A Substation Fault Diagnosis Method Based on Hyperbolic Procedural Reasoning

Abstract

This invention belongs to the field of text reasoning technology and provides a substation fault diagnosis method based on hyperbolic procedure reasoning. The method includes: obtaining a matching score between structured symptom text and procedure clause text based on hyperbolic distance; filtering procedure clause texts that meet preset conditions based on the matching score to form a candidate clause set; if any procedure clause in the candidate clause set has a directed edge with a fault node in a pre-constructed causal clause network, outputting the corresponding fault node to form a candidate fault set; filtering the candidate clause set using a proof-of-contrast masking mechanism to obtain a valid clause set; obtaining a comprehensive causal score for each candidate fault in the candidate fault set; and selecting the candidate fault with the highest comprehensive causal score as the final diagnosed fault type. This invention effectively eliminates interfering procedure clauses and enhances the reliability of the diagnostic results by constructing a causal clause network and introducing a proof-of-contrast masking mechanism for causal consistency verification.

Description

Technical Field

[0001] This invention relates to the field of text reasoning technology, specifically to a substation fault diagnosis method based on hyperbolic reasoning. Background Technology

[0002] With the continuous expansion of power system scale and the increasing complexity of its operation, the monitoring of substation equipment operation status and fault diagnosis are of great significance to ensuring power grid security. In actual operation and maintenance, inspection personnel need to analyze and judge equipment anomalies based on on-site observation information and in accordance with power regulations. However, traditional diagnostic methods mainly rely on manual experience and rule matching, which is not only inefficient but also prone to judgment errors under complex operating conditions, making it difficult to meet the development needs of intelligent inspection.

[0003] Currently, wearable devices such as smart glasses are increasingly used in substation inspections, making on-site data collection more convenient and real-time. However, in practical applications, multimodal data often exhibits characteristics such as high unstructuredness, non-standard semantic expression, and strong noise interference, increasing the difficulty of subsequent analysis and processing. Meanwhile, existing fault diagnosis methods are mostly based on shallow semantic matching or empirical rules, making it difficult to fully characterize the hierarchical relationships between regulations and clauses, and also difficult to accurately reflect the deep semantic connections between on-site symptoms and potential faults, resulting in room for improvement in the accuracy and stability of diagnostic results. On the other hand, power regulations typically have a clear hierarchical structure, with subordinate and related relationships between different clauses. Traditional semantic modeling methods based on Euclidean space are unable to effectively characterize this hierarchical feature, leading to limited matching accuracy between symptom information and regulations. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a substation fault diagnosis method based on hyperbolic reasoning, which improves the accuracy of fault identification and decision support capabilities, and enhances the accuracy and interpretability of inspection and diagnosis.

[0005] This invention provides a substation fault diagnosis method based on hyperbolic reasoning, comprising: S1. Collect multimodal data from the substation inspection site and convert it into text data. After preprocessing the text data, obtain structured symptom text. S2. Semantically encode the structured symptom text and the pre-normalized procedure clause text respectively, and map them to hyperbolic space to calculate hyperbolic distance. Based on the hyperbolic distance, obtain the matching score between the structured symptom text and the procedure clause text. S3. Based on the matching score, filter out the procedural clause texts that meet the preset conditions to form a candidate clause set; S4. If any procedural clause in the candidate clause set has a directed edge with a fault node in the pre-constructed causal clause network, then output the corresponding fault node to form a candidate fault set. S5. Filter the candidate clause set through the proof-of-contrast masking mechanism to obtain the valid clause set; S6. For each candidate fault in the candidate fault set, a comprehensive causal score is obtained by associating it with the procedural clauses in the valid clause set. S7. The candidate fault with the highest comprehensive causal score is taken as the final fault type for diagnosis.

[0006] Optionally, matching score The calculation is performed in the following manner: , in, Represents hyperbolic distance. It is a hyperbolic distance function. To prevent tiny constants with a denominator of zero; This represents the structured symptom text in hyperbolic space. For structured symptom text; For the first The representation of the provisions of the regulations in hyperbolic space. For the first The text of the pre-standardized procedure clauses; This is a hyperbolic space embedding mapping function.

[0007] Optionally, the method for constructing a network of causal clauses includes: Build a node set ;in, This is a set of symptom nodes used to represent device objects, abnormal parts, abnormal phenomena, alarm information, operating parameter information, and historical association information in structured symptom text; This is a set of procedural clause nodes used to represent the pre-normalized procedural clause text; This is a set of faulty nodes used to represent candidate fault types; Construct edge set ;in, These are directed edges from symptom nodes to procedure clause nodes, used to represent the triggering relationship between symptom elements and procedure clauses. A directed edge from a procedure clause node to a fault node, used to represent the causal relationship between the procedure clause and the fault type; These are directed edges from symptom nodes to fault nodes, used to represent the direct correlation between symptoms and faults in historical inspection records, historical work orders, or maintenance and reconsideration results; Construct the edge weight set ;in, The trigger edge weight between the symptom node and the procedure clause node; The causal edge weight between the node of the procedure clause and the node of failure; The historical association edge weights between symptom nodes and faulty nodes; Forming a network of causal clauses .

