A method for tracing the source of anomalies in smart power equipment

By using unified identifier mapping, field standardization, and time alignment methods, combined with knowledge graph and verification strategy optimization, the problem of anomaly tracing caused by multi-source data heterogeneity is solved. This improves the accuracy and reliability of anomaly tracing in smart power plant equipment, reduces computing resource costs, and enables the graph relationships to be continuously corrected as operation and maintenance change.

CN123116071APending Publication Date: 2026-11-17SHANDONG DESHENG ELECTRIC POWER CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611293958.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, multi-source data from smart power plants suffers from inconsistent field units, different sampling periods, out-of-order arrival and missing data, and inconsistent recording granularity of the same event by different systems. This makes it difficult to directly correlate and synchronize the data, affecting the reliability of anomaly analysis. Furthermore, the unit's operating status is affected by load, start-up and shutdown, valve position combinations, and environmental conditions. The same anomaly can manifest differently under different operating conditions, which can easily lead to miscorrelation and misjudgment.

Method used

By unifying identifier mapping, field standardization, and time alignment, multi-source heterogeneous data are aligned under the same caliber and time grid. A knowledge graph is constructed, and a set of effective time periods and applicable working condition labels for relation edges is introduced. During the subgraph extraction and inference stages, time intersection, working condition inclusion, and time sequence consistency constraints are executed. A verification strategy space is constructed and the verification strategy is optimized. A consistency score is generated to eliminate hypotheses that do not meet the time sequence and working condition logic, and the confidence of relation edges in the knowledge graph is updated.

Benefits of technology

It improves the accuracy and reliability of anomaly tracing in smart power equipment, reduces false associations and misjudgments, enhances data consistency and availability, reduces computing resource costs, and enables continuous correction of graph relationships as operation and maintenance change, resulting in more stable tracing performance over long-term operation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116071A_ABST
    Figure CN123116071A_ABST
Patent Text Reader

Abstract

This invention provides a method for tracing the source of anomalies in intelligent power equipment, belonging to the field of intelligent diagnostic technology for power plant equipment. The method includes: acquiring multi-source operating data from the power plant and performing unified identification mapping, field standardization, time alignment, and operating condition label generation; forming structured event data based on event extraction rules; constructing a knowledge graph of equipment entity nodes, measurement point nodes, event nodes, and their relational edges based on the structured event data, and setting effective time periods and applicable operating condition label sets for the relational edges; filtering target anomaly events, generating target equipment ranges and target time windows, and extracting candidate anomaly subgraphs in the knowledge graph based on time intersection, operating condition inclusion, and range filtering. This method can improve the accuracy and interpretability of anomaly root cause localization under temporal and operating condition constraints, and enhance the continuous evolution capability of the knowledge graph.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for power plant equipment, and in particular to a method for tracing the source of anomalies in intelligent power equipment. Background Technology

[0002] Under safe and stable unit operation, smart power plants typically deploy multi-source data platforms, including monitoring systems, equipment management systems, maintenance work order systems, and ledger topology systems, to collect data such as measurement point curves, alarm and protection actions, control commands, maintenance defect records, and equipment topology information. Analyzing this data to pinpoint the root causes of equipment anomalies is a crucial means of power plant operation and maintenance diagnosis and accident prevention.

[0003] Existing anomaly analysis methods are mostly based on single-source data or the fusion of a small amount of data. However, multi-source data suffers from problems such as inconsistent field units, different sampling periods, out-of-order arrival and missing data, and inconsistent granularity of records for the same event by different systems. This makes it difficult to directly correlate and synchronize data, thus affecting the reliability of anomaly analysis. At the same time, the operating status of the unit is affected by load, start-up and shutdown, valve position combination and environmental conditions. The same anomaly can show significant differences under different operating conditions. If there is a lack of operating condition constraints and time sequence consistency constraints, it is easy to make incorrect correlations and misjudgments.

[0004] Therefore, we propose a method for tracing anomalies in smart power equipment; the information disclosed above in the background section is only for enhancing the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for tracing the source of anomalies in smart power equipment, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for tracing the source of anomalies in smart power equipment includes the following steps:

[0008] S1. Acquire multi-source operation data of the power plant, perform unified identification mapping and field standardization on the data, and perform time alignment and operation condition label generation. Generate structured event data according to event extraction rules.

[0009] S2. Construct a knowledge graph based on structured event data. The knowledge graph includes equipment entity nodes, measurement point nodes, event nodes and their relationship edges, and set effective time periods and applicable working condition label sets for the relationship edges; filter target abnormal events, generate target equipment range and target time window, and extract abnormal candidate subgraphs in the knowledge graph by time intersection, working condition inclusion and range filtering.

[0010] S3. Perform reachability, timing and operating condition consistency constraints on the abnormal candidate subgraph to obtain the source reasoning space, generate candidate root cause hypotheses and extract evidence chains, and construct observable response descriptions based on the evidence chains.

[0011] S4. Construct a verification strategy space and define consistency score, discrimination and verification cost. Use the fruit fly algorithm to optimize the verification strategy to obtain the optimal verification strategy. Under the optimal verification strategy, compare the predicted observable response with the actual observable response to generate a consistency score and perform proof by contradiction elimination to converge the root cause hypothesis set.

[0012] S5. Determine the target root cause based on the converged root cause hypothesis set and output the anomaly tracing results; update the confidence of the knowledge graph relation edges based on the evidence chain corresponding to the target root cause and write back the version; and construct an active learning task queue based on information gain scoring to achieve self-updating of the knowledge graph.

[0013] S1 specifically includes: configuring multi-source connections and acquiring real-time measurement points, alarms and protection actions, control commands and operation tickets, maintenance work orders and equipment ledger data, generating original records and generating acquisition failure records when acquisition fails;

[0014] Perform a unified mapping from the original identifier to the global device identifier and the global measurement point identifier on the original record, and perform unit completion, unit conversion and valid range verification according to the field standardization rule table to generate standardized records;

[0015] Establish a unified time base and generate a time grid, align standardized records to generate time-aligned records and time-aligned sequences, and generate working condition labels for time-aligned records;

[0016] Based on the event type table, perform event-based extraction on time-aligned records and event-type data, continuously determine measurement point over-limit events, establish event deduplication keys, generate structured event data, and output event identifiers.

[0017] S2 specifically includes: constructing a knowledge graph storage and writing device entity nodes, measurement point nodes and event nodes based on node deduplication keys; establishing relationship edges such as topological connections, control dependencies, material or energy flow directions and time sequence triggers; and writing initial confidence values ​​for relationship edges to form an initial knowledge graph.

[0018] Configure valid time periods for relation edges and perform version updates of relation edges during maintenance or topology changes. At the same time, configure a set of applicable working condition labels for relation edges to form a time-series working condition constraint knowledge graph.

[0019] Target abnormal events are filtered based on abnormal triggering conditions, and target device range and target time window are generated to form target abnormal event input;

[0020] In the time-series operating condition constraint knowledge graph, subgraph extraction is performed starting from the target abnormal event. Time intersection filtering, operating condition inclusion filtering and range filtering are performed in sequence, and candidate abnormal subgraphs are obtained by connectivity expansion.

[0021] S3 specifically includes: limiting the allowed relation edge types on the anomaly candidate subgraph and executing topological reachability constraints according to directional rules, generating node reference times and performing temporal sequence constraints and working condition consistency constraints to obtain the source inference space;

[0022] Within the source tracing reasoning space, a candidate pool of cause points is constructed. Based on abnormal signs, candidate root cause hypotheses are generated and a set of associated paths is formed. The initial root cause score is calculated and normalized. Evidence chains are extracted from candidate root cause hypotheses as the starting point and target abnormal events as the ending point. After the inclusion condition is met, scoring, deduplication, and long path suppression are performed to generate evidence chain identifiers. Based on the evidence chains, observable response descriptions are constructed to determine the target measurement point set and its change direction, change range, and prediction linkage sequence. A candidate root cause hypothesis-observable response description mapping table is output.

