Cross-period tracing analysis system for supervision rule error data

Abstract

The invention discloses a supervision rule error data cross-period traceability analysis system, which belongs to the field of energy supervision data governance and comprises a data acquisition module, a traceability graph construction module, a periodic error analysis module and a multi-stage early warning processing module. The data acquisition module constructs a compliance tracing data set; the traceability graph construction module integrates the compliance traceability data set, the equipment association logic and the personnel operation association rule, constructs a multi-dimensional weighted association relation table, and generates a cross-cycle association graph and a preliminary error evolution sequence; the periodic error analysis module is used for combining the cross-periodic correlation map and the preliminary error evolution sequence, analyzing frequency change, periodic rules and response characteristics through dynamic trend fitting, and constructing a periodic error trend characteristic library and a periodic repetition error list; and the multi-stage early warning processing module is used for constructing a to-be-early-warned error set, triggering multi-stage early warning and processing the multi-stage early warning in combination with comprehensive early warning grade scores, thereby realizing effective monitoring and management of power plant operation errors.

Description

Technical Field

[0001] This invention relates to a cross-period traceability analysis system for erroneous data in regulatory rules, belonging to the field of energy regulatory data governance technology. Background Technology

[0002] Chinese patent application CN120452449A discloses a power operation supervision method and system based on AI behavior recognition and semantic analysis. The method includes deploying a power system control architecture and data acquisition equipment to continuously collect voiceprint data and video stream data; verifying operator identity through voiceprint recognition and performing semantic analysis on voice commands based on a power standard terminology database to determine the consistency between command content and system-issued tasks; identifying operator positions through visual positioning and spatially matching them with target equipment interval numbers, while simultaneously detecting the standardization of operational behavior through video analysis; generating voice interaction prompts based on command ambiguity; triggering tiered alarms through multimodal data fusion; and simultaneously recording and storing data on violations. This invention improves the efficiency of power system supervision, ensures authorized operations and reduces human error, provides safety through visual monitoring, triggers alarms through multimodal fusion, responds promptly to violations and records evidence, and supports traceable safety management.

[0003] Although existing power operation supervision methods and systems based on AI behavior recognition and semantic analysis have achieved continuous voiceprint and video stream data collection, operator identity voiceprint verification, voice command semantic analysis, operator location and equipment interval spatial matching, operation behavior standardization detection, multimodal data fusion to trigger hierarchical alarms, storage of illegal operation data and generation of audio and video evidence chains before and after, there is a lack of correlation dimensions between the subject and error type of cross-cycle error data, and the problem of periodic recurring errors cannot be identified. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a cross-period traceability analysis system for regulatory rule error data, enabling full-dimensional feature extraction, trend tracing, and hierarchical early warning of periodic error data, accurately locating maintenance-related causes, optimizing maintenance strategies, improving response efficiency and equipment reliability, and enhancing the proactive prevention and control level of power plant operation risks.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The system for cross-period traceability analysis of regulatory rule error data includes: a data acquisition module, a traceability map construction module, a periodic error analysis module, and a multi-level early warning processing module;

[0007] The data acquisition module is used to collect, clean, and analyze raw power plant operation data, label it with four-dimensional tags, and construct a compliance traceability dataset.

[0008] The traceability graph construction module is used to integrate the compliance traceability dataset, equipment association logic and personnel operation association rules to construct a multi-dimensional weighted association table and generate a cross-cycle association graph and a preliminary error evolution sequence.

[0009] The periodic error analysis module is used to combine the cross-period correlation map with the preliminary error evolution sequence, and to analyze frequency changes, periodic patterns and response characteristics through dynamic trend fitting analysis, thereby constructing a periodic error trend feature library and determining a list of periodic recurring errors.

[0010] The multi-level early warning processing module is used to construct a set of errors to be warned, and combined with the comprehensive early warning level score, to trigger multi-level early warnings and process them.

[0011] Specifically, the steps for constructing a multidimensional weighted association table include:

[0012] Obtain the power plant equipment association logic, identify equipment types and extract association conditions to form a triplet of equipment association: equipment A, equipment B, and association conditions;

[0013] Filter time period arrive The complete log is grouped by operator ID and device number, retaining the basic data associated with each operator and device.

