SER event diagnosis method based on causal verification and semantic generalization

Abstract

The invention discloses an SER event diagnosis method based on causal verification and semantic generalization, and the method comprises the steps: firstly collecting historical event records of a converter station, and carrying out the priority division and dynamic time window aggregation; then, the events are constructed into a time-space-description triple, and a frequent item set is mined through correlation analysis; redundant rule pruning is carried out through confidence coefficient improvement, and false correlation is eliminated by adopting hybrid variable control; the causal intensity is verified based on ratio statistics, and reliable causal rules are screened; calling an intelligent language model to perform semantic analysis and tagging on the rule, extracting a core factor and generating a cross-scene generalization template; and finally, importing events, rules, tags and factors into a graph database, and constructing a visual causal knowledge graph. According to the method, the defects of redundancy and false association of the traditional association rule are effectively overcome, and the accuracy, interpretability and cross-scene migration capability of fault root cause positioning are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of converter station event diagnosis technology, and in particular relates to a SER event diagnosis method based on causal verification and semantic generalization. Background Technology

[0002] Converter stations, as a key component of high-voltage direct current (HVDC) transmission systems, are characterized by numerous devices, complex structures, and variable operating conditions. With the increasing scale and intelligence of systems, higher demands are placed on the intelligent operation and maintenance and fault diagnosis capabilities of converter stations. Existing event analysis methods for converter stations largely rely on experience-based judgment or traditional monitoring methods, limiting their ability to identify deep causal relationships between events and between events and faults. When a sudden fault occurs in the DC system and cannot be located and handled in a timely manner, it may lead to serious consequences such as equipment damage and grid disconnection, posing significant risks and economic losses to the operation of the power system.

[0003] During the evolution of faults or anomalies in converter stations, the system generates Sequential Event Recording (SER) data with millisecond-level precision. This type of data not only reflects equipment status and protection actions but also contains temporal and logical information about fault evolution, making it an important data source for fault diagnosis and event tracing. However, SER data is large in volume and complex in event types and triggering mechanisms, and manual analysis and traditional methods are significantly insufficient in quickly and accurately identifying key fault factors from massive amounts of events.

[0004] Recent technological advancements have demonstrated that intelligent language models based on semantic understanding and graph-based knowledge representation technologies offer significant advantages in information abstraction, semantic generalization, and relationship visualization. Intelligent language models can semantically process event texts and rule descriptions, enabling the standardization of event descriptions, extraction of key factors, and generation of cross-scenario generalization templates from specific rules. Graph databases, on the other hand, can organize events, rules, semantic tags, and core factors in the form of nodes and edges, supporting visualized queries and interactive analysis of causal links. While these technologies have shown promising applications in other fields, effectively combining the semantic generalization capabilities of intelligent language models with the graph representation of graph databases, along with redundant rule pruning and confounding variable control to ensure the reliability of analysis results, remains an unresolved technical challenge in the causal diagnosis of converter station SER data.

[0005] Therefore, there is an urgent need for a SER event diagnosis method based on causal verification and semantic generalization, which can retain the causal screening mechanism of statistical methods and improve the rule transferability and interpretability through semantic generalization and graphing, so as to support the accurate identification and visual tracing of key fault events in complex converter station operation scenarios. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a SER event diagnosis method based on causal verification and semantic generalization. It aims to solve the problem that existing methods for analyzing and predicting abnormal events in converter stations mostly focus on the superficial correlation between events and fail to fully identify the potential causal relationship between events and fault information, as well as between events and manual switching operations, resulting in insufficient diagnostic coverage and accuracy.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A SER event diagnosis method based on causal verification and semantic generalization includes the following steps: S1 collects historical sequential event records (SER messages), simulated fault information, and sequential control operation information of the converter station during the fault simulation and sequential control operation phases. It prioritizes the events and aggregates them based on a dynamic time window. S2, perform data layering and feature extraction on the aggregated event data, and construct a structured event representation in the form of a "time-space-description" triple; S3 aggregates events based on dynamic time windows, uses a support hierarchical strategy to filter event combinations, utilizes the Apriori algorithm to mine frequent itemsets, and generates candidate association rules. S4. Based on the confidence enhancement method, redundant rules are pruned for candidate association rules to remove weakly correlated or duplicate rules. S5 controls the rule set by defining and hierarchically processing confounding variables between events, thereby eliminating spurious associations. S6. The odds ratio (OR) causal statistics method is used to analyze the causal strength of the rules and screen out causal rules that meet the preset threshold. S7 converts causal rules into structured descriptive text, calls an intelligent language model for semantic parsing, generates semantic tags, and enhances the semantics of the event description; S8 generalizes causal rules based on an intelligent language model, extracts core factors and generates cross-scenario transferable generalized rule templates, and imports events, rules, tags, core factors and generalized templates into a graph database to construct a visualized causal knowledge graph.