[0008] Optionally, the method for constructing the edge weight set includes: For the edge set Any directed edge , Boundary rights ;in, For directed edges The right of the border; Nodes in historical samples and nodes The normalized value of co-occurrence frequency; For nodes and nodes Semantic matching score between them; For the confirmation value corresponding to manual annotation, prior verification of procedures, or maintenance and re-verification results; Let be the weight coefficient, and satisfy... ; Based on directed edges The type will determine the edge weight The set of trigger edge weights that are respectively assigned to symptom nodes and procedure clause nodes The set of causal edge weights from the procedure clause node to the fault node The set of historical associated edge weights from symptom nodes to faulty nodes .

[0009] Alternatively, the set of valid terms is as follows: ; in, For the first The set of candidate clauses corresponding to each inspection. For the first Standardized procedure clause text; For the first Structured symptom text corresponding to the second inspection With the Candidate Procedure Clauses The disproven mask values ​​between them are determined based on the conflict resolution result. Sure; , in, For identifying device object conflicts; Indicates conflict in abnormal areas; This is an indicator of anomalies and conflicts. This serves as an indicator of a conflict between the alarm threshold and the status direction. To negate the conflict indicator; For identifying time-related conflicts; symbol This represents a logical OR operation; if any conflict flag is true, then... ,otherwise ; .

[0010] Optionally, the comprehensive causal score of the candidate fault is as follows: , in, For the first Candidate faults during the second inspection The comprehensive causal score; For the first The credibility score of the structured symptom text corresponding to the second inspection; This is the set of valid clauses after being eliminated by evidence of contradiction. For structured symptom text With standardized procedures and clauses The mask values ​​for the proof of contradiction between them; For structured symptom text With standardized procedures and clauses Semantic matching score between them; Symptom nodes To the nodes of the procedural clauses The trigger edge weight; For the nodes of the procedural clauses To candidate fault node Causal boundary weight; Symptom nodes To candidate fault node Historically related border rights; For the first The set of symptom nodes contained in the structured symptom text corresponding to the next inspection; This is the adjustment coefficient for the historically related edge weights.

[0011] Optionally, the first The credibility score of the structured symptom text corresponding to the second inspection. , This represents the credibility of the data source, used to characterize the confidence level of recognition from sources such as speech recognition, text recognition, and historical work order parsing. This refers to the completeness of fields, which characterizes the degree of completeness of fields such as device object, abnormal location, abnormal phenomenon, alarm information, and historical association information. To ensure consistency across multiple sources, this is used to characterize whether the verbal statements of inspection personnel, on-site alarm texts, and historical work orders corroborate each other; Time validity is used to characterize the degree of correlation between alarm information, historical work orders, and the current inspection time; Let be the weight coefficient, and satisfy... .

[0012] Optionally, the final diagnosis result is , in, The final diagnosed fault type corresponds to the r-th inspection. For candidate fault set The first in One candidate fault, For the first Candidate faults during the second inspection The comprehensive causal score.

[0013] Optionally, it also includes generating diagnostic logs, specifically:

[0014] in, For the first Diagnostic log of the next inspection task; For the first The structured symptom text corresponding to the next inspection; For a set of candidate terms; This is the set of valid clauses after being eliminated by evidence of contradiction. For the set of candidate faults; This is the mask value for the proof of contradiction; To ultimately diagnose the type of fault; For the basis of regulations; To demonstrate against the evidence; Recommendations for handling this matter.

[0015] By adopting the above technical solution, this application has the following beneficial effects: (1) This invention achieves standardized integration of inspection voice, image and operation parameters by uniformly textualizing and structuring multimodal data, thereby improving the consistency and integrity of information.

[0016] (2) This invention effectively characterizes the hierarchical structure of procedure clauses through hyperbolic space semantic mapping and matching, thereby improving the accuracy of matching between symptoms and procedures.

[0017] (3) The present invention introduces a causal clause network and performs causal consistency verification on candidate clauses through a counter-evidence masking mechanism, effectively eliminating interference information that does not conform to actual working conditions and improving diagnostic reliability.

[0018] (4) This invention integrates semantic matching score, causal weight and symptom credibility to refine the screening and sorting of fault diagnosis results, thereby improving the accuracy and interpretability of the diagnosis conclusions.