[0023] S4 specifically includes: constructing a verification strategy space based on candidate root cause hypotheses and observable response description mapping tables, generating a set of verification time window numbers, a set of measurement point numbers, a set of verification model numbers, and a candidate set of verification order, and performing feasibility determination on strategy items;

[0024] Define consistency scores, discrimination scores, and verification costs, set computational resource constraints, and determine the merits of strategies according to fitness comparison rules;

[0025] The verification strategy is encoded as a strategy configuration object and optimized using the fruit fly algorithm. The optimal verification strategy is then iteratively updated and output based on the discrete perturbation rule.

[0026] Under the optimal verification strategy, predictive and actual observable responses are generated. Based on the baseline window location, the actual change time and magnitude are extracted. The consistency of direction, the consistency of magnitude range and the consistency of sequence are checked in turn. A consistency score is synthesized and the evidence is rejected by contradiction according to the threshold.

[0027] S5 specifically includes: forming a candidate ranking list based on the consistency score of the root cause hypothesis set that has converged after elimination by contradiction; calculating the difference between the first two and determining the target root cause when the difference threshold is met; or determining the target root cause according to the decision chain rule and generating a decision path record when the difference between the first two is insufficient.

[0028] Read the evidence chain identifier set and observable response description corresponding to the target root cause, generate an anomaly tracing result object containing sub-scores and predictions, actual comparison summaries, and output and store it in the database;

[0029] For the evidence chain relationship, the request list is updated while it is being generated. The confidence level is increased or decreased and upper and lower limits are applied according to the update step size parameter. Multiple requests on the same side are merged and conflict adjudicated. The graph elements are written back in a versioned manner, written to the graph change log and the updated knowledge graph version number is output.

[0030] A pool of candidate triples to be verified is constructed, and the coverage, uncertainty and impact are calculated to synthesize the information gain score. High information gain triples are selected to generate an active learning task queue. The verification results are written back and updated in version according to the verification action type.

[0031] The beneficial effects of this invention are as follows:

[0032] This invention aligns multi-source heterogeneous data under the same caliber and time grid by unifying identifier mapping, field standardization, and time alignment. This reduces the problem of "fragmented evidence that cannot be pieced together" caused by out-of-order, missing, and inconsistent granularity data, improving the consistency and usability of source tracing input data. It introduces effective time periods and applicable working condition label sets for relation edges into the knowledge graph, and performs time intersection, working condition inclusion, and time sequence consistency constraints during subgraph extraction and inference stages. This avoids misassociating events from different working conditions / time periods with changes in measurement points, improving the accuracy of root cause localization.

[0033] This invention constructs a structured output of the "cause-path-expected response-actual response" evidence chain through candidate root cause hypotheses, evidence chain extraction, and observable response description. This makes the conclusions traceable and auditable, reducing the subjectivity of relying on human experience. A verification strategy space is constructed, and strategies are optimized based on consistency scores, discriminative power, and verification costs. Under computational resource constraints, more discriminative and lower-cost verification schemes are prioritized, reducing simulation / model calls and measurement point reading costs.

[0034] This invention performs component-by-component comparisons and synthesizes a consistency score based on directional consistency, amplitude range consistency, and linkage sequence consistency. This allows for the elimination of hypotheses that do not meet the timing and operational logic requirements through disproven evidence, improving convergence speed and the final root cause confidence level. Based on the target root cause evidence chain, the confidence level of relational edges is incrementally updated and versioned back, forming a traceable graph change record. This ensures that graph relationships are continuously corrected as operation and maintenance change, resulting in more stable traceability performance over long-term operation. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of a method for tracing the source of anomalies in intelligent power equipment according to the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1: As Figure 1 As shown, this embodiment provides a method for tracing the source of anomalies in smart power equipment, including the following steps:

[0038] S1. Acquire multi-source operation data from the power plant, perform unified identifier mapping and field standardization on the data, and generate time alignment and operating condition labels. Generate structured event data according to event extraction rules. This includes the following sub-steps:

[0039] S110, Multi-source data acquisition: Configure multi-source connections in the data access layer; where “multi-source connection” refers to the communication configuration items used to obtain data from different business systems or acquisition systems, including at least connection address, authentication information, data acquisition method, and acquisition cycle or subscription topic;

[0040] Multi-source connections include at least: real-time measurement point data connection, alarm and protection action data connection, control command and operation ticket data connection, maintenance defect and work order data connection, and equipment ledger and topology relationship data connection.

[0041] Data acquisition is performed based on multi-source connections: when the data acquisition method is polling, queries are triggered according to the acquisition cycle; when the data acquisition method is subscription, push notifications are received according to the subscribed topics; a "raw record" is generated for each received data.

[0042] The original record must include at least: data source type, original timestamp, original device identifier, original measurement point identifier, original field key-value pairs, and receiving time; the receiving time is used to handle out-of-order data arrival and audit traceability.

[0043] When any multi-source connection times out, fails authentication, returns a null value, or fails to resolve, a "collection failure record" is generated. The collection failure record includes at least: the type of the failed data source, the time of failure, the reason for failure, the number of retries, and the corresponding query range or subscription offset. The collection failure record is written to the "failure cache", which is a collection of collection failure records stored by data source type for subsequent retries and auditing.

[0044] The successfully parsed original records are written to the "original data cache", which refers to the collection of original records stored in partitions according to the data source type. For each original record, both the original text of the original field key-value pairs and the parsed field key-value pairs are saved to ensure that subsequent calculations and traceability are possible.

[0045] S120. Unified Identifier and Field Standardization: Create a "Unified Identifier Mapping Table", which is a mapping dictionary that maps the original device identifier and the original measurement point identifier to the "Global Device Identifier" and the "Global Measurement Point Identifier". The Global Device Identifier and the Global Measurement Point Identifier are unique and stable identifiers within this method and are used to associate the same device and the same measurement point across data sources.

[0046] Traverse each original record in the original data cache and perform a unified identification: prioritize matching based on complete consistency of asset codes in the equipment ledger; if the asset code is missing, use matching based on both equipment name and installation location; if multiple matching candidates appear, select a unique candidate in the order of "system affiliation consistency priority, historical occurrence frequency highest priority"; if the above matching fails, perform a secondary matching based on the alias table; if the secondary matching still fails, write the original record to the "pending manual verification queue" and mark it as an unmapped record, where the pending manual verification queue refers to the set of original records for which a unique mapping relationship cannot be determined.

[0047] Create a "Field Standardization Rule Table", which refers to a set of rules that define the uniformity of fields in various data sources. The Field Standardization Rule Table should include at least: standard field name, target unit, unit conversion method, numerical precision retention rules, valid value range, and default unit for measurement points.

[0048] For successfully mapped original records, perform field standardization and generate "standardized records": when the original unit is missing, fill it in according to the default unit of the measurement point; when the original unit and the target unit are convertible, convert the value to the target unit according to the unit conversion method, and round or truncate according to the numerical precision retention rules; when the original unit is not convertible or the default unit of the measurement point is missing, mark the corresponding field as a non-standardized field and write it to the "standardized anomaly cache", where the standardized anomaly cache refers to the collection that stores non-standardized fields and their reasons; perform a valid value range check on the converted value, and mark it as an anomaly field if it exceeds the limit, while retaining both the original value and the standardized value; output the standardized record, which at least includes: global device identifier, global measurement point identifier (can be empty), unified timestamp (can be empty), standard field key-value pairs, working condition label (can be empty), and anomaly mark set.

[0049] S130, Time Alignment and Operating Condition Label Generation: Create a "Unified Time Baseline", which refers to converting timestamps from different data sources into time expressions with the same time zone and the same precision; generate unusable time records for unresolved timestamps and write them to a "Time Anomaly Queue", which refers to a collection that stores unusable time records and their reasons; unusable time records do not participate in subsequent time alignment calculations.

[0050] A "time grid" is generated, where a time grid refers to a time point sequence discretized according to a fixed sampling period. The sampling period is preferentially taken from the sampling period configured in the measurement point ledger. When the ledger is missing, the time difference between two adjacent records of the measurement point is counted in the statistical window specified by the system configuration parameters, and the time difference with the most occurrences is selected as the sampling period. Continuous time points are generated from the start to the end of the target time period according to the sampling period to form a time grid. Each standardized record is then assigned to the corresponding time grid point or between two adjacent time grid points according to its timestamp.