[0008] Preferably, the dynamic time window in step S1 includes a fast fault window, a slow fault window, and a manual sequential operation window, and various events are aggregated according to different time scales.

[0009] Preferably, the "time-space-description" triple in step S2 includes the timestamp of the event, the location identifier of the device or area, and the text description of the event.

[0010] Preferably, the support stratification strategy in step S3 sets multi-level support thresholds based on event priority to improve the mining accuracy of important event combinations.

[0011] Preferably, the confidence enhancement method in step S4 includes calculating the original confidence, enhancement, and correlation direction between the antecedent and consequent of the rule, thereby eliminating redundant rules.

[0012] Preferably, the confounding variables in step S5 include changes in operating mode, external disturbances, human intervention, and measurement noise. The causal relationship is debiased by performing hierarchical processing on the confounding variables.

[0013] Preferably, the OR causal statistical method in step S6 includes calculating the ratio of the influence of the antecedent event on the triggering probability of the consequent event, and screening causal rules based on a significance threshold.

[0014] Preferably, the semantic labels generated by the intelligent language model in step S7 include at least one of the following: root cause labels, propagation link labels, protective action labels, measurement interference labels, sequential control operation labels, noise labels, and scene factor labels.

[0015] Preferably, the knowledge graph constructed in step S8 includes the following node types: event nodes, causal rule nodes, generalization rule template nodes, label nodes, and core factor nodes; the relationship edges include event causal relationships, rule application relationships, attribution relationships between rules and core factors, label assignment relationships, and mapping relationships between generalization rules and original rules.

[0016] Preferably, step S8 further includes verifying historical rules based on the constructed causal knowledge graph, and evaluating the stability, credibility, and cross-scenario transferability of the causal rules.

[0017] The beneficial effects of this invention are as follows: This method introduces an intelligent language model to perform semantic processing and generalization on candidate rules, including standardizing event descriptions in the rules, extracting key factors, and generating cross-scenario general templates to improve the transferability and interpretability of the rules. The generalization results, along with the original rules, event nodes, and semantic tags, are imported into a graph database. By constructing multi-type nodes and relational edges of events, rules, factors, and tags, and utilizing graph storage and display based on Neo4j, a visualized causal knowledge graph is formed. Based on the above causal statistical methods and graph representation, the causal evolution logic of events during the fault occurrence process and manual switching operation process of the converter station can be revealed, thereby achieving accurate identification and visual tracking of abnormal and important events, significantly improving the accuracy and efficiency of diagnosis. Attached Figure Description

[0018] Figure 1 This is a general framework diagram of important events in the correlation causal analysis of embodiments of the present invention. Detailed Implementation

[0019] Example 1: like Figure 1 As shown, a SER event diagnosis method based on causal verification and semantic generalization includes the following steps: S1 collects historical sequential event records (SER messages), simulated fault information, and sequential control operation information of the converter station during the fault simulation and sequential control operation phases. It prioritizes the events and aggregates them based on a dynamic time window. S2, perform data layering and feature extraction on the aggregated event data, and construct a structured event representation in the form of a "time-space-description" triple; S3 aggregates events based on dynamic time windows, uses a support hierarchical strategy to filter event combinations, utilizes the Apriori algorithm to mine frequent itemsets, and generates candidate association rules. S4. Based on the confidence enhancement method, redundant rules are pruned for candidate association rules to remove weakly correlated or duplicate rules. S5 controls the rule set by defining and hierarchically processing confounding variables between events, thereby eliminating spurious associations. S6. The odds ratio (OR) causal statistics method is used to analyze the causal strength of the rules and screen out causal rules that meet the preset threshold. S7 converts causal rules into structured descriptive text, calls an intelligent language model for semantic parsing, generates semantic tags, and enhances the semantics of the event description; S8 generalizes causal rules based on an intelligent language model, extracts core factors and generates cross-scenario transferable generalized rule templates, and imports events, rules, tags, core factors and generalized templates into a graph database to construct a visualized causal knowledge graph.