[0014] Based on the aforementioned basic data, the temporal co-occurrence frequency of error types within the statistical time window is calculated, and the correlation strength is calculated in conjunction with the total number of operations, and then normalized to the correlation confidence level.

[0015] Based on the three association types—device-to-device, operator-to-device, and operator-to-error—a structured association rule library containing device association triples and association confidence is constructed.

[0016] Specifically, the steps for constructing a multidimensional weighted association table also include:

[0017] Iterate through each compliance traceability record in the aforementioned compliance traceability dataset;

[0018] For the device-to-device association type, the operation timing and state sequence are extracted. Through alignment analysis, the condition compliance is calculated by combining the timing synchronization score and the state response matching score.

[0019] For the operator-equipment and operator-error type association types, the number of matching fields between each compliance traceability record and the structured association rule base is counted. Combined with the total number of matching fields, the association matching degree and behavior matching degree are calculated.

[0020] The overall matching score is calculated by weighting the condition compliance, the association matching score, and the behavior matching score, and an association intermediate table is generated.

[0021] For each intermediate matching record in the intermediate association table, the association weight is calculated by combining the association confidence and the overall matching degree.

[0022] By integrating the core dimensions, association types, and association weights of each intermediate matching record, a multidimensional weighted association table is constructed.

[0023] Specifically, the steps for generating cross-period correlation maps include:

[0024] Based on the multidimensional weighted association table, the operator, equipment, and error type nodes are initialized.

[0025] Directed edges are constructed for the three types of association to form a basic association graph, which is then split into periodic subgraphs.

[0026] Calculate the directed edge weight distribution characteristics of each periodic subgraph, generate a set of subgraphs, and filter out similar node pairs by traversing and comparing adjacent periodic subgraphs.

[0027] If the edge weight of the same type of node pair in the adjacent periodic subgraph is less than the preset weight threshold, it is determined to be a non-cross-period association candidate pair; otherwise, it is determined to be a cross-period association candidate pair. The preceding bridging weight and the following bridging weight are calculated, and the cross-period association strength is obtained by summing them. Bridging edges are generated and an intermediate graph is formed.

[0028] Clusters are divided by passing labels between nodes. Strongly related edges with weights exceeding a preset maximum threshold are filtered out and related paths are retained. Weakly related edges with weights below a preset minimum threshold are filtered out and accidental associations are eliminated.

[0029] Add time dimension labels to the intermediate graph to generate a cross-cycle correlation graph.

[0030] Specifically, the steps for generating the initial error evolution sequence include:

[0031] The association edges and auxiliary features are extracted from the cross-period association graph to generate the original temporal association dataset;

[0032] For the time-series data in the operator ID, device, and error type triplet of operation error, divide the sliding window in ascending order of period;

[0033] Based on the sliding window, normalized numerical features and one-hot encoded category features are concatenated into a fixed-length feature vector to generate a set of time-series feature sequences.

[0034] A self-attention mechanism is introduced to calculate attention weights, and the top n windows are retained as key sequence segments by sorting them in descending order.

[0035] Based on the aforementioned operation error triplet, the key sequence segments are traversed, and target segments with the same operator ID, matching device segments with the same device number, and matching error segments with the same error type are selected respectively. These segments are then assembled in chronological order to generate a preliminary error evolution sequence.

[0036] Specifically, the steps for constructing a periodic error trend feature library include:

[0037] The system constructs primary groups based on equipment type, operator ID, cycle label, and error type, counts error frequency, and extracts equipment maintenance status.

[0038] Based on the first-level grouping, locate the sequence segments in the preliminary error evolution sequence, extract the effective error occurrence timestamp and the effective repair completion timestamp, and calculate the response time;

[0039] Based on the primary grouping, secondary groupings are constructed using equipment type, job position, cycle label, and error type;

[0040] By associating the error frequency, the equipment maintenance status, and the response time, basic characteristic data of periodic errors are formed;

[0041] Using the historical error incidence and fluctuation range of similar equipment and positions as a historical baseline, and associating it with the secondary group, a periodic error basic dataset is generated.