[0019] (5) This invention relies on the closed-loop feedback of diagnostic results and the iterative optimization mechanism of the model to achieve continuous improvement of system performance and enhance the robustness and engineering applicability of the method. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0021] Figure 1 The following is a flowchart of one of the substation fault diagnosis methods based on hyperbolic procedural reasoning provided by an embodiment of the present invention; Figure 2 The second flowchart of a substation fault diagnosis method based on hyperbolic procedural reasoning provided by an embodiment of the present invention is shown; Figure 3 A schematic diagram of hyperbolic space mapping in S2 provided by an embodiment of the present invention is shown; Figure 4 A schematic diagram of a causal clause network provided in an embodiment of the present invention is shown. Detailed Implementation

[0022] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention. In one embodiment, such as Figure 1-2 As shown, a substation fault diagnosis method based on hyperbolic reasoning is provided, including: S1. Collect multimodal data from the substation inspection site and convert it into text data. After preprocessing the text data, obtain structured symptom text.

[0023] Inspection personnel carry data acquisition equipment into the substation site to carry out inspection work. After entering the equipment area or actively triggering the inspection task, the multimodal data collected includes: the inspection personnel's verbal description of the status of the equipment on site; on-site image information such as equipment display interface, alarm panel, indicator light status, nameplate and identification plate; equipment operating parameters or background alarm parameters; and historical work orders, maintenance records and historical alarm records retrieved according to the current equipment identification.

[0024] No. The multi-source data collected during the second inspection is represented as follows: , in, Voice data dictated by inspection personnel; This includes images of on-site equipment, alarm panel images, or on-site text data. These are equipment operating parameters or background alarm parameters; This refers to historical work orders, maintenance records, or historical alarm data associated with the current equipment.

[0025] To integrate the aforementioned heterogeneous data from multiple sources into a unified semantic processing chain, the system converts data from different sources into text data using speech recognition functions, text recognition functions, runtime parameter parsing functions, and historical work order parsing functions, respectively. , in, This is a speech recognition function used to convert the spoken words of inspection personnel into spoken text; This is a text recognition function used to convert on-site images, alarm panels, equipment labels, or screen display content into on-site text. This is a parameter parsing function used to convert device operating parameters, alarm codes, threshold status, or background monitoring data into parameter description text. This is a function for parsing historical work orders, used to extract historical text information related to the current equipment from historical work orders, maintenance records, and historical alarm records.

[0026] After completing various text data conversions, the system performs time alignment, equipment identification alignment, and contextual association processing on spoken text, on-site text, operating parameter text, and historical related text, enabling multi-source text from the same equipment and within the same inspection period to form a unified text input. , in, This is a text fusion function. For the first The unified text input generated during each inspection.

[0027] When a certain type of data cannot be reliably acquired due to on-site environment, equipment obstruction, voice noise, or system communication abnormalities, the system can set the corresponding data item to a null value or a low confidence value, and continue to form a unified text input based on other acquired text data, thereby ensuring the continuity of the inspection and diagnosis process.

[0028] Through the above steps, the originally discrete, heterogeneous, and inconsistently expressed multimodal information in on-site inspections is transformed into a unified text input that can be extracted, standardized, matched, and reasoned about.

[0029] The collected raw text content is preprocessed to form a unified, structured symptom text. This is done after obtaining the unified text input. Afterwards, the system first performs text cleaning. The cleaning process includes removing invalid interjections, repetitive descriptions, environmental noise words, irrelevant background information, and descriptive content unrelated to the current device; correcting obvious errors generated during speech recognition and text recognition; merging repeated descriptions of the same device and the same abnormal phenomenon; and normalizing synonyms for colloquial expressions used in inspections, alarm text expressions, and historical work order expressions.

[0030] The cleaned and normalized text is further processed through word segmentation, entity recognition, attribute extraction, alarm element extraction, parameter status recognition, and time-related information extraction to extract key symptom elements highly correlated with fault identification. These key symptom elements include at least the equipment object, abnormal location, abnormal phenomenon, alarm information, operating parameter information, historical correlation information, and negative information.

[0031] In this step, the generation of structured symptom text references the pre-normalized procedure clause text field system mentioned in S2, ensuring comparability between the on-site symptom description and the procedure clause text in key fields such as equipment object, abnormal location, abnormal phenomenon, alarm information, threshold status, and historical correlation information. This processing does not mechanically rewrite the on-site symptom text into procedure clause text, nor does it alter the original on-site symptom facts. Instead, it performs field-based, standardization, and semantic normalization processing on non-standard, colloquial, or complex-sourced on-site text, thereby reducing expression differences between different data sources.

[0032] Structured symptom text can be generated by a structured transformation function: , in, The structured symptom text corresponding to the r-th inspection; A structuring transformation function for symptom text; For the first The unified text input generated during the second inspection; This is a standardized set of procedural clauses. (By introducing...) As a field reference, it enables the structured symptom text to conform as closely as possible to the standardized field system of the procedure clause text, which facilitates subsequent hyperbolic space semantic encoding and clause matching.