[0020] Preferably, the dynamic time window in step S1 includes a fast fault window, a slow fault window, and a manual sequential operation window, and various events are aggregated according to different time scales.

[0021] Preferably, the "time-space-description" triple in step S2 includes the timestamp of the event, the location identifier of the device or area, and the text description of the event.

[0022] Preferably, the support stratification strategy in step S3 sets multi-level support thresholds based on event priority to improve the mining accuracy of important event combinations.

[0023] Preferably, the confidence enhancement method in step S4 includes calculating the original confidence, enhancement, and correlation direction between the antecedent and consequent of the rule, thereby eliminating redundant rules.

[0024] Preferably, the confounding variables in step S5 include changes in operating mode, external disturbances, human intervention, and measurement noise. The causal relationship is debiased by performing hierarchical processing on the confounding variables.

[0025] Preferably, the OR causal statistical method in step S6 includes calculating the ratio of the influence of the antecedent event on the triggering probability of the consequent event, and screening causal rules based on a significance threshold.

[0026] Preferably, the semantic labels generated by the intelligent language model in step S7 include at least one of the following: root cause labels, propagation link labels, protective action labels, measurement interference labels, sequential control operation labels, noise labels, and scene factor labels.

[0027] Preferably, the knowledge graph constructed in step S8 includes the following node types: event nodes, causal rule nodes, generalization rule template nodes, label nodes, and core factor nodes; the relationship edges include event causal relationships, rule application relationships, attribution relationships between rules and core factors, label assignment relationships, and mapping relationships between generalization rules and original rules.

[0028] Preferably, step S8 further includes verifying historical rules based on the constructed causal knowledge graph, and evaluating the stability, credibility, and cross-scenario transferability of the causal rules.

[0029] Example 2: This embodiment provides a specific execution method, including the following steps: Step 1: Collect historical message information during the converter station fault simulation period and sequential control operation period, as well as simulated fault information and sequential control operation information. Prioritize the collected events and aggregate them based on dynamic time windows. Step 2: Perform data layering on the processed event data, extract the feature information of the events, construct the feature model of Sequential Event Record (SER), and form a time-space-description triple representation; Step 3: Aggregate important events within different time windows based on the dynamic time window mechanism, filter event combinations using a support hierarchical strategy, and mine frequent itemsets using the Apriori algorithm; Step 4: Based on the confidence enhancement method, perform redundant rule pruning on the causal rules generated from frequent itemsets, and filter out weakly correlated or duplicate rules; Step 5: By defining and stratifying the confounding variables between events, the candidate rules are controlled for confounding variables in order to remove spurious associations caused by confounding factors; Step 6: Perform causal strength analysis on the screened rule set based on the odds ratio (OR) causal statistics method to discover causal rules that meet the causal determination threshold; Step 7: Construct the causal rules into structured descriptive text, and perform semantic parsing of the causal rules by calling the intelligent language model to achieve the labeling of causal rules and semantic enhancement of event descriptions; Step 8: Based on the intelligent language model, further generalize the causal rules, extract the core factors of the rules, and form a generalized rule template that can be transferred across scenarios; import the event nodes, causal rule nodes, tag nodes and core factor nodes into the graph database to build a visualized causal knowledge graph, so as to realize the display and diagnosis of the causal links of important events in the converter station.

[0030] Furthermore, the dynamic time window in step 1 includes a fast fault window, a slow fault window, and a manual sequential control operation window, with different event categories aggregated according to different time scales.

[0031] Furthermore, the time-space-description triple constructed in step 2 includes the timestamp of the event, the device or region location identifier, and the textual description of the event.

[0032] Furthermore, the support hierarchy strategy adopted in step 3 includes multi-level support thresholds based on event priority to improve the mining accuracy of important event combinations.