[0051] Generate "Time Alignment Records" and "Time Alignment Sequences": "Time Alignment Records" refer to grid-by-grid records indexed by time grids. Each time alignment record contains at least the timestamp of the grid point, the alignment value of each measurement point at the grid point, and the working condition label corresponding to the grid point.

[0052] "Time-aligned sequence" refers to a single-point sequence view extracted from a time-aligned record according to a certain measurement point, which is used to perform scanning judgment and duration calculation for that measurement point;

[0053] For each measurement point, a time-aligned sequence is generated on the time grid: when a time grid point has a valid value for that measurement point, the value is directly taken; when it is missing, the value is filled in according to the interpolation or filling method specified by the system configuration parameters, and the same process is maintained for the same measurement point throughout the entire process; when the number of consecutive missing time grid points exceeds the upper limit of the missing segment specified by the system configuration parameters, no value is filled in but it is marked as a missing segment to avoid long gaps causing incorrect value filling.

[0054] Generate "operating condition tags" and write them to time-aligned records: Operating condition tags are discrete markers of the operating status of a specific time grid segment, and are obtained by concatenating at least the following components: Load segment: obtained by mapping the unit load to the interval specified by the system configuration parameters; Start-up / Stop status: determined on the time grid based on the order of occurrence of start-up events, grid connection events, and shutdown events; Key valve position combination: obtained by discretizing the valve position openings of the key valve position set specified by the system configuration parameters on the time grid according to intervals and then concatenating and encoding them in a fixed order; Environmental conditions: obtained by mapping ambient temperature and humidity to intervals; Write the operating condition tags to all time-aligned records within the corresponding time grid segment, so that subsequent event extraction is performed under a unified time base and a unified operating condition context.

[0055] S140. Event-based extraction and structured event generation: Create an "Event Type Table", which refers to the set of event extraction rules; the Event Type Table should at least include: alarm events, protection action events, control command events, operation ticket events, maintenance events, defect events, and measurement point over-limit events; and configure trigger conditions, event attribute field sets, and event start and end time generation rules for each event type.

[0056] Event-based extraction is performed on time-aligned records and original event records: for alarm and protection action records, events of the corresponding event type are generated according to their event codes or action codes; for control instructions and operation tickets, events are generated according to the instruction issuance time or ticket execution time; for maintenance and work order records, events are generated according to their effective time and shutdown time; and the operating condition label of the generated event is written into the event.

[0057] For measurement point over-limit events, a continuous determination is performed: the alignment value of the measurement point is traversed on the time alignment sequence to find a continuous grid point segment that satisfies the condition that "the measurement point value exceeds the over-limit threshold"; the duration of the continuous segment is defined as the time difference between the grid point time at the beginning of the segment and the grid point time at the end of the segment; when the duration is not less than the duration threshold specified by the system configuration parameters, a measurement point over-limit event is generated, and the time at the beginning of the segment is taken as the event occurrence time and the time at the end of the segment is taken as the event end time; the over-limit threshold and the duration threshold are fixed to be derived from the measurement point ledger or rule base.

[0058] Generate and deduplicate "structured event data": Structured event data refers to event records in a unified format. Each record must contain at least a global device identifier, a global measurement point identifier (which can be empty), an event type, an event occurrence time, an event end time (which can be empty), standardized event attributes, operating condition tags, and a set of anomaly markers. To avoid duplicate generation, an "event deduplication key" is established. The event deduplication key is generated by concatenating the global device identifier, event type, event occurrence time, and key event attributes according to a fixed concatenation rule. When the event deduplication keys are the same, they are considered the same event and the same "event identifier" is reused. The event identifier refers to the unique identifier of the event in the event storage. The structured event data is written to the event storage and the event identifier is output for reference in subsequent knowledge graph modeling.

[0059] S2. Construct a knowledge graph based on structured event data. The knowledge graph includes equipment entity nodes, measurement point nodes, event nodes, and their relationship edges. Valid time periods and applicable operating condition label sets are set for the relationship edges. Target abnormal events are filtered, target equipment ranges and target time windows are generated, and candidate abnormal subgraphs are extracted from the knowledge graph based on time intersection, operating condition inclusion, and range filtering. Specifically, this includes the following sub-steps:

[0060] S210. Knowledge Graph Element Modeling: Create a "Knowledge Graph Storage", which refers to the storage structure used to store graph data, and includes at least a node table, an edge table, an attribute table, and an index table. The index table should at least support locating nodes by node type and node identifier, and locating edges by edge type and start node identifier or end node identifier, so as to ensure that subgraphs can be quickly queried and extracted according to conditions.

[0061] Create a "node deduplication key", where the node deduplication key is a unique field used to determine whether two nodes belong to the same entity. For device entity nodes, the node deduplication key is composed of the global device identifier; for measurement point nodes, the node deduplication key is composed of the global measurement point identifier; for event nodes, the node deduplication key is composed of the event identifier. When writing a node, if the node deduplication key already exists, the existing node is reused and its attribute version number is updated, without creating a duplicate node.

[0062] Write the global device identifier corresponding object as a device entity node, write the global measurement point identifier corresponding object as a measurement point node, and write the structured event data as an event node; each node must at least write the node type, node identifier, node deduplication key, and node attribute set, and the node attribute set must at least include the source data source type and writing time.

[0063] Create and write "relationship edges", where a relationship edge refers to a record that connects two nodes and expresses the meaning of association; relationship edges include at least topological connection relationship edges, control dependency relationship edges, material or energy flow relationship edges, and time-triggered relationship edges; for each relationship edge, at least write the edge type, starting node identifier, ending node identifier, relationship source, and generation rule version.

[0064] Write an "initial confidence value" for each relation edge, where the initial confidence value refers to the reliability value of the relation edge when it is first written. The rules for the initial confidence value are fixed to the system configuration parameters and at least satisfy the following: when the relation source is a strong source such as equipment ledger or design topology, take a higher initial confidence value; when the relation source is a weak source such as log inference or statistical association, take a lower initial confidence value. Write the initial confidence value and the relation source together into the edge attribute to ensure that the source of the confidence is traceable.

[0065] Output "Initial Knowledge Graph", where the initial knowledge graph refers to the graph data version after the nodes and relation edges have been written, and write this version number as the input version for subsequent time-series updates and subgraph extraction.

[0066] S220. Temporalization and Constraint of Relationship Edges: Add an "effective time period" to each relationship edge, where the effective time period refers to the time range in which the relationship edge is valid, and is represented by at least two fields: effective start time and effective end time; when the effective end time is empty, it means that the relationship edge is still valid at the current time.

[0067] When a maintenance event or topology change event occurs, perform a versioning update of the relation edges: determine the set of affected relation edges, and write the time of the change to the valid end time of these relation edges, so that they are no longer valid after the time of the change; then create an alternative relation edge version, which reuses the same edge type and endpoint node identifier and writes a new valid start time as the time of the change, thereby ensuring that there are no two conflicting and simultaneously valid relation edge versions at the same point in time.

[0068] Write the "Applicable Working Condition Label Set" for relation edges that require working condition constraints. The Applicable Working Condition Label Set refers to the set of working condition labels that allow the relation edge to participate in reasoning. When the Applicable Working Condition Label Set is empty, it means that the relation edge has no restrictions on working conditions. When the Applicable Working Condition Label Set is not empty, the relation edge is allowed to enter the subsequent subgraph extraction and reasoning only if the working condition label of the target abnormal event belongs to the set.

[0069] Output "Time-series working condition constraint knowledge graph", where the time-series working condition constraint knowledge graph refers to the version of the knowledge graph that includes the constraints of the effective time period and the set of applicable working condition labels, and this version number is used as the unique input version number for the subsequent S240 subgraph extraction.

[0070] S230. Target Abnormal Event Determination: Read structured event data from the event store and filter target abnormal events according to "abnormal triggering conditions";

[0071] The abnormal triggering conditions include at least the following: the alarm level reaches the level threshold specified by the system configuration parameters, a protection action event occurs, the duration of the measurement point over-limit event is not less than the duration threshold specified by the system configuration parameters, and the number of alarm events for the same global device identifier within the time window specified by the system configuration parameters is not less than the aggregation number threshold specified by the system configuration parameters.