[0042] Specifically, the steps for constructing a periodic error trend feature library also include:

[0043] For each group of equipment type, job position, and error type in the periodic error basic dataset, calculate the year-on-year growth rate of error frequency in adjacent periods in ascending order of period, and obtain the dynamic trend slope through linear fitting;

[0044] If the year-on-year growth rate of the error frequency exceeds the fluctuation range, the corresponding period is marked as an abnormal period; otherwise, it is marked as a normal period.

[0045] Obtain the error frequency periodic sequence, obtain the spectrum through signal decomposition, filter the dominant frequency with the largest amplitude, and calculate the period length by combining the sampling period;

[0046] If the amplitude corresponding to the main frequency exceeds the preset frequency threshold, it is determined that there is a periodic pattern; otherwise, it is determined to be non-periodic.

[0047] Obtain the maintenance execution cycle, combine the frequency before and after maintenance, calculate the frequency decrease rate after maintenance, count the number of errors with response time lower than the preset response time threshold, calculate the response time target rate, and obtain the response efficiency score through weighted calculation;

[0048] Construct frequency variation units, periodic pattern units, and response feature units to store various calculated indicators, and build a periodic error trend feature library.

[0049] Specifically, the steps for identifying a list of periodically recurring errors include:

[0050] Acquire frequency change unit data and filter high-risk candidate errors that have been continuously abnormal for more than m periods;

[0051] If the slope of the dynamic trend is positive and the year-on-year growth rate of the average error frequency exceeds the historical baseline, then the trend is determined to be a deviation from the error; otherwise, it is determined to be an accidental fluctuation.

[0052] The number of consecutive abnormal periods is counted, the trend deviation index is calculated by weighting, and the deviation level is determined by combining the preset trend deviation index threshold to form a preliminary error list.

[0053] Based on the aforementioned periodicity unit, errors exhibiting periodicity in the initial error list are retained. Based on the period length, the pattern matching degree is calculated to determine the repetition pattern, and the list is updated to include repetition errors.

[0054] If the response efficiency score is lower than the preset score value within the corresponding period of delayed maintenance, and the frequency decrease rate after maintenance is negative, then a maintenance-related candidate cause is marked; otherwise, a non-maintenance-related period error is marked.

[0055] The power plant operation and maintenance logs are retrieved to query the erroneous maintenance records in the recurring error list. If maintenance is missing or resources are insufficient, the potential cause is marked; otherwise, it is marked as pending verification. Finally, the periodic recurring error list is determined.

[0056] Specifically, the steps for handling multi-level early warnings include:

[0057] Based on the periodic error trend feature library and the periodic recurring error list, a set of errors to be warned is constructed;

[0058] The deviation level quantification value, repetition pattern intensity quantification value, and potential trigger indicator quantification value are calculated separately, and a comprehensive early warning level score is calculated by weighting. ;

[0059] Set three threshold levels , , ,in The warning level is mapped by combining the set of errors to be warned;

[0060] when When a Level 3 warning is issued, an immediate response is required.

[0061] when At that time, a level-two warning was issued, requiring close monitoring;

[0062] when At that time, a Level 1 warning was issued, indicating a potential risk that requires continuous monitoring.

[0063] when At that time, it was determined that there was no warning.

[0064] Specifically, the steps for handling multi-level early warnings also include:

[0065] Implement differentiated handling strategies based on different warning levels;

[0066] A dynamic feedback mechanism is activated, a re-inspection mechanism is set up, and the number of recurrences and total monitoring times during the execution of the differentiated processing strategy are counted to calculate the error recurrence rate.

[0067] The number of affected devices before and after implementing the differentiated processing strategy is counted, and the reduction ratio of the error impact range is calculated.

[0068] An effectiveness index is obtained by combining the error recurrence rate and the reduction ratio of the error impact scope through weighted calculation. ;

[0069] like If the value exceeds the preset threshold, the warning is deemed effective; otherwise, the warning is deemed ineffective. Common characteristics of ineffective warnings are statistically analyzed, and the weight parameters are dynamically adjusted.

[0070] The beneficial effects of this invention are:

[0071] 1. This invention enables in-depth source tracing and temporal evolution analysis of power plant operation errors by constructing a cross-period correlation map and a multi-dimensional weighted correlation table; by integrating equipment logic, personnel behavior and historical data, it breaks through the limitations of traditional methods that analyze independently by period, effectively identifies cross-period correlation patterns, and improves the interpretability of error causes.