[0033] Furthermore, the confidence enhancement method in step 4 includes: calculating the original confidence, enhancement, and correlation direction between the antecedent and consequent of the rule, thereby eliminating redundant rules.

[0034] Furthermore, the confounding variables in step 5 include changes in operating mode, external disturbances, human intervention, and measurement noise. By performing hierarchical processing of the confounding variables, the causal relationship can be debiased.

[0035] Furthermore, the OR causal statistical method used in step 6 includes calculating the ratio of the influence of the antecedent event on the triggering probability of the consequent event, and screening causal rules based on a set significance threshold.

[0036] Furthermore, the semantic labels generated by the Deepseek open model of intelligent language in step 7 include at least one of the following: root cause labels, propagation link labels, protective action labels, measurement interference labels, sequential control operation labels, noise labels, and scene factor labels.

[0037] Furthermore, the semantic enhancement process in step 7 includes extracting key event factors, standardizing event text descriptions, and generating a unified semantic representation of the rules.

[0038] Furthermore, the generalization rule templates generated by the intelligent language model in step 8 are represented by an "if-then" structure.

[0039] Furthermore, the knowledge graph constructed in step 8 includes the following node types: event nodes, causal rule nodes, generalization rule template nodes, tag nodes, and core factor nodes.

[0040] Furthermore, the relational edges constructed in step 8 include event causal relationships, rule application relationships, attribution relationships between rules and core factors, label assignment relationships, and mapping relationships between generalized rules and original rules.

[0041] Furthermore, the knowledge graph is stored based on a graph database, and the event propagation chain is visualized through node types and causal relationship edges.

[0042] Furthermore, the rule labeling and rule generalization of the intelligent language model support inference computation via external API calls.

[0043] Furthermore, the method further includes validating historical rules based on the constructed causal knowledge graph to assess the stability, credibility, and cross-scenario transferability of the mined causal rules.

[0044] Example 3: Before proceeding with the organization, we need to define "important events" and "exceptionally important events" in SER.

[0045] Critical events: Events classified as "Urgent" or "Alarm" in the SER (Search Engine Registry) are the core clues for fault diagnosis. Critical events originate from equipment failures, operational anomalies, and planned manual switching operations within the converter station, and are key to fault analysis.

[0046] Critical events refer to false alarms or missed alarms generated in the SER system due to secondary system anomalies such as communication interruptions or data acquisition device malfunctions. These events can be considered "pseudo-events." Such events indicate hidden vulnerabilities that may lead to misjudgments by maintenance personnel, directly threatening the accuracy and reliability of maintenance work.

[0047] Step 1: Event Prioritization and Dynamic Time Window Aggregation: Historical SER messages and related simulation data generated by the converter station during the fault simulation and sequential control operation phases are collected. To improve the quality of event screening, SER events are first prioritized based on their impact on system security, including: P1 type events: serious fault signals that may lead to equipment damage or converter station shutdown, requiring immediate action; P2 type events: These indicate that the system is in an abnormal or degraded state and needs to be checked promptly. P3 events: These include minor anomalies or status alerts, used for daily recording and monitoring.

[0048] To accommodate the varying durations of different fault types, fault events are clustered based on historical SER data, categorizing faults into fast and slow faults, and different dynamic time windows are set according to the fault category. The time of the first P1 type event is used as the starting point of the window; fast faults use an aggregation window of approximately 1 minute, and slow faults use an aggregation window of approximately 2.5 minutes; manual sequential control operations use independent window durations.

[0049] During the aggregation process, P1 and P2 type events are retained first, while low-confidence events that are not obviously related to P1 type events are removed, in order to construct an event set around the "fault evolution cycle" for subsequent frequent event set mining.

[0050] Step 2: SER Data Modeling The original SER messages are formatted and parsed to extract key event attributes, including timestamp, event level, system identifier, alarm group information, etc., and the structured features of the events are constructed based on time characteristics, device location features and event text descriptions.

[0051] To facilitate subsequent model processing, each event record is converted into a "time-space-description" triple representation, where: Time component: Describes the sequence of events; Spatial components: identify the location of equipment or systems; Description component: expresses the event category, nature, and alarm level.

[0052] This modeling approach preserves the temporal characteristics, device relationships, and semantic information of SER data, forming a unified event representation suitable for event mining.