[0033] Structured symptom text can be further represented as: , in, For equipment objects, used to identify the name, number, or type of equipment corresponding to the current inspection; This refers to the abnormal location, used to identify the specific component or area where the abnormality occurred. An abnormal phenomenon, used to describe an abnormal state that can be observed or identified on site; This refers to alarm information or alarm threshold status, used to indicate background alarms, content displayed on the on-site alarm panel, or threshold exceeding status. This refers to equipment operating parameter information, used to indicate the status of parameters such as temperature, current, voltage, pressure, and oil level; Historical information is used to represent information related to the current symptoms in historical work orders, past maintenance records, or historical alarms. As negative information, key information such as "no abnormalities observed", "no unusual noises", "normal oil temperature", and "pressure not exceeding limits" that can support subsequent counter-evidence is retained; This is time information, used to help determine whether there is continuity or correlation between the current anomaly and historical anomalies.

[0034] S2. Semantically encode the structured symptom text and the pre-normalized procedure clause text respectively, and map them to hyperbolic space to calculate the hyperbolic distance. Based on the hyperbolic distance, obtain the matching score between the structured symptom text and the procedure clause text.

[0035] It should be noted that the standardized clause text in this step is obtained by loading the procedure clause library and pre-normalizing the original procedure clause text in the procedure clause library. The normalization process includes clause number parsing, hierarchical path extraction, equipment object identification, abnormal part identification, abnormal phenomenon extraction, trigger condition extraction, alarm information or threshold information extraction, associated fault type extraction, handling action extraction, and synonym terminology normalization, thereby forming a standardized procedure clause text with clear fields and a clear hierarchy.

[0036] Let the original set of regulations be: , in, For the first The original text of the procedure clauses, This represents the total number of clauses in the regulations.

[0037] The original procedure clauses are processed using the procedure clause normalization function: , in, For the normalization function of the procedure clauses, For the first The standardized text of the regulations and provisions is formed after the regulations and provisions have been standardized.

[0038] The standardized set of procedural clauses is represented as follows: , in, This is a collection of standardized procedures and clauses.

[0039] Each standardized procedure clause can be represented as: , in, For the clause number; The hierarchical path of the clauses within the procedural system; For device objects; This is an abnormal area; This is an abnormal phenomenon; As a triggering condition; This is either alarm information or threshold information; To associate fault types; Recommendations for handling this matter.

[0040] Through the above-mentioned standardization process, the original natural language clauses are converted into standardized procedure clause texts that can be aligned with and semantically matched with the structured symptom text from the field. Simultaneously, the system establishes a hierarchical relationship between procedure clauses based on clause number, directory level, equipment type, anomaly category, and handling logic, providing a hierarchical structural foundation for subsequent hyperbolic space mapping.

[0041] After obtaining the structured symptom text in step S1, the next step is to determine which procedural clauses have a strong semantic association with the symptom. Considering that procedural clauses typically have a clear hierarchical structure, this invention maps both the symptom text and the procedural clauses to a hyperbolic space to better characterize their hierarchical semantic features. For example... Figure 2 As shown, the structured symptom text and procedural clauses are mapped to vector representations in hyperbolic space using hyperbolic embedding functions, respectively: ; ; in, For the first The representation of the structured symptom text corresponding to the second inspection in hyperbolic space; For the first The representation of the text of a standardized procedure clause in hyperbolic space; Embedding mapping functions in hyperbolic space; For the first Standardized procedure clause text.

[0042] After mapping, hyperbolic distance is used to measure the semantic similarity between field symptoms and procedural clauses. Hyperbolic distance reflects both semantic similarity and the correspondence between field symptoms and procedural clauses in the hierarchical system. Hyperbolic distance is expressed as: , The smaller the hyperbolic distance, the closer the current on-site symptoms are to the corresponding procedural clause in terms of semantic content and hierarchical structure. To facilitate subsequent screening of candidate clauses, a matching score is calculated between the structured symptom text and the text of each procedural clause based on the hyperbolic distance: , in, To prevent the use of tiny constants with a denominator of zero. A higher matching score indicates that the corresponding procedural clause is more likely to apply to the current on-site symptoms.

[0043] S3. Based on the matching score, select candidate clauses that meet the preset conditions to form a set of candidate clauses.

[0044] Specifically, relying solely on the semantic similarity between structured symptom texts and procedural clause texts may lead to misjudgments where the superficial semantics are similar but the actual operating conditions do not hold true. Therefore, after obtaining candidate procedural clauses, this invention, based on hyperbolic space semantic matching, further introduces a causal clause network with a counter-evidence mask to perform causal consistency verification, counter-evidence elimination, and comprehensive scoring on candidate clauses and their corresponding candidate faults.

[0045] Based on the matching score calculated in step S2, the procedural clauses that meet the threshold conditions are selected to form a candidate clause set: , in, For the first The set of candidate clauses corresponding to each inspection; For the first Standardized procedure clause text; For the first Structured symptom text corresponding to the second inspection With the Articles and Terms Match score between them; This is the preset matching threshold.