[0072] 2. This invention establishes a periodic error trend feature library and a multi-level early warning mechanism, and combines dynamic trend slope, periodicity, and response characteristics for comprehensive scoring to achieve intelligent early warning of repetitive and trending errors; it evaluates the effectiveness of processing through a closed-loop feedback mechanism, dynamically optimizes weight parameters, enhances the system's adaptability, and effectively supports the hierarchical management and proactive prevention and control of power plant operation risks. Attached Figure Description

[0073] Figure 1 A structural diagram of a system for tracing and analyzing cross-period errors in regulatory rules;

[0074] Figure 2 This is a flowchart of the process for generating cross-period correlation maps in this invention;

[0075] Figure 3 This is a flowchart illustrating the generation of the initial error evolution sequence in this invention;

[0076] Figure 4 This is a flowchart illustrating the construction of a periodic error trend feature library in this invention. Detailed Implementation

[0077] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0078] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0079] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a cross-period traceability analysis system for regulatory rule error data, including a data acquisition module, a traceability map construction module, a periodic error analysis module, and a multi-level early warning processing module;

[0080] The data acquisition module is used to collect raw data from power plant operations. Combined with a pre-set power plant regulatory rule library containing detailed error type standards and periodic regulatory requirements, the raw data is cleaned and semantically parsed. Based on the matching results of error judgment logic and periodic attribution rules in the pre-set power plant regulatory rule library, the data is labeled with four dimensions: operator, equipment, error type, and period, to construct a compliance traceability dataset. The raw data includes, but is not limited to, operator ID, error operation log, timestamp, and equipment number.

[0081] The traceability graph construction module is used to integrate compliance traceability datasets, power plant equipment association logic, and personnel operation association rules to construct a multi-dimensional weighted association table; through graph association analysis, a cross-cycle association graph of the power plant scenario is generated; and combined with sequence aggregation processing, a preliminary error evolution sequence is generated, grouped according to operator, equipment, and error type; among them, the power plant equipment association logic is obtained from the power plant operation procedure text, and the personnel operation association rules are obtained from historical logs;

[0082] Specifically, the steps for constructing a multidimensional weighted association table include:

[0083] The association logic of power plant equipment is extracted from the power plant operation procedure text, the text is preprocessed, the equipment is identified by the BERT-based NER model, and the association conditions are extracted from the preprocessed power plant operation procedure text by dependency parsing to form equipment association triplet of equipment A, equipment B and association conditions.

[0084] Filter time periods from historical logs arrive The complete logs are grouped by operator ID and device number, and each group retains the operation time, error type and device status, which are defined as the basic data associated with operators and devices.

[0085] Based on personnel operation association rules, a time window is set for the retained basic data on operator and equipment associations. Statistics on the time after the operator operates the equipment The frequency of time-series co-occurrence of error types within the device is combined with the total number of operations performed by the same operator on the same device. The ratio of the frequency of time-series co-occurrence to the total number of operations is used to obtain the association strength. The association strength is then normalized and set as the association confidence level.

[0086] Based on the three types of associations—equipment-equipment combinations, operator-equipment combinations, and operator-error type combinations—a structured association rule base is constructed. Each type of rule includes the rule source and core association elements, such as equipment association triples and association confidence.

[0087] The process iterates through each compliance traceability record in the compliance traceability dataset and matches it against a structured association rule base based on association type. This includes: for device-to-device association types, extracting the operation sequence and status sequence of device A and device B in the compliance traceability dataset, performing alignment analysis with preset linkage patterns in the structured association rule base, and calculating the condition compliance degree based on the timing synchronization score and status response matching score; for operator-to-device association types, counting the number of matching fields between operator ID and device number in each record of the compliance traceability dataset and the structured association rule base, and calculating the association matching degree by combining the total number of matching fields; for operator-to-error type association types, matching the number of matching fields between operator ID and error type in each record of the compliance traceability dataset and the structured association rule base, and calculating the behavior matching degree by combining the total number of matching fields.

[0088] The overall matching score is calculated by weighting the comprehensive condition compliance, association matching score, and behavior matching score, and then generating an intermediate association table with matching score.