[0053] Step 3: Frequent event set mining: Within the dynamic window constructed in step 1, frequent item mining is performed on the event set. To avoid missing key low-frequency events due to a single support threshold, a hierarchical support strategy is adopted, setting basic and scenario thresholds according to different fault types, so that low-frequency but important events can be retained.

[0054] Based on this strategy, the improved Apriori algorithm is used to obtain frequent combinations of events, and the corresponding support and confidence are calculated. Simultaneously, association rules between events are generated from the frequent itemset, and only candidate rules with confidence levels meeting preset requirements are retained.

[0055] Definition 4 Frequent itemsets: If the support of event set X is... If the set is frequently used, then it is considered a frequent itemset, and the support is calculated using the following formula: ; In the formula The number of transactions containing event set X. This represents the total number of transactions.

[0056] Extract association rules from frequent itemsets ( For the set of events mentioned above, For subsequent event sets (such as "protection trip → switch action"), only the confidence level is retained. The rule, the confidence level calculation formula is: ; Step 4: Redundant rule pruning based on confidence enhancement: Since Apriori mining may generate a large number of redundant rules containing duplicate or irrelevant factors, it is necessary to further filter out redundant items based on confidence enhancement metrics.

[0057] Specifically, by comparing the confidence gain of a rule with its simplified form, we can evaluate whether a rule contains effective factors that can provide additional information. A rule is only retained if it has a significant confidence increase relative to its simplified form. By adjusting the confidence increase threshold, the stringency of rule selection can be controlled, resulting in a higher-quality set of rules.

[0058] Confidence Boost: Calculate the confidence boost for the remaining rules Only retain The rules.

[0059] ; in, Representation rules Confidence level, All redundant rules were considered in the denominator. ,Require The preceding term for A subset of, and for A subset of the latter term.

[0060] Confidence Improvement The rules are described Advantages in providing additional information. As a set parameter, it only applies when Only then are the rules considered sufficiently non-redundant to be used for causal inference. This is achieved through adjustment. This allows us to obtain the number of rules corresponding to different levels of redundancy control, select the optimal confidence enhancement threshold, and obtain a new rule set.

[0061] Step 5: Further control confounding variables in the rule set through confounding variable definition and hierarchical structure. To reduce the interference of events that are unrelated to faults but overlap in time on the rules, the mixed variables in the events are processed in layers according to the functional correlation of the converter station alarm group.

[0062] Event variables are categorized into the following types based on their frequency of occurrence, distribution range, and physical correlation: Irrelevant variables: Events that occur infrequently in all samples and are not statistically significant; Specific variables: Events that occur only in a specific experiment or scenario and may lead to imbalance between groups; Function-independent variables: Events in which the alarm group has no physical connection with the functional area where the target event is located.

[0063] According to this hierarchical strategy, rules containing specific variables and functionally irrelevant variables are removed, while rules that are only affected by irrelevant variables are retained, thus obtaining a cleaned rule set that has been noise-controlled.

[0064] Control flow: rules Extract all non- Based on the above hierarchical logic, irrelevant variables, specific variables, and function-irrelevant variables are filtered from these variables; rules containing specific variables and function-irrelevant variables are removed from the rule set, and rules containing only irrelevant variables are retained to form a purified set of association rules.

[0065] Step 6: Causal rule mining based on OR causal statistics: For rules after redundancy control and confounding variable control, the causal strength between events is measured based on the ratio (OR) statistical method.

[0066] By dividing transactions into groups containing antecedent events and groups not containing antecedent events, the occurrence of consequent events in both groups is statistically analyzed, and the OR value and its confidence interval are calculated. A rule is considered to have a stable causal relationship only when the OR value of a rule reaches a preset threshold and the confidence interval meets the significance requirement.

[0067] This method uses statistically enhanced constraints to eliminate spurious causal rules caused by accidental co-occurrence or weak correlation, ultimately forming a reliable set of causal rules.

[0068] Calculate OR using a grouping approach: 1. Division between the exposure group and the control group: Exposure group ( ): A set of transactions containing the preceding itemset; control group ( ): A set of transactions that does not contain the preceding itemset.