[0046] S4. If any procedural clause in the candidate clause set has a directed edge with a fault node in the pre-constructed causal clause network, then output the corresponding fault node to form a candidate fault set.

[0047] After obtaining the candidate clause set, the system determines the candidate fault set based on the associated fault type field in the candidate procedure clauses and the causal relationship between the procedure clause node and the fault node.

[0048] Specifically, if the candidate procedure terms With faulty nodes If there is a directed edge between a procedure clause node and a faulty node, then... Included in the The set of candidate faults corresponding to each inspection. The set of candidate faults is represented as: , in, For the first The set of candidate faults corresponding to each inspection; For the first One candidate faulty node; For the set of candidate clauses, A directed edge from a clause node to a faulty node; It is the set of directed edges from the clause nodes to the faulty nodes, determined based on the pre-constructed causal clause network.

[0049] The causal clause network is a heterogeneous directed weighted graph structure built based on substation inspection business relationships in this embodiment of the invention, rather than directly calling existing general neural network models. This network uses structured symptom elements, standardized procedure clauses, and candidate fault types as different types of nodes, and symptom-triggered clauses, clause-directed faults, and symptom-historically associated faults as different types of directed edges, used to represent the triggering relationships, causal pointing relationships, and historical association relationships between symptoms, procedure clauses, and fault types.

[0050] The following describes the method for constructing the causal clause network in the embodiments of the present invention, including steps T1-T4.

[0051] T1. Construct a node set ;in, This is a set of symptom nodes used to represent device objects, abnormal parts, abnormal phenomena, alarm information, operating parameter information, and historical association information in structured symptom text; This is a set of procedural clause nodes used to represent the pre-normalized procedural clause text; This is a set of faulty nodes used to represent candidate fault types; T2. Construct the edge set ;in, These are directed edges from symptom nodes to procedure clause nodes, used to represent the triggering relationship between symptom elements and procedure clauses. A directed edge from a procedure clause node to a fault node, used to represent the causal relationship between the procedure clause and the fault type; These are directed edges from symptom nodes to fault nodes, used to represent the direct correlation between symptoms and faults in historical inspection records, historical work orders, or maintenance and reconsideration results; T3. Construct the edge weight set ;in, The trigger edge weight between the symptom node and the procedure clause node; The causal edge weight between the node of the procedure clause and the node of failure; The historical association edge weights between symptom nodes and faulty nodes; For the edge set Any directed edge Its corresponding edge weight ;in, For directed edges The right of the border; Nodes in historical samples and nodes The normalized value of co-occurrence frequency; For nodes and nodes Semantic matching score between them; For the confirmation value corresponding to manual annotation, prior verification of procedures, or maintenance and re-verification results; Let be the weight coefficient, and satisfy... ; Based on directed edges The type will determine the edge weight The set of trigger edge weights that are respectively assigned to symptom nodes and procedure clause nodes The set of causal edge weights from the procedure clause node to the fault node The set of historical associated edge weights from symptom nodes to faulty nodes The above three types of edge weights together constitute the edge weight set, which is then used in the comprehensive causal scoring of subsequent candidate faults.

[0052] T4. Forming a network of causal clauses ,like Figure 3 As shown.

[0053] S5. Filter the candidate clause set through the proof-of-contrast masking mechanism to obtain the valid clause set.

[0054] After obtaining the candidate clause set in step S4, the system introduces a rebuttal masking mechanism to exclude candidate clauses that conflict with the current on-site symptoms. This rebuttal masking mechanism does not filter solely based on semantic similarity, but rather judges based on the conflict relationships between key fields between the structured symptom text and the candidate procedure clause text. These key field conflict relationships include at least conflicts related to equipment objects, abnormal locations, abnormal phenomena, alarm threshold directions, negative descriptions, and time associations.

[0055] For the Structured symptom text generated during the second inspection Candidate Procedure Terms Define the proof by contradiction function as follows: , in, To prove the judgment function by contradiction, For the first Structured symptom text corresponding to the second inspection With the Candidate Procedure Clauses The mask value for the proof of contradiction between them.

[0056] The proof-of-contradiction function is expressed as follows: , in, For the conflict determination result; when When this occurs, it indicates that there is key counter-evidence between the current structured symptom text and the candidate procedure clause, and that candidate clause should be eliminated; when If the candidate clause is not refuted by evidence, it can continue to participate in subsequent causal scoring.

[0057] The conflict determination result can be expressed as: , in, For identifying device object conflicts; Indicates conflict in abnormal areas; This is an indicator of anomalies and conflicts. This serves as an indicator of a conflict between the alarm threshold and the status direction. To negate the conflict marker; For identifying time-related conflicts; symbol This represents a logical OR operation. If any conflict flag is true, then the candidate clause is considered to have a rebuttal relationship with the current on-site symptoms.