[0072] For each target anomaly event, a "target device range" is generated: the device entity node corresponding to the global device identifier associated with the target anomaly event is used as the central node, and the expansion is performed layer by layer along the topology connection edge. The number of expansion layers does not exceed the preset number of hops specified by the system configuration parameters. The layer-by-layer expansion process is to add the first layer of adjacent devices to the set, then add the second layer of adjacent devices to the set, until the preset number of hops is reached or no new devices are added. The final set of device entity nodes is the target device range.

[0073] For each target abnormal event, a "target time window" is generated: the target time window is obtained by extending forward by the time of occurrence of the target abnormal event and backward by the time of occurrence of the target abnormal event as specified by the system configuration parameters; the target time window is bound to the working condition label of the target abnormal event for subsequent filtering of relationship edges according to the same working condition.

[0074] Output "Target Abnormal Event Input", which includes at least the event identifier, global device identifier, target device range, target time window and operating condition label, and serves as the set of input parameters extracted from the S240 subgraph.

[0075] S240, Anomaly Candidate Subgraph Extraction: Using the target anomaly event as input as a parameter, subgraph extraction is performed from the time-series operating condition constraint knowledge graph. The subgraph extraction starts from the event node corresponding to the target anomaly event and limits the extraction results to only cover nodes and edges within the target device range.

[0076] Perform "Time Intersection Filtering": Read the valid time period of each candidate relationship edge and calculate whether there is an intersection with the target time window; the intersection determination rule is fixed as follows: when the valid start time of the relationship edge is less than or equal to the end time of the target time window and the valid end time of the relationship edge is empty or greater than or equal to the start time of the target time window, it is determined that there is an intersection and the relationship edge is retained; otherwise, the relationship edge is removed.

[0077] Perform "Working Condition Include Filtering": When the set of applicable working condition labels for a candidate relationship edge is empty, retain it directly; when the set of applicable working condition labels for a candidate relationship edge is not empty, retain the relationship edge only if the working condition label of the target abnormal event belongs to the set of applicable working condition labels, otherwise remove the relationship edge.

[0078] Perform "range filtering": retain only the relationship edges of endpoint nodes that belong to device entity nodes within the target device range; for event nodes and measurement point nodes, require that their associated global device identifier belongs to the target device range, otherwise remove the node and its related edges.

[0079] Perform "connectivity expansion": On the set of nodes and edges that have passed the above filtering, start from the starting node and perform a layer-by-layer traversal expansion along the topological connection relationship edges and control dependency relationship edges, with the traversal depth not exceeding the expansion depth specified by the system configuration parameters; on the basis of the connected structure obtained by the priority expansion, supplement the material or energy flow relationship edges and timing trigger relationship edges within the same connected structure to ensure that the extracted subgraph has an interpretable main connected structure as its skeleton and to suppress isolated nodes.

[0080] Output an “anomaly candidate subgraph”, which contains at least a set of nodes and a set of edges. For each edge, retain the edge type, relation source, generation rule version and initial confidence value for subsequent evidence chain scoring and write-back updates.

[0081] S3. Apply reachability, temporal consistency, and operational condition consistency constraints to the anomaly candidate subgraph to obtain the source inference space, generate candidate root cause hypotheses and extract evidence chains, and construct an observable response description based on the evidence chains; specifically including the following sub-steps:

[0082] S310, Limitation of the source reasoning space: A "source reasoning space" is created with the abnormal candidate subgraph and the target abnormal event input as inputs. The source reasoning space refers to the inferable subgraph version obtained after constraint filtering on the abnormal candidate subgraph, and this inferable subgraph version serves as the sole input for subsequent candidate root cause hypothesis generation and evidence chain extraction.

[0083] Define a “set of allowed relation edge types”, which refers to a whitelist of edge types in the source reasoning space that are allowed to participate in traversal and reachability determination. The whitelist includes at least topological connection relation edges, control dependency relation edges, material or energy flow relation edges, and time-triggered relation edges. And specify the direction rules: material or energy flow relation edges are only allowed to be traversed along their defined flow direction, and time-triggered relation edges are only allowed to be traversed from the direction of the first event to the direction of the second event.

[0084] Execute "topology reachability constraints": Using the event node corresponding to the target anomaly event as the endpoint node, perform reachability determination on all nodes in the anomaly candidate subgraph; the reachability determination is implemented using constraint-type traversal, during which only the set of allowed relation edge types is allowed and the direction rules are followed, and the maximum traversal depth is set to the upper limit of the inference depth specified by the system configuration parameters; nodes that are unreachable endpoint nodes and their associated edges are removed as a whole.

[0085] Execute "Time Sequence Constraints": Generate a "Node Reference Time" for each node in the candidate subgraph of anomalies, where the node reference time refers to the time base used for time sequence comparison; the node reference time of an event node is taken as the time of its event occurrence; the node reference times of the measurement point node and the device entity node are generated according to the time of the first appearance of the anomaly sign as defined in S320; when a node cannot obtain a node reference time, the node is marked as a time undeterminable node and removed from the source inference space.

[0086] "Time reversal" is defined as: in any path there exists a relation edge whose starting node reference time is later than the ending node reference time, or whose ending node reference time is later than the time when the target abnormal event occurs; paths containing time reversal are eliminated, and nodes and edges that can only be reached through time reversal paths are further eliminated.

[0087] Execute "Consistency Constraints": For each relation edge in the source inference space, read its applicable condition label set. If the applicable condition label set is empty, retain the edge. If the applicable condition label set is not empty, retain the edge only if the condition label of the target abnormal event belongs to the set, otherwise remove the edge. After removal, perform connectivity pruning on the resulting unreachable structure and output the source inference space.

[0088] S320, Candidate Root Cause Hypothesis Generation: Construct a "Candidate Cause Pool" in the source reasoning space, where the candidate cause pool refers to the set of nodes that may serve as abnormal causes, including at least device entity nodes, event nodes, and measurement point nodes.

[0089] "Abnormal signs" are defined as detectable abnormal signals that appear at a node within the target time window: For event nodes, an abnormal sign is that the event node exists and its event occurs earlier than the target abnormal event; for measurement point nodes, an abnormal sign is that the measurement point has a measurement point over-limit event, or the measurement point has a continuous range of values ​​that deviate from the baseline within the target time window; for equipment entity nodes, an abnormal sign is that the equipment is associated with at least one alarm event, protection action event, maintenance event, or defect event, and these events occur earlier than the target abnormal event.

[0090] Define a "baseline window" and fix the criteria for judging continuous deviation from the baseline: The baseline window refers to a continuous time segment before the occurrence of the target abnormal event, and its length is the system configuration parameter; calculate the baseline mean and baseline fluctuation level within the baseline window; when the deviation of the measurement point value from the baseline mean continuously exceeds the deviation threshold specified by the system configuration parameter and the duration is not less than the duration threshold specified by the system configuration parameter, it is determined that the measurement point has a continuous deviation from the baseline, and the grid point time that first meets the continuous condition is taken as the first time that the abnormal sign of the measurement point appears.

[0091] Candidate root cause hypotheses are generated based on anomaly signs: For each node in the candidate pool of cause points, if it meets the definition of anomaly signs and its node reference time is earlier than the time of occurrence of the target anomaly event, a candidate root cause hypothesis is generated; where the candidate root cause hypothesis refers to a structured hypothesis record that a certain cause point may lead to the target anomaly event, and at least includes the cause point identifier, cause point type and node reference time.

[0092] For each candidate root cause hypothesis, generate a "set of associated paths": starting from the cause point and ending at the target abnormal event, perform a finite-depth constrained traversal in the source reasoning space to obtain the set of associated paths; the finite depth is limited by the maximum associated path length threshold specified by the system configuration parameters, and each associated path must contain at least one event node and at least one measurement point node or measurement point exceeding the limit event node.