[0089] For each intermediate matching record in the intermediate table with matching degree, the association weight is obtained by calculating the product of the association confidence degree and the overall matching degree;

[0090] By integrating the core dimensions, association types, and association weights of each intermediate matching record, a structured multidimensional weighted association table is constructed to ensure that the information of each dimension corresponds one-to-one with the weight; among which, the core dimensions include operator, equipment, error type, and cycle.

[0091] Specifically, the steps for generating cross-period correlation maps include:

[0092] Based on a multidimensional weighted association table, three types of nodes are defined: operator, equipment, and error type. The power plant's basic database is then called to supplement node attributes, such as operator position, equipment operating years, and error type severity level, to complete node initialization.

[0093] For three types of associations—device-device combinations, operator-device combinations, and operator-error type combinations—corresponding directed edges are constructed. The initial attributes of the directed edges include association weights and corresponding period labels. The NetworkX library is used to store the data, forming a basic association graph containing node attributes, directed edge weights, and period information. The basic association graph is then divided into N periodic subgraphs according to the period labels, with each subgraph retaining only the nodes and directed edges within the corresponding period.

[0094] For each subgraph, calculate the weight distribution characteristics of directed edges of the same type, such as standard deviation. Then, use Z-score standardization to convert the edge weights into relative strengths within a period, enhancing the comparability of associations within the same period. Label each subgraph with a period identifier, such as a period number, forming a set of subgraphs with periodic attributes. ,in For the first Subgraphs of each period, Furthermore, the periods of each subgraph are arranged in chronological order.

[0095] For adjacent periodic subgraphs in the subgraph set and By traversing and comparing and The node set and directed edge set are used to filter out pairs of nodes of the same type that exist simultaneously, such as operator a and device B;

[0096] Set a weight threshold; if similar node pairs are in adjacent periodic subgraphs... or If the edge weight is greater than the weight threshold, it indicates that the same type of node pair has a valid association in adjacent periods and is judged as a cross-period association candidate pair; otherwise, it indicates that the same type of node pair does not exist in adjacent periods and is judged as a non-cross-period association candidate pair.

[0097] For cross-period correlation candidate pairs, by calculating the first... Subgraphs of each period The middle edge weight and the bridge weight of the preceding term with the preset first weight coefficient, the first edge weight Subgraphs of each period The bridging strength is obtained by summing the middle and side weights with the subsequent bridging weights and the preset second weight coefficient, and then summing the preceding and subsequent bridging weights.

[0098] The bridging strength is defined as the cross-period correlation strength, generating a subgraph connecting adjacent periods. and The bridging edges are then saved to the subgraph set to form an intermediate graph with cross-period connections;

[0099] The label propagation algorithm is adopted, which enables adjacent nodes in the intermediate graph to gradually share the same label through the transfer of labels between nodes, thereby dividing them into closely related clusters and marking the core community nodes within the clusters.

[0100] Based on edge weights and cross-cycle association strength, a maximum threshold is preset using statistical methods to filter strongly associated edges with weights exceeding the preset maximum threshold, prioritizing the retention of associated paths containing core community nodes, and fully recording node pairs, cycle sequences, and intensity change information; a minimum threshold is preset using statistical methods to simultaneously filter weakly associated edges with weights below the preset minimum threshold, eliminating one-off accidental associations and reducing redundancy in intermediate graphs.

[0101] The label propagation algorithm is adopted, which enables adjacent nodes in the intermediate graph to gradually share the same label through the transfer of labels between nodes, thereby dividing them into closely related clusters and marking the core community nodes within the clusters.

[0102] Based on edge weights and cross-cycle association strength, strong association edges with weights higher than the preset maximum threshold are selected within the cluster, and association paths containing core community nodes are retained first. Node pairs, cycle sequences, and strength change information are fully recorded. Simultaneously, weak association edges with weights lower than the preset minimum threshold are selected to eliminate one-off accidental associations and reduce redundancy in intermediate graphs.

[0103] Add time dimension labels to the intermediate graph to supplement the core association range attributes of the community core nodes, and finally generate a cross-period association graph; the cross-period association graph includes, but is not limited to, node attributes, cross-period association strength, and time series features.