[0058] The candidate clause set is filtered using a proof-of-contrast masking mechanism to obtain the valid clause set: , in, This is the set of valid clauses after being eliminated by proof of contradiction.

[0059] S6. For each candidate fault in the candidate fault set, a comprehensive causal score is obtained by relating it to the procedural clauses in the valid clause set.

[0060] The comprehensive causal score for candidate faults is as follows: , in, For the first Candidate faults during the second inspection The comprehensive causal score; For the first The credibility score of the structured symptom text corresponding to the second inspection; This is the set of valid clauses after being eliminated by evidence of contradiction. For structured symptom text With standardized procedures and clauses The mask values ​​for the proof of contradiction between them; For structured symptom text With standardized procedures and clauses Match score between them; Symptom nodes To the nodes of the procedural clauses The trigger edge weight; For the nodes of the procedural clauses To candidate fault node Causal boundary weight; Symptom nodes To candidate fault node Historically related border rights; For the first The set of symptom nodes contained in the structured symptom text corresponding to the next inspection; This is the adjustment coefficient for the historically related edge weights.

[0061] Among them, the The credibility score of the structured symptom text corresponding to the second inspection is: , This represents the credibility of the data source, used to characterize the confidence level of recognition from sources such as speech recognition, text recognition, and historical work order parsing. This refers to the completeness of fields, which characterizes the degree of completeness of fields such as device object, abnormal location, abnormal phenomenon, alarm information, and historical association information. To ensure consistency across multiple sources, this is used to characterize whether the verbal statements of inspection personnel, on-site alarm texts, and historical work orders corroborate each other; Time validity is used to characterize the degree of correlation between alarm information, historical work orders, and the current inspection time; Let be the weight coefficient, and satisfy... .

[0062] Based on the comprehensive causal scoring described above, the system can further utilize the causal relationships between procedural clauses, historical work orders, maintenance records, and manual review results to verify the causal consistency of candidate faults, building upon the semantic matching results. For candidate clauses that appear semantically similar but have key conflicts with the on-site symptoms, the system eliminates them using a counter-evidence masking mechanism. For valid clauses that are not counter-evidenced, the system performs a comprehensive scoring based on their semantic matching score, network causal weight, historical association weight, and symptom text credibility, thereby obtaining a candidate fault ranking result that better reflects the current on-site situation.

[0063] S7. The candidate fault with the highest comprehensive causal score is taken as the final fault type for diagnosis.

[0064] The final diagnosed fault type is , in, The final diagnosed fault type corresponds to the r-th inspection. For candidate fault set The first in One candidate fault, For the first Candidate faults during the second inspection The comprehensive causal score.

[0065] Based on the final diagnosed fault type It also outputs the corresponding procedural basis, counter-evidence explanation, and handling suggestions. The final diagnosed fault type is used to indicate the most likely fault category of the current equipment. The procedural basis is used to indicate the valid procedural clauses that support the fault judgment. The counter-evidence explanation is used to indicate the candidate procedural clauses that have been eliminated by the counter-evidence masking mechanism and the reasons for their conflict. The handling suggestions are used to indicate the follow-up handling measures corresponding to the final diagnosed fault type.

[0066] If the highest comprehensive causal score is lower than the preset diagnostic confidence threshold, the system outputs a suspected fault result or a result requiring manual review, and retains the candidate fault ranking based on the comprehensive causal score, the corresponding procedural basis, the record of evidence of rebuttal and elimination, and the comprehensive score result for maintenance personnel to make further judgments.

[0067] Record the structured symptom text, candidate clause set, valid clause set, candidate fault set, disproven evidence masking results, final diagnosed fault type, procedural basis, disproven evidence explanation, and handling suggestions generated during this inspection process, and generate a diagnostic log: , in, For the first Diagnostic log of the next inspection task; For the first The structured symptom text corresponding to the next inspection; For a set of candidate terms; This is the set of valid clauses after being eliminated by evidence of contradiction. For the set of candidate faults; This is the mask value for the proof of contradiction; To ultimately diagnose the type of fault; For the basis of regulations; To demonstrate against the evidence; Recommendations for handling this matter.

[0068] After obtaining actual repair results, manual verification results, or further test results, compare the actual results with the current diagnostic results, and update the model parameters based on the comparison results:

[0069] in, For the first The set of parameters to be updated; For the updated parameter set; The learning rate; The loss function; For the first The actual repair results, manual verification results, or further test results corresponding to each inspection; For the first The diagnostic results corresponding to the next inspection.