[0093] Calculate and normalize the "initial root cause score": For each candidate root cause hypothesis, calculate the number of paths, path length, edge confidence, and temporal proximity based on its associated path set; deduct points for paths whose path length exceeds the long path threshold specified by the system configuration parameters according to the attenuation rules specified by the system configuration parameters; perform fixed normalization processing on each item within the same target abnormal event input range; synthesize the normalized items into the initial root cause score according to the weights specified by the system configuration parameters.

[0094] Output a set of candidate root cause hypotheses. The set of candidate root cause hypotheses should include at least the cause point identifier, cause point type, node reference time, initial root cause score and related path set reference, and write the system configuration parameter version number to ensure the traceability of the scoring criteria.

[0095] S330, Evidence Chain Extraction: Define an "evidence chain" as a path instance that connects the candidate root cause hypothesis cause point with the target abnormal event. The evidence chain consists of a node sequence and a relation edge sequence, and the path instance is required to satisfy the allowed relation edge type set, direction rules, temporal sequence constraints and working condition consistency constraints.

[0096] Evidence chain extraction is performed on each candidate root cause hypothesis set: starting from the cause point and ending with the target abnormal event, a constraint-based traversal is used to generate candidate evidence chains in the source reasoning space; the constraint-based traversal uses a traversal queue to store the current node, node sequence, and relation edge sequence, and performs time validity, working condition consistency, direction legality, and loop inhibition checks on each expansion; the loop inhibition rule is fixed to reject expansion if a node that already exists in the node sequence is introduced.

[0097] The evidence chain candidates are required to meet the inclusion condition: the node sequence must contain at least one test point node or a test point exceeding the limit event node, and at least one event node; and a maximum path length threshold specified by the system configuration parameters is set, and paths exceeding the maximum path length threshold are not included in the evidence chain candidate pool.

[0098] Evidence chain candidate pool is subjected to evidence chain scoring, deduplication, and long path suppression: the evidence chain score is synthesized by edge confidence, path length, and temporal proximity components according to the synthesis rules specified by the system configuration parameters; deduplication uses an "evidence chain deduplication key", which is generated by the node identifier sequence and the relation edge type sequence according to a fixed splicing rule; for evidence chains with the same deduplication key, only the one with the higher evidence chain score is retained; when the path length exceeds the long path threshold specified by the system configuration parameters, the evidence chain score is deducted according to the attenuation coefficient specified by the system configuration parameters, and the deduction rule is fixed to deduct once for each step exceeding the threshold.

[0099] The processed evidence chain is written into the evidence chain storage and a unique evidence chain identifier is generated, so that the evidence chain identifier is linked to the candidate root cause hypothesis in an indexable way, which can be used for subsequent construction of observable response descriptions and consistency comparison.

[0100] S340. Construction of Observable Response Description: Define "observable response description" as a structured description of "what changes should occur at which measurement points within the verification time window if a candidate root cause hypothesis is true". The structured description shall include at least the target measurement point set, the direction of change of each measurement point, the range of change amplitude, and the order of prediction linkage.

[0101] For each candidate root cause hypothesis, a "target measurement point set candidate pool" is constructed: measurement point nodes that have appeared are extracted from the evidence chain set associated with the candidate root cause hypothesis and merged and deduplicated to obtain a measurement point candidate list; availability verification is performed on each measurement point, which includes at least the existence of valid data, completion of time alignment, and completion of field standardization; measurement points that fail the verification are removed; data integrity is calculated for measurement points that pass the verification, which is the ratio of the number of valid sampling points to the number of theoretical sampling points, and measurement points that are lower than the lower limit specified by the system configuration parameters are removed.

[0102] Calculate "Sound Point Health" and filter out bad points: The sound point health is synthesized by the missing rate, dead rate and noise level according to the synthesis rules specified by the system configuration parameters; the missing rate is the proportion of invalid sampling points; the dead rate is the proportion of multiple consecutive points with the same value; the noise level is the statistical value of the fluctuation amplitude within the baseline window, and the baseline window follows the definition of S320; sound points with sound point health below the lower limit specified by the system configuration parameters are removed.

[0103] Calculate the "measurement point contribution" and determine the target measurement point set: The measurement point contribution is synthesized by at least the frequency of the evidence chain occurrence, the shortest reachable path length to the target anomaly event, and the degree of deviation from the baseline according to the synthesis rules specified by the system configuration parameters; the degree of deviation from the baseline is generated according to the relative degree of the maximum deviation of the measurement point within the target time window and the baseline fluctuation level, and the baseline window follows the definition of S320; select the number of measurement points specified by the system configuration parameters as the target measurement point set according to the measurement point contribution from high to low, and save the target measurement point set as the measurement point set number for subsequent strategy optimization to switch within the discrete number range.

[0104] Generate the direction and range of change: The direction of change is primarily determined based on whether there is information on target value adjustment or reduction in the control command event in the evidence chain. An upward adjustment is mapped to an upward direction, and a downward adjustment is mapped to a downward direction. When target value information is missing, the direction of change is determined by the historical trend rule of the same working condition, and a degradation mark is recorded. The range of change is primarily generated according to the quantile of historical samples of the same working condition. When historical samples are insufficient, the range of change is generated by adding or subtracting a multiple of the baseline fluctuation level specified by the system configuration parameters to the baseline mean, and a degradation mark is recorded.

[0105] Generate the predicted linkage sequence: Calculate the shortest reachable path length from the cause point to each measuring point, with shorter path lengths ranked higher; if path lengths are the same, rank them according to the order in which the measuring point first appears in the evidence chain; if still the same, rank them according to the upstream to downstream order of material or energy flow relationship edges; write the change direction, change range, and linkage sequence number into the observable response description, and output the candidate root cause hypothesis-observable response description mapping table as input for subsequent verification strategy space construction.

[0106] S4. Construct a verification strategy space and define consistency score, discriminant, and verification cost. Use the fruit fly algorithm to optimize the verification strategy to obtain the optimal verification strategy. Under the optimal verification strategy, compare the predicted observable response with the actual observable response to generate a consistency score and perform proof by contradiction elimination to converge the root cause hypothesis set. Specifically, this includes the following sub-steps:

[0107] S410, Validation Strategy Space Construction: Define "Validation Strategy" as a structured configuration of a validation execution plan; the validation strategy shall at least include the candidate root cause hypothesis validation order, validation time window number, measurement point set number and validation model number; wherein the validation time window number and measurement point set number are discrete value fields, used to ensure that the strategy space is enumerable and reproducible.

[0108] Generate a "Verification Time Window Number Set": Using the target time window in the target abnormal event input as the parent window, divide the parent window into multiple sub-windows according to the sub-window length and sliding step size specified by the system configuration parameters; generate a verification time window number for each sub-window; when the number of sub-windows after division is zero, degenerate to using the parent window as the only sub-window and generating a unique verification time window number, and record the degeneration mark.

[0109] Generate a "Set of Measurement Point Numbers": For each candidate root cause hypothesis, read one or more measurement point set numbers corresponding to that candidate root cause hypothesis from the Candidate Root Cause Hypothesis-Observable Response Description Mapping Table; the set of measurement point set numbers is only allowed to be taken from the output range of this mapping table.

[0110] Define the "Verification Model Number Set": Register available verification models as the verification model number set; available verification models must include at least digital twin models or rapid mechanism simulation models; when a model lacks necessary input fields or the model health check fails, mark the model as unavailable and remove it from the verification model number set.

[0111] Constructing a "verification order candidate set": First, obtain a basic sorting from high to low based on the initial root cause scores; then, generate a limited number of order variants based on the basic sorting according to the upper limit of the number of adjacent swaps specified by the system configuration parameters, forming a verification order candidate set; if the number of candidate root cause hypotheses is one, the verification order candidate set only contains that single-element order.

[0112] Construct a "validation strategy space index": For each candidate root cause hypothesis, generate a strategy entry by combining the candidate set of validation order, the set of validation time window numbers, the set of measurement point numbers, and the set of validation model numbers, and write a unique strategy identifier for each strategy entry; the validation strategy space index supports enumerating strategies by candidate root cause hypothesis and supports filtering unexecutable strategies by feasibility judgment.