[0104] Specifically, the steps for generating the initial error evolution sequence include:

[0105] Using graph database queries, such as Cypher, we filter the associated edges containing operator, equipment, and error type nodes from the cross-cycle association graph, extract the cycle label and cross-cycle association strength of each associated edge, and associate them with auxiliary features such as equipment status parameters, error details, and operator operation duration. The data is then structured and stored in tabular form to form the original time-series association dataset.

[0106] For each unique operator ID, device number, and error type triplet, arrange the time series data in the original time series association dataset in ascending order of period, so as to continuously... The time series data is divided into multiple overlapping sliding windows, with each period serving as the window size and a single period serving as the step size.

[0107] For numerical features within each sliding window, such as correlation strength, normalization is performed to eliminate the influence of dimensions; for categorical features within each sliding window, such as error type, one-hot encoding is used to convert them into binary vectors; the numerical and categorical features within the same sliding window are concatenated into a fixed-length feature vector to form a window feature vector sequence corresponding to the operation error triplet; combined with the window feature vector sequences corresponding to all operation error triplets, a set of encoded temporal feature sequences is formed; in this embodiment, the fixed length is set to L.

[0108] Based on the window feature vector sequence of each operation error triplet in the temporal feature sequence set, a self-attention mechanism is introduced to calculate the attention weight of each sliding window in the window feature vector sequence. The windows are sorted in descending order according to the attention weight, and the first n windows are retained as key sequence segments. Redundant segments are filtered out, and the attention weight and period of the key sequence segments are recorded synchronously.

[0109] Group by operator, traverse key sequence segments, filter target segments with the same operator ID, and splice them in chronological order to generate an error evolution sequence centered on the operator;

[0110] Group by device, traverse key sequence segments, filter matching device segments with the same device number, and splice them in chronological order to generate an error evolution sequence centered on the device;

[0111] Grouping by error type, filtering matching error segments with the same error type by traversing key sequence segments, and splicing them in cyclical order to generate an error evolution sequence with the error type as the core;

[0112] The final result is a preliminary error evolution sequence grouped by operator, equipment, and error type.

[0113] The periodic error analysis module is used to comprehensively utilize cross-period correlation maps and preliminary error evolution sequences. Using the error occurrence rate and corresponding fluctuation range of similar equipment and the same job as the historical baseline, it analyzes frequency changes and periodic patterns through dynamic trend fitting, and combines the response characteristics of equipment maintenance status correlation analysis to form a periodic error trend feature library containing frequency changes, periodic patterns and response characteristics. It also identifies a list of periodic recurring errors with marked trend deviations, repetition patterns and potential causes.

[0114] Specifically, the steps for constructing a periodic error trend feature library include:

[0115] By comprehensively utilizing cross-cycle correlation maps and preliminary error evolution sequences, basic groups are constructed with equipment type, operator ID, cycle label, and error type as core dimensions, and these are recorded as first-level groups. This ensures that each group corresponds to a unique combination of equipment type, operator, cycle label, and error type.

[0116] The frequency of edges associated with each error type node within each group is counted and defined as the error frequency. This ensures that only a unique error frequency record is generated for each group, avoiding duplicate counting. At the same time, the equipment maintenance status within the corresponding period, such as normal maintenance and delayed maintenance, is extracted from the equipment node attributes of the cross-period association graph.

[0117] Based on the first-level grouping, in the initial error evolution sequence, the sequence segments with consistent device type, same operator ID, matching cycle label, and corresponding error type are located, and the valid error occurrence timestamp and valid repair completion timestamp for each error type in the corresponding cycle are extracted from the time series field of the sequence segments;

[0118] The response time is calculated based on the valid error occurrence timestamp and the valid repair completion timestamp, including: if the error type is repaired within the same cycle, the response time is recorded as follows. One cycle; if the error type requires cross-cycle repair, then calculate the number of cycles required to cross the cycle. Response time is recorded as One cycle; if there is no corresponding timestamp record in the initial error evolution sequence, the error handling log of the associated power plant operation and maintenance management system will be supplemented and marked as to be supplemented;

[0119] The system calls upon the employee job configuration database to match the corresponding job information. Combining the equipment type and operator ID of the primary group, it uses equipment type, job position, periodic label, and error type as core dimensions to form a secondary group. Error frequency, equipment maintenance status, and response time are then associated with the secondary group to form structured periodic error basic feature data.