[0070] The set of parameters to be updated This includes parameters for the hyperbolic embedding mapping function, semantic matching score calculation parameters, causal clause network edge weight calculation coefficients, structured symptom text credibility score weights, and historical association adjustment coefficients. Specifically, the hyperbolic embedding mapping function parameters are used to optimize the vector representation of structured symptom text and standardized procedure clause text in hyperbolic space; the semantic matching score calculation parameters are used to optimize the matching relationship between symptom text and procedure clause text; the causal clause network edge weight calculation coefficients are used to adjust the contribution ratio of historical co-occurrence frequency, semantic matching score, and manual confirmation information in edge weight calculation; the structured symptom text credibility score weights are used to adjust the impact of data source credibility, field completeness, multi-source consistency, and time validity on the final score; and the historical association adjustment coefficients are used to adjust the impact of historical work orders, maintenance records, or manual review results on candidate fault scores.

[0071] Among these, the hyperparameters in the model include the preset matching threshold for candidate terms. The thresholds for rebuttal conflict determination, diagnostic confidence, and scoring screening are not obtained directly through a single gradient update. Instead, they can be adjusted periodically based on manually reviewed samples, statistical results of verification samples, or the experience of operation and maintenance experts to ensure the accuracy, stability, and interpretability of the diagnostic results.

[0072] In practical applications, the method provided in this invention can collect voice recordings from inspection personnel, on-site image recognition, and equipment operating parameters through smart glasses. Combined with historical work order records, multi-source heterogeneous inspection information is converted into text and fused to form structured symptom text. Further, the symptom text and procedural clause text are semantically encoded and mapped together into a hyperbolic space. Hyperbolic distance is used to characterize the hierarchical semantic relationship between symptom descriptions and procedural clauses, thereby achieving preliminary matching and screening of candidate fault clauses. Based on this, a causal clause network is constructed, and a counter-evidence masking mechanism is used to verify causal consistency and eliminate counter-evidence from candidate clauses, suppressing interference clauses that are semantically similar but inconsistent with the actual situation on-site, thus improving the accuracy and reliability of fault judgment. Subsequently, the semantic matching score, causal network weight, and symptom text credibility are combined to comprehensively score candidate faults, determine the final fault type, and output the corresponding procedural basis, counter-evidence explanation, and handling suggestions. This method can be deployed in smart glasses terminals or their supporting edge computing modules for substation inspection scenarios where images or voiceprints are difficult to obtain stably, and can support inspection result recording and maintenance decision support.

[0073] The method provided in this invention is applicable to both initial inspection scenarios where image or voice information is difficult to obtain reliably, and substation intelligent diagnostic scenarios where conventional multimodal information is available. When image acquisition is limited (e.g., insufficient lighting, severe surface glare on equipment, environmental obstruction) or voice acquisition is interfered with (e.g., strong electromagnetic noise, noisy environment), traditional methods relying on visual recognition or voiceprint analysis struggle to obtain stable input. This invention unifies the textual processing of brief verbal descriptions from inspection personnel, equipment alarm information, and historical work order data. Even under single or incomplete modal input conditions, it can still construct effective structured symptom text, thereby ensuring the continuity and usability of the diagnostic process. By performing hyperbolic space semantic mapping between the structured symptom text and procedural clauses, it can still uncover hierarchical semantic relationships between the text and procedural knowledge, even when information is insufficient or expression is not standardized, achieving effective matching of candidate fault clauses. For long-term inspections or multi-round task scenarios, where the quality of image or voice data fluctuates significantly, historical inspection text records and hierarchical structures of procedures can be further utilized to perform correlation analysis on symptom information from different times and different devices, achieving cross-temporal compensation and optimization of diagnostic results. This invention can adapt to complex inspection environments where image or voice data is difficult to acquire stably, possessing good environmental adaptability and engineering practical value, and can be widely applied to intelligent inspection and auxiliary diagnosis scenarios under complex operating conditions in substations.

[0074] The above embodiments are only used to provide a detailed description of the technical solutions of this application. However, the descriptions of the above embodiments are only for the purpose of helping to understand the methods of the embodiments of the present invention and should not be construed as limiting the embodiments of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A substation fault diagnosis method based on hyperbolic reasoning, characterized in that, include: S1. Collect multimodal data from the substation inspection site and convert it into text data. After preprocessing the text data, obtain structured symptom text. S2. Semantically encode the structured symptom text and the pre-normalized procedure clause text respectively, and map them to hyperbolic space to calculate hyperbolic distance. Based on the hyperbolic distance, obtain the matching score between the structured symptom text and the procedure clause text. S3. Based on the matching score, filter out the procedural clause texts that meet the preset conditions to form a candidate clause set; S4. If any procedural clause in the candidate clause set has a directed edge with a fault node in the pre-constructed causal clause network, then output the corresponding fault node to form a candidate fault set. S5. Filter the candidate clause set through the proof-of-contrast masking mechanism to obtain the valid clause set; S6. For each candidate fault in the candidate fault set, a comprehensive causal score is obtained by associating it with the procedural clauses in the valid clause set. S7. The candidate fault with the highest comprehensive causal score is taken as the final fault type for diagnosis.