[0113] Perform "Feasibility Determination": For each strategy entry, check whether there is valid data in the time-aligned sequence of the target measurement point set within its verification time window, whether the data integrity is not lower than the data integrity lower limit specified by the system configuration parameters, and whether the verification model is in a usable state; if any check fails, the strategy entry is marked as an unusable strategy and written to the unusable reason field; only usable strategy entries are output to enter the subsequent strategy optimization process.

[0114] S420, Strategy Optimization Objective Definition: Define "Consistency Score" as the score for the degree of matching between the predicted observable response and the actual observable response under a certain verification strategy for a certain candidate root cause hypothesis; The calculation method of consistency score is fixed as the synthesis of the sub-scores of the directional consistency check, amplitude interval consistency check and sequential consistency check in S440 according to the weights specified by the system configuration parameters.

[0115] Define the calling method for the optimization phase: In the strategy optimization phase, for each available verification strategy, calculate the consistency score of each candidate root cause hypothesis according to the fixed caliber of S440, and use it as input for the discrimination calculation to ensure that the caliber of the strategy optimization phase is consistent with that of the final verification phase.

[0116] The "discrimination index" is defined as the degree of separability of the consistency score distribution of a certain verification strategy for the candidate root cause hypothesis set. The discrimination index calculation is performed in a fixed order: first, calculate the consistency score for each candidate root cause hypothesis under the verification strategy; sort the consistency scores from high to low, and take the difference between the first and second consistency scores as the first two differences; calculate the absolute difference between the consistency scores of any two candidate root cause hypotheses and sum them up, then divide by the number of pairwise combinations to obtain the average difference; combine the first two differences and the average difference according to the synthesis rules specified by the system configuration parameters to form the discrimination index.

[0117] The "verification cost" is defined as the resource consumption for executing the verification strategy. The verification cost is at least composed of the number of model calls, simulation duration, number of measurement points read, and concurrent usage according to the synthesis rules specified by the system configuration parameters. The number of measurement points read is calculated by multiplying the number of target measurement points by the number of time grid points within the verification time window. The synthesis rules satisfy the monotonicity constraint, that is, if any one of them increases, the verification cost will not decrease.

[0118] Define "computational resource constraints" as a set of resource limits, including at least the maximum number of model calls, the maximum total simulation time, the maximum number of measurement points read, and the maximum concurrent usage. When the resource consumption of any verification strategy exceeds the corresponding limit, the verification strategy is determined to be an unusable strategy, and in subsequent optimizations, it is assigned the minimum fitness and is not allowed to be output as the optimal strategy.

[0119] The "fitness comparison rule" is defined as a lexicographical comparison rule: under the premise of satisfying computational resource constraints, priority is given to selecting the one with higher discrimination; when the discrimination is the same, the one with lower verification cost is selected; when both discrimination and verification cost are the same, the one with fewer target measurement points and shorter simulation time is selected; the comparison rule serves as the sole fitness evaluation basis for the subsequent fruit fly algorithm.

[0120] S430, Fruit Fly Algorithm Strategy Optimization: Encode the verification strategy as a "strategy configuration object", where the strategy configuration object refers to an executable configuration record composed of discrete fields, including at least the verification order field, the verification time window number field, the measurement point set number field, and the verification model number field; and stipulate that the parsing rules from the strategy configuration object to the verification strategy are fixed.

[0121] Define "Discrete Disturbance Rules" as a set of rules that allow only limited discrete changes to the strategy configuration object: the verification time window number field can only switch between adjacent numbers and must not exceed the boundary; the measurement point set number field can only switch within the range of numbers output by the candidate root cause hypothesis-observable response description mapping table; the verification order field can only allow adjacent exchanges or local segment reversals and the number of exchanges cannot exceed the upper limit specified by the system configuration parameters; the verification model number field can only switch within the set of verification model numbers; the maximum number of fields that can be modified per round of disturbance cannot exceed the upper limit of field modifications specified by the system configuration parameters.

[0122] Define "Drosophila population" as a set of policy configuration objects; when initializing the Drosophila population, use a random seed obtained by fixed encoding of the event identifier in the target abnormal event input to drive pseudo-random sampling to generate the initial set of policy configuration objects; where the random seed refers to the fixed input value used to drive pseudo-random sampling, and the pseudo-random sampling algorithm and the method of using the random seed remain unchanged to ensure that the initialization results can be reproduced under the same input.

[0123] For each policy configuration object in the fruit fly population, perform feasibility determination and fitness calculation: First, check the availability of its policy entries according to S410 feasibility determination; if unavailable, directly assign the minimum fitness and write the reason for unavailability; if available, calculate the consistency score set, discrimination, and verification cost according to S420, and obtain the fitness ranking basis according to the fitness comparison rules; write each calculation result to the "optimization audit log", in which the optimization audit log records at least the field values ​​of the policy configuration object, discrimination, verification cost, resource constraint check results, and final ranking basis.

[0124] Perform iterative updates: Using the current best policy configuration object as the center, generate a new set of policy configuration objects according to the discrete perturbation rule; calculate the fitness of each policy configuration object in the new set of policy configuration objects; if there is a policy configuration object that is better than the current best, replace the current best; repeat the iteration until the termination condition is met.

[0125] The termination condition is defined as reaching the maximum number of iterations specified by the system configuration parameters or reaching the upper limit specified by the system configuration parameters for the number of consecutive iterations without improvement; after termination, the optimal strategy configuration object is output and parsed into the optimal verification strategy; the output process is to optimize the verification strategy space, and the strategy optimization method is explicitly defined as the fruit fly algorithm.

[0126] S440, Prediction-Actual Consistency Comparison and Counter-Evidence Elimination: Generating Predicted and Actual Observable Responses: For each candidate root cause hypothesis under the optimal verification strategy, the verification model is invoked to output the predicted observable response; the predicted observable response includes at least the target measurement point set, the predicted change direction for each measurement point, the predicted change range, and the predicted linkage sequence; the time segment corresponding to the same verification time window number is extracted from the time-aligned record to form the actual observable response; the time-aligned sequence extracted from this time segment is used for single measurement point scanning and continuous determination.

[0127] The baseline window positioning rules are retained and expanded: the baseline window definition follows S320, which is a continuous time segment used to calculate the baseline mean and baseline fluctuation level, with the length being a system configuration parameter; during the validation phase, when the predicted observable response gives the predicted start time of change, the continuous time segment before the predicted start time of change is used as the baseline window; when the predicted start time of change is missing, it degenerates to using the time segment that most recently meets the baseline window length requirement before the start of the validation time window as the baseline window, and a degradation marker is recorded.

[0128] Extracting the actual change time and actual change magnitude: For each target measurement point, first calculate the baseline mean and baseline fluctuation level within the baseline window; then scan the time alignment sequence of the measurement point grid by grid within the verification time window to find the first significant deviation from the baseline. The significant deviation judgment rule is that the deviation magnitude continuously exceeds the deviation threshold specified by the system configuration parameters and reaches the duration threshold specified by the system configuration parameters; the difference between the mean in the neighborhood of the change time and the baseline mean is defined as the actual change magnitude, and the neighborhood length is the system configuration parameter; if no significant deviation occurs within the verification time window, the measurement point is marked as not triggered and the actual change magnitude is recorded as zero.

[0129] Perform a direction consistency check: determine whether the signs of the predicted change direction and the actual change magnitude are consistent or inconsistent according to a fixed matching rule; if the prediction is to remain unchanged, the direction is considered consistent if the absolute value of the actual change magnitude does not exceed the deviation threshold, otherwise the direction is considered inconsistent; the direction score is determined by the system configuration parameters and remains unchanged throughout the process.

[0130] Perform a consistency check on the amplitude range: determine whether the actual change amplitude falls within the predicted change amplitude range. If it does, it is considered consistent; otherwise, it is considered inconsistent. When the predicted change amplitude range is missing, it is treated as inconsistent and a missing mark is recorded. The amplitude score value is determined by the system configuration parameters and remains unchanged throughout the process.

