1, characterized in that, The output of the fault diagnosis results specifically includes: When the uncertainty index indicated by the dialogue control determination result is less than or equal to the preset uncertainty threshold, obtain the fault feature set and the candidate root factor graph node set. A set of fault feature retrieval terms is generated based on the fault feature set, and a set of root cause node retrieval terms is generated based on the candidate root factor graph node set. The set of fault feature search terms and the set of root cause node search terms are assembled to obtain the joint search conditions. The joint search conditions include fault feature matching conditions, root cause node matching conditions, and joint constraint conditions of fault features and root cause nodes. The joint search criteria are used to perform a search in the equipment operation and maintenance knowledge base to obtain a set of target documents, and a document relevance value is generated for each target document in the set of target documents. Sort the target document set in descending order of document relevance value, select the top-ranked target documents by a predetermined number of relevance values, and assemble them into the document input set; The document input set and the node identifiers of the candidate root factor graph node set are input into the large language model for semantic reasoning to generate fault diagnosis conclusions and fault troubleshooting and repair solutions.

10. Associate and encapsulate the fault diagnosis conclusions, fault troubleshooting and repair plans with the document input set, and output the fault diagnosis results.

Images