[0069] 2. Statistical analysis using a 2D contingency table: : The number of transactions in the exposed group that contain the subsequent itemset; The number of transactions in the exposure group that do not include subsequent itemsets; The number of transactions in the control group that include the subsequent itemset; The number of transactions in the control group that do not include subsequent itemsets.

[0070] 3. OR calculation and confidence interval verification: The formula for calculating OR is as follows: ; 95% confidence interval (CI) calculation: ; Experimentally screened the OR threshold m, retaining only and The rules form the initial causal set. . OR reflects the degree to which the probability of B occurring increases or decreases when A occurs, compared to the probability of A not occurring.

[0071] In this invention, the OR value of a rule is used to characterize the strength of the causal relationship between the preceding and following event sets. When the OR value is close to 1, it can be considered that there is no significant dependency between the preceding and following events; when the OR value is greater than 1, the occurrence of the preceding event increases the probability of the following event occurring; when the OR value is less than 1, there may be an inhibitory relationship.

[0072] To avoid noisy rules caused by weak correlations or accidental co-occurrences being misjudged as causal, this invention introduces an OR threshold m during the causal rule selection process to constrain the strength of the OR value. Only when the rule... Furthermore, a rule is only considered a candidate causal rule if its lower confidence interval is higher than a preset standard. This screening mechanism effectively improves the reliability and robustness of causal rules.

[0073] Step 7: Causal rule labeling and semantic enhancement based on intelligent language model; After completing the initial causal association rules obtained from event sequence mining, this embodiment further introduces an intelligent language model to perform semantic analysis, label generation, and semantic enhancement on the rules, thereby improving the standardization and retrievalability of the rule expression. This step includes: Step 7.1: Rule-based structured coding For each causal rule, the antecedent event, consequent event, event description, alarm level, occurrence time, and statistical indicators (support, confidence, OR value, etc.) are organized into structured data, and a rule description text is constructed.

[0074] Step 7.2: Intelligent Language Model Invocation The structured rule text is taken as input, and the Deepseek open-source model is deployed locally. The intelligent language model and inference interface are called to obtain the semantic parsing results of the rules.

[0075] Intelligent language models, based on their pre-trained domain knowledge and language understanding capabilities, perform semantic classification and intelligent interpretation of rules.

[0076] Step 7.3: Rule Tag Generation Based on the model output, at least one semantic label is automatically generated for each causal rule, and the semantic labels include, but are not limited to: Fault root cause tags (e.g., converter valve faults, current loop abnormalities); propagation link tags (e.g., protection action trigger chain, measurement interference chain); manual operation tags; noise / false triggering tags; scenario factor tags; event type tags (e.g., "protection action", "switch operation", "communication abnormality"), etc.

[0077] Intelligent language models automatically select or combine multiple labels based on the logical relationships between events, semantic features, and pattern features in the domain corpus.

[0078] Step 7.4: Semantic Enhancement Processing The intelligent language model further enhances the semantics of the event text in the rules, including: 1. Extract the key factors of the event; 2. Eliminate redundant or ambiguous expressions; 3. Generate standardized event descriptions; 4. Output a unified semantic representation of the rules.

[0079] The above steps make the rules more universal and interpretable, providing a standardized information foundation for subsequent map construction.

[0080] Step 8: Rule generalization and visualization knowledge graph construction based on intelligent language model: After labeling the causal rules, this embodiment further utilizes an intelligent language model to perform generalized reasoning on the rules, generating a universal causal template that can be transferred across scenarios, and finally constructing a causal knowledge graph that spans events and rules, including the following steps: Step 8.1: Core Factor Extraction: Taking the semantically enhanced description of the rule as input, and leveraging the factor extraction capability of the intelligent language model, the core causal elements in the rule are identified, including: Equipment type factors; failure mechanism factors; protection / control action factors; environmental or operating condition factors, etc.

[0081] The core factor is used to describe the causal structure that the rule maintains a stable state across different sites and topological environments.

[0082] Step 8.2: Rule generalization generation: Based on the aforementioned core factors, the intelligent language model generates generalized expressions of rules, forming a universal causal rule template that can be transferred across scenarios.

[0083] Generalized content includes: Remove inherent features associated with the site; Retain common characteristics related to equipment category and failure mechanism; Output a standardized causal expression in "if-then" format.