[0131] Perform a sequence consistency check: Sort the target measurement point set according to the actual change time to obtain the actual linkage sequence, and compare it with the predicted linkage sequence; when there are untriggered measurement points, place the untriggered measurement points at the end of the sequence and deduct points according to the number of untriggered points; when there is a reversed sequence, deduct points according to the logarithm of the reverse; the deduction rules are system configuration parameters and satisfy monotonicity constraints.

[0132] Synthesize a consistency score and perform disprovenance elimination: Combine the direction score, magnitude score, and ordinal score with the weights specified by the system configuration parameters to synthesize a consistency score, with the sum of the weights always being one; generate a consistency threshold based on the quantiles specified by the system configuration parameters of the consistency score distribution of historically confirmed root cause samples under the same working condition label. If there are insufficient historical samples, degenerate to the system default threshold and record the degeneration mark; candidate root cause hypotheses with a consistency score lower than the consistency threshold are judged as disprovenance failures and eliminated; candidate root cause hypotheses with a consistency score not lower than the consistency threshold but lower than the high confidence threshold specified by the system configuration parameters are subject to weight reduction and retained; output the converged root cause hypothesis set and write sub-scores, missing value marks, and degeneration marks for each retained hypothesis for auditing.

[0133] S5. Determine the target root cause based on the converged root cause hypothesis set and output the anomaly tracing results; update the confidence of the knowledge graph relation edges based on the evidence chain corresponding to the target root cause and write back the version; and construct an active learning task queue based on information gain scoring to achieve self-updating of the knowledge graph. Specifically, this includes the following sub-steps:

[0134] S510, Target Root Cause Determination: Using the converged root cause hypothesis set output from S440 as input, construct a "candidate ranking list"; where the candidate ranking list refers to the list obtained by sorting the consistency scores from high to low; for each candidate root cause hypothesis, simultaneously read the consistency score, direction score, magnitude score, order score, evidence chain identifier set, as well as the cumulative confidence value of the evidence chain edge and the evidence chain length.

[0135] Calculate the "first two difference", where the first two difference refers to the difference between the consistency score of the first candidate and the consistency score of the second candidate in the candidate ranking list; when there are fewer than two candidates, directly identify the unique candidate as the target root cause and skip the subsequent difference judgment.

[0136] When the difference between the first two values ​​is not less than the difference threshold specified by the system configuration parameters, the first candidate in the ranking list is determined as the target root cause. The source of the difference threshold is fixed as the quantile of the distribution of the first two differences of historically confirmed root cause samples under the same working conditions, as specified by the system configuration parameters. When there are insufficient historical samples, the difference threshold is reduced to the system default difference threshold and a degradation mark is recorded.

[0137] When the difference between the first two values ​​is less than the difference threshold, the "adjudication chain rule" is executed to determine the target root cause. The adjudication chain rule is a fixed order comparison rule, which is as follows: the one with the higher order score takes priority; if the order scores are the same, the one with the higher amplitude score takes priority; if the amplitude scores are the same, the one with the higher direction score takes priority; if they are still the same, the one with the shorter evidence chain length takes priority; if they are still the same, the one with the higher cumulative confidence value of the evidence chain edge takes priority; if they are still the same, the one with the higher initial root cause score takes priority; if they still cannot be distinguished, the one with the smaller lexicographical order of the candidate root cause hypothesis identifier is selected to ensure deterministic output.

[0138] Generate an "adjudication path record", which is a structured log that records the process of determining the root cause of the target. It includes at least the first two differences, whether the difference threshold was directly determined, the triggered adjudication chain rule level, the comparison input values ​​of each level, the degradation flag and the final root cause identifier. Write the adjudication path record to the result storage for auditing.

[0139] S520, Anomaly Tracing Result Output: Taking the target root cause determination result output by S510 as input, read the set of evidence chain identifiers corresponding to the target root cause from the evidence chain storage, and obtain the node sequence and relation edge sequence of each evidence chain according to the evidence chain identifier.

[0140] Read the observable response description corresponding to the target root cause from the candidate root cause hypothesis-observable response description mapping table; read the actual observable response fragments and sub-item score details corresponding to the target root cause from the consistency comparison results.

[0141] Generate an "anomaly tracing result object", where the anomaly tracing result object refers to the structured result record used for output and storage; the anomaly tracing result object shall at least include: target anomaly event identifier, target root cause identifier, final consistency score, direction score, magnitude score, sequence score, verification time window number, target measurement point set, prediction and actual comparison summary of key measurement points, evidence chain identifier set, and adjudication path record reference.

[0142] Output the anomaly tracing results to the human-machine interface or linkage interface, and write them to the result storage in a structured format; if the output interface fails, generate an output failure record and retain the number of retries and the reason for failure to avoid loss of results due to output failure.

[0143] S530, Source Knowledge Back-Write: Define "back-write object" as a set of knowledge graph elements that need to be updated. The back-write object includes at least: relation edges on the evidence chain, association edges between candidate root cause hypotheses and target abnormal events, and rule parameters used to generate time-series triggering relation edges; where rule parameters refer to configuration items used to generate time-series triggering relation edges from event sequences.

[0144] Generate an "edge update request list": For each relation edge in the target root cause evidence chain, generate an update request to increase confidence; for each relation edge in the key evidence chain corresponding to the candidate root cause hypothesis that was eliminated by disproven evidence, generate an update request to decrease confidence; each update request includes at least the edge identifier, update direction, update step size parameter, working condition label, and tracing task identifier; where the update step size parameter is a system configuration parameter.

[0145] Perform update magnitude calculation and apply upper and lower limits: increase or decrease the edge confidence according to the update step size parameter, and limit the confidence between the minimum and maximum confidence specified by the system configuration parameters; if the confidence exceeds the upper or lower limits after increase or decrease, it will be truncated according to the upper or lower limits to ensure that the confidence is always within the legal range.

[0146] Perform same-edge merge update: When there are multiple update requests with the same edge identifier in the edge update request list, merge them first according to the update direction; after the update magnitude of the requests in the same direction is accumulated, perform a one-time update; when there is a conflict between promotion and demotion, adjudicate according to the fixed rule that promotion of the target root cause evidence chain takes precedence over demotion of the elimination chain, and record the conflict adjudication in the graph change log to ensure that the update result of the same edge is uniquely determined by a single tracing task.

[0147] Perform incremental write-back and versioning: update attributes only for the edges and nodes involved in this write-back object; for relationship edges that need to be replaced due to topological or logical error correction, generate replacement edges in a versioned manner: write the old edge to the valid end time, and write the new edge version to the valid start time as the write-back time to avoid double validity at the same time point.

[0148] Write to the "Graph Change Log": The graph change log should include at least the value before the change, the value after the change, the reason for the change, the direction of the update, the magnitude of the update, the working condition label, the traceability task identifier, and the reference to the associated abnormal traceability result object; output the updated time-series working condition constraint knowledge graph version number as the only input version number for subsequent processes.

[0149] S540, Active Learning and Update: Constructing a "Candidate Triple Pool for Verification": Adding triples on the target root cause evidence chain and triples on the key evidence chains of eliminated candidates to the candidate triple pool; where triples refer to edge instances of head nodes, relations, and tail nodes in the knowledge graph; for each triple, record its triple identifier, occurrence count, corresponding working condition label, corresponding tracing task identifier, and the identifier of the evidence chain in which it belongs.

[0150] Calculate "Coverage Count": The coverage count is defined as the number of times the triple is referenced by the evidence chain in the most recent source tracing tasks specified by the system configuration parameters; if the number of source tracing tasks is insufficient, it is calculated based on the actual number.

[0151] Calculate the "uncertainty": The uncertainty is calculated by how close the current confidence level of the triplet is to the intermediate confidence level specified by the system configuration parameters. The uncertainty calculation rule is fixed as the closer the confidence level is to the intermediate confidence level, the higher the uncertainty. The uncertainty is then normalized to the range specified by the system configuration parameters to ensure comparability.

[0152] Calculate "Impact": Impact is defined as the potential increase in discrimination of the strategy optimization objective when the confidence of the triple is increased or decreased by a specified range according to the system configuration parameters. Impact estimation is performed according to fixed simulation rules: keeping the confidence of other edges unchanged, only the confidence of the triple is increased or decreased by a preset range, and the consistency score and discrimination of the candidate root cause hypothesis related to the triple are recalculated. The maximum increase in discrimination is taken as the impact. The recalculation range and the preset range are both system configuration parameters.