2. The method according to claim 1, characterized in that, Match score The calculation is performed in the following manner: ， in, Represents hyperbolic distance. It is a hyperbolic distance function. To prevent tiny constants with a denominator of zero; This represents the structured symptom text in hyperbolic space. For structured symptom text; For the first The representation of the provisions of the regulations in hyperbolic space. For the first The text of the pre-standardized procedure clauses; This is a hyperbolic space embedding mapping function.

3. The method according to claim 2, characterized in that, Methods for constructing causal clause networks include: Build a node set ;in, This is a set of symptom nodes used to represent device objects, abnormal parts, abnormal phenomena, alarm information, operating parameter information, and historical association information in structured symptom text; This is a set of procedural clause nodes used to represent the pre-normalized procedural clause text; This is a set of faulty nodes used to represent candidate fault types; Construct edge set ;in, These are directed edges from symptom nodes to procedure clause nodes, used to represent the triggering relationship between symptom elements and procedure clauses. A directed edge from a procedure clause node to a fault node, used to represent the causal relationship between the procedure clause and the fault type; These are directed edges from symptom nodes to fault nodes, used to represent the direct correlation between symptoms and faults in historical inspection records, historical work orders, or maintenance and reconsideration results; Construct the edge weight set ;in, The trigger edge weight between the symptom node and the procedure clause node; The causal edge weight between the node of the procedure clause and the node of failure; The historical association edge weights between symptom nodes and faulty nodes; Forming a network of causal clauses .

4. The method according to claim 3, characterized in that, Methods for constructing edge weight sets include: For the edge set Any directed edge , Boundary rights ;in, For directed edges The right to the side; Nodes in historical samples and nodes The normalized value of co-occurrence frequency; For nodes and nodes Semantic matching score between them; For the confirmation value corresponding to manual annotation, prior verification of procedures, or maintenance and re-verification results; Let be the weight coefficient, and satisfy... ; Based on directed edges The type will determine the edge weight The set of trigger edge weights that are respectively assigned to symptom nodes and procedure clause nodes The set of causal edge weights from the procedure clause node to the fault node The set of historical associated edge weights from symptom nodes to faulty nodes .

5. The method according to claim 4, characterized in that, The set of valid terms is as follows: ； in, For the first The set of candidate clauses corresponding to each inspection. For the first Standardized procedure clause text; For the first Structured symptom text corresponding to the second inspection With the Candidate Procedure Clauses The disproven mask values ​​between them are determined based on the conflict resolution result. Sure; ， in, For identifying device object conflicts; Indicates conflict in abnormal areas; This is an indicator of anomalies and conflicts. This serves as an indicator of a conflict between the alarm threshold and the status direction. To negate the conflict indicator; For identifying time-related conflicts; symbol This represents a logical OR operation; if any conflict flag is true, then... ,otherwise ; .

6. The method according to claim 5, characterized in that, The comprehensive causal score for candidate faults is as follows: ， in, For the first Candidate faults during the second inspection The comprehensive causal score; For the first The credibility score of the structured symptom text corresponding to the second inspection; This is the set of valid clauses after being eliminated by evidence of contradiction. For structured symptom text With standardized procedures and clauses The mask values ​​for the proof of contradiction between them; For structured symptom text With standardized procedures and clauses Semantic matching score between them; Symptom nodes To the nodes of the procedural clauses The trigger edge weight; For the nodes of the procedural clauses To candidate fault node Causal boundary weight; Symptom nodes To candidate fault node Historically related border rights; For the first The set of symptom nodes contained in the structured symptom text corresponding to the next inspection; This is the adjustment coefficient for the historically related edge weights.

7. The method according to claim 6, characterized in that, No. The credibility score of the structured symptom text corresponding to the second inspection. , This represents the credibility of the data source, used to characterize the confidence level of recognition from sources such as speech recognition, text recognition, and historical work order parsing. This refers to the completeness of fields, which characterizes the degree of completeness of fields such as device object, abnormal location, abnormal phenomenon, alarm information, and historical association information. To ensure consistency across multiple sources, this is used to characterize whether the verbal statements of inspection personnel, on-site alarm texts, and historical work orders corroborate each other; Time validity is used to characterize the degree of correlation between alarm information, historical work orders, and the current inspection time; Let be the weight coefficient, and satisfy... .

8. The method according to claim 6, characterized in that, The final diagnosis was ， in, The final diagnosed fault type corresponds to the r-th inspection. For candidate fault set The first in One candidate fault, For the first Candidate faults during the second inspection The comprehensive causal score.

9. The method according to claim 8, characterized in that, This also includes generating diagnostic logs, specifically: in, For the first Diagnostic log of the next inspection task; For the first The structured symptom text corresponding to the next inspection; For a set of candidate terms; This is the set of valid clauses after being eliminated by evidence of contradiction. For the set of candidate faults; This is the mask value for the proof of contradiction; To ultimately diagnose the type of fault; For the basis of regulations; To demonstrate against the evidence; Recommendations for handling this matter.

Images