[0084] The generalized rules are used for consistency diagnosis and knowledge reuse in different engineering scenarios.

[0085] Step 8.3: Knowledge Graph Node Generation: Based on information such as rule tags, event nodes, and core factors, knowledge graph nodes are generated, including: Event nodes; causal rule nodes; tag nodes; core factor nodes; rule template nodes, etc.

[0086] Each node contains corresponding attribute fields (such as category, timestamp, confidence level, statistical indicators, etc.).

[0087] Step 8.4: Knowledge Graph Relationship Construction: Based on the semantic structure of the rules, establish multiple types of relationships between nodes, including: Causal relationships between events; correlation between rules and events; attribution relationships between rules and core factors; mapping relationships between generalized rules and original rules; and label relationships between labels and events / rules.

[0088] The relationships are stored in the graph database Neo4j and persistently managed using a graph structure.

[0089] Step 8.5: Visualization of the knowledge graph: Based on the graph query and visualization capabilities of graph databases, causal links, tag information, core factors and generalization templates are presented in the form of visual graphs, which facilitates operation and maintenance personnel to perform fault location, root cause analysis and rule reuse.

Claims

1. A SER event diagnosis method based on causal verification and semantic generalization, characterized in that, Includes the following steps: S1 collects historical sequential event records (SER messages), simulated fault information, and sequential control operation information of the converter station during the fault simulation and sequential control operation phases. It prioritizes the events and aggregates them based on a dynamic time window. S2, perform data layering and feature extraction on the aggregated event data, and construct a structured event representation in the form of "time-space-description" triples; S3 aggregates events based on dynamic time windows, uses a support hierarchical strategy to filter event combinations, utilizes the Apriori algorithm to mine frequent itemsets, and generates candidate association rules. S4. Based on the confidence enhancement method, redundant rules are pruned for candidate association rules to remove weakly correlated or duplicate rules. S5 controls the rule set by defining and hierarchically processing confounding variables between events, thereby eliminating spurious associations. S6. The odds ratio (OR) causal statistics method is used to analyze the causal strength of the rules and screen out causal rules that meet the preset threshold. S7 converts causal rules into structured descriptive text, calls an intelligent language model for semantic parsing, generates semantic tags, and enhances the semantics of the event description; S8 generalizes causal rules based on an intelligent language model, extracts core factors and generates cross-scenario transferable generalized rule templates, and imports events, rules, tags, core factors and generalized templates into a graph database to construct a visualized causal knowledge graph.

2. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, The dynamic time window in step S1 includes a fast fault window, a slow fault window, and a manual sequential control operation window, and various events are aggregated according to different time scales.

3. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, The "time-space-description" triple in step S2 includes the timestamp of the event, the location identifier of the device or area, and the text description of the event.

4. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, The support stratification strategy in step S3 sets multi-level support thresholds based on event priority to improve the mining accuracy of important event combinations.

5. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, The confidence enhancement method in step S4 includes calculating the original confidence, enhancement, and correlation direction between the antecedent and consequent of the rule, thereby eliminating redundant rules.

6. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, The confounding variables in step S5 include changes in operating mode, external disturbances, human intervention, and measurement noise. By performing hierarchical processing on the confounding variables, the causal relationship can be debiased.

7. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, The OR causal statistical method in step S6 includes calculating the ratio of the influence of the antecedent event on the triggering probability of the consequent event, and filtering causal rules based on a significance threshold.

8. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, The semantic labels generated by the intelligent language model in step S7 include at least one of the following: root cause labels, propagation link labels, protective action labels, measurement interference labels, sequential control operation labels, noise labels, and scene factor labels.

9. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, The knowledge graph constructed in step S8 includes the following node types: event nodes, causal rule nodes, generalization rule template nodes, label nodes, and core factor nodes; the relationship edges include event causal relationships, rule application relationships, attribution relationships between rules and core factors, label assignment relationships, and mapping relationships between generalization rules and original rules.

10. The SER event diagnosis method based on causal verification and semantic generalization according to claim 1, characterized in that, Step S8 further includes verifying historical rules based on the constructed causal knowledge graph, and evaluating the stability, credibility, and cross-scenario transferability of the causal rules.

Images