[0153] Calculate the "information gain score": Combine the coverage count, uncertainty, and impact into an information gain score according to the synthesis rules specified in the system configuration parameters, and satisfy the monotonicity constraint; Sort the information gain scores from high to low and select the number of triples specified in the system configuration parameters to enter the active learning task queue.

[0154] Specify the "verification action type" and add it to the queue: The verification action type should at least include expert confirmation, supplementary data collection, and mandatory inclusion in the measurement point set for the next strategy optimization; The action selection rules are fixed as follows: when the uncertainty and impact are high, expert confirmation should be given priority; when the coverage is high but the data integrity is low, supplementary data collection should be given priority; when the impact is high and the health of the measurement point that the observable response depends on can be improved, mandatory inclusion in the measurement point set should be given priority; Write the triplet identifier, action type, priority, and reason into the active learning task queue.

[0155] Verification result write-back and versioning: For triples that are verified correctly, increase the confidence of their relation edges and solidify the applicable working condition label set; for triples that are verified incorrectly, set the relation edges to invalid and create a new version of the replacement relation edges, writing the new valid time period and source into the replacement relation edges; all write-backs are written to the knowledge graph change log and include the verification basis, action type, working condition label and corresponding tracing task reference; the updated knowledge graph version number is used as the only input version number for subsequent anomaly tracing processes to achieve traceable incremental self-updating.

[0156] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0157] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for tracing the source of anomalies in intelligent power equipment, characterized in that, Includes the following steps: S1. Acquire multi-source operation data of the power plant, perform unified identification mapping and field standardization on the data, and perform time alignment and operation condition label generation. Generate structured event data according to event extraction rules. S2. Construct a knowledge graph based on structured event data. The knowledge graph includes equipment entity nodes, measurement point nodes, event nodes and their relationship edges, and set a set of valid time periods and applicable working condition labels for the relationship edges. Filter target abnormal events, generate target equipment range and target time window, and extract abnormal candidate subgraphs in the knowledge graph by time intersection, working condition inclusion and range filtering; S3. Perform reachability, timing and operating condition consistency constraints on the abnormal candidate subgraph to obtain the source reasoning space, generate candidate root cause hypotheses and extract evidence chains, and construct observable response descriptions based on the evidence chains. S4. Construct a verification strategy space and define consistency score, discrimination, and verification cost. Use the fruit fly algorithm to optimize the verification strategy to obtain the optimal verification strategy. Under the optimal verification strategy, compare the predicted observable response with the actual observable response to generate a consistency score and perform proof by contradiction elimination to converge the root cause hypothesis set.

2. The method for tracing the source of anomalies in intelligent power equipment according to claim 1, characterized in that, Also includes: S5. Determine the target root cause based on the converged root cause hypothesis set and output the anomaly tracing results; The confidence of the knowledge graph relation edges is updated and versioned back based on the evidence chain corresponding to the target root cause. Furthermore, an active learning task queue is constructed based on information gain scoring to achieve self-updating of the knowledge graph.

3. The method for tracing the source of anomalies in intelligent power equipment according to claim 1, characterized in that, S1 specifically includes: Configure multi-source connections and acquire real-time measurement points, alarms and protection actions, control commands and operation tickets, maintenance work orders and equipment ledger data, generate original records and generate acquisition failure records when acquisition fails; Perform a unified mapping from the original identifier to the global device identifier and the global measurement point identifier on the original record, and perform unit completion, unit conversion and valid range verification according to the field standardization rule table to generate standardized records; Establish a unified time base and generate a time grid, align standardized records to generate time-aligned records and time-aligned sequences, and generate working condition labels for time-aligned records; Based on the event type table, perform event-based extraction on time-aligned records and event-type data, continuously determine measurement point over-limit events, establish event deduplication keys, generate structured event data, and output event identifiers.

4. The method for tracing the source of anomalies in intelligent power equipment according to claim 1, characterized in that, S2 specifically includes: Construct a knowledge graph storage and write device entity nodes, measurement point nodes and event nodes based on node deduplication keys. Establish topological connections, control dependencies, material or energy flow directions and time-series triggering relationship edges, and write initial confidence values ​​for relationship edges to form an initial knowledge graph. Configure valid time periods for relation edges and perform version updates of relation edges during maintenance or topology changes. Configure applicable working condition label sets for relation edges to form a time-series working condition constraint knowledge graph. Target abnormal events are filtered based on abnormal triggering conditions, and target device range and target time window are generated to form target abnormal event input.

5. The method for tracing the source of anomalies in intelligent power equipment according to claim 4, characterized in that, In the time-series operating condition constraint knowledge graph, subgraph extraction is performed starting from the target abnormal event. Time intersection filtering, operating condition inclusion filtering and range filtering are performed in sequence, and candidate abnormal subgraphs are obtained by connectivity expansion.

6. The method for tracing the source of anomalies in intelligent power equipment according to claim 1, characterized in that, S3 specifically includes: On the candidate subgraph of anomalies, the types of allowed relation edges are restricted and topological reachability constraints are executed according to the direction rules. Node reference times are generated and temporal sequence constraints and working condition consistency constraints are applied to obtain the source inference space. Within the source tracing reasoning space, a candidate pool of cause points is constructed. Based on abnormal signs, candidate root cause hypotheses are generated and a set of associated paths is formed. The initial root cause score is calculated and normalized. Starting with candidate root cause hypotheses and ending with target anomalous events, evidence chains are extracted. After the inclusion condition is met, scoring, deduplication, and long path suppression are performed to generate evidence chain identifiers. Based on the evidence chain, an observable response description is constructed, the target measurement point set and its change direction, change range and prediction linkage order are determined, and candidate root cause hypotheses and observable response description mapping table are output.

7. The method for tracing the source of anomalies in intelligent power equipment according to claim 1, characterized in that, S4 specifically includes: Based on the candidate root cause hypothesis and the observable response description mapping table, a verification strategy space is constructed, generating a set of verification time window numbers, a set of measurement point numbers, a set of verification model numbers, and a candidate set of verification order, and a feasibility judgment is performed on the strategy items. Define consistency scores, discrimination scores, and verification costs, set computational resource constraints, and determine the merits of strategies according to fitness comparison rules; The verification strategy is encoded as a strategy configuration object and optimized using the fruit fly algorithm. The strategy is iteratively updated based on the discrete perturbation rule and the optimal verification strategy is output.

8. The method for tracing the source of anomalies in intelligent power equipment according to claim 7, characterized in that, Under the optimal verification strategy, predictive and actual observable responses are generated. Based on the baseline window location, the actual change time and magnitude are extracted. The consistency of direction, the consistency of magnitude range and the consistency of sequence are checked in turn. A consistency score is synthesized and the evidence is rejected by contradiction according to the threshold.

9. The method for tracing the source of anomalies in intelligent power equipment according to claim 1, characterized in that, S5 specifically includes: The root cause hypothesis set that converges after elimination by contradiction is used to form a candidate ranking list based on the consistency score. The difference between the first two is calculated and the target root cause is determined when the difference threshold is met, or the target root cause is determined according to the decision chain rule and a decision path record is generated when the difference between the first two is insufficient. Read the evidence chain identifier set and observable response description corresponding to the target root cause, generate an anomaly tracing result object containing sub-scores and predictions, actual comparison summaries, and output and store it in the database; As the evidence chain relationship is generated, the request list is updated. The confidence level is increased or decreased according to the update step parameter and upper and lower limits are applied. Multiple requests on the same side are merged and conflict adjudicated. The graph elements are written back in a versioned manner, written to the graph change log, and the updated knowledge graph version number is output.

10. A method for tracing the source of anomalies in intelligent power equipment according to claim 9, characterized in that, A pool of candidate triples to be verified is constructed, and the coverage, uncertainty and impact are calculated to synthesize the information gain score. High information gain triples are selected to generate an active learning task queue. The verification results are written back and updated in version according to the verification action type.