[0120] Using similar equipment and the same job as the core dimensions of the historical baseline, the error occurrence rate and fluctuation range of similar equipment and the same operation are extracted from the preset historical database as the historical baseline. Using equipment type, job position and error type as dimensions, the historical baseline is associated with secondary groups and integrated into a structured periodic error basic dataset. The periodic error basic dataset includes, but is not limited to, error frequency, equipment maintenance status, response time, historical fluctuation range and error frequency periodic sequence.

[0121] Specifically, the steps for constructing a periodic error trend feature library also include:

[0122] For each group of equipment type, job position, and error type in the periodic error dataset, calculate the month-on-month growth rate of error frequency in adjacent periods in ascending order of period to obtain the month-on-month growth rate sequence of error frequency for each period; use a linear regression model to fit the dynamic trend of the month-on-month growth rate sequence of error frequency for each period and output the dynamic trend slope; if the dynamic trend slope is positive, it is determined that the frequency is on an upward trend; otherwise, it is determined that the frequency is on a downward trend.

[0123] The error frequency is judged based on the historical baseline fluctuation range. If the month-on-month growth rate of the error frequency exceeds the historical baseline fluctuation range, the corresponding period is marked as an abnormal period, and the number of abnormalities and the deviation are recorded; otherwise, it is marked as a normal period, and the number of consecutive normal periods is counted.

[0124] Error frequency periodic sequences are extracted from the periodic error dataset. Fourier transform is used to decompose the error frequency periodic sequences into a frequency domain spectrum. The frequency component with the largest amplitude is selected from the spectrum and defined as the dominant frequency. The ratio of the sampling period to the dominant frequency is calculated to obtain the period length. In this embodiment, the sampling period is set to 1.

[0125] Based on expert experience, a preset frequency threshold is established. If the amplitude corresponding to the main frequency exceeds the preset frequency threshold, it is determined that the erroneous frequency periodic sequence has a periodic pattern, and the period length is recorded; otherwise, it is determined that the erroneous frequency periodic sequence does not have a periodic pattern and is marked as non-periodic.

[0126] The equipment maintenance execution cycle is obtained from the equipment node attributes of the cross-cycle correlation graph. Taking the equipment maintenance execution cycle as the boundary, the error frequency of the first 'a' cycles before maintenance is selected as the pre-maintenance frequency, and the error frequency of the last 'a' cycles after maintenance is selected as the post-maintenance frequency. The difference between the pre-maintenance frequency and the post-maintenance frequency is calculated, and then the difference is divided by the pre-maintenance frequency to obtain the post-maintenance frequency reduction rate.

[0127] By setting a response time threshold according to industry standards, the number of errors with a response time lower than the threshold is counted, and the ratio of the number of errors to the total number of errors is calculated to obtain the response time compliance rate.

[0128] The response efficiency score for each group of equipment type, job position, and error type is obtained by weighting the post-maintenance frequency reduction rate and the response time target rate.

[0129] The feature library is divided into units based on equipment type and job position. Each feature library unit is further classified by error type, including: a frequency variation unit for storing error frequency for each period, the year-on-year growth rate sequence of error frequency for each period, the dynamic trend slope, and abnormal periods; a periodicity pattern unit for storing period length, main frequency, the amplitude corresponding to the main frequency, and the periodicity pattern judgment results; and a response feature unit for storing equipment maintenance status, response time, and response efficiency score.

[0130] By integrating frequency variation units, periodicity pattern units, and response characteristic units, a periodic error trend feature library with full dimensions of frequency, period, and response is formed.

[0131] Specifically, the steps for identifying a list of periodically recurring errors include:

[0132] Extract frequency change unit data from the periodic error trend feature library, filter out errors in each group of data that have been continuously marked as abnormal cycles for more than m cycles, and mark them as high-risk candidate errors;

[0133] For high-risk candidate errors, the dynamic trend slope is used to verify the dynamic trend. If the dynamic trend slope is positive and the year-on-year growth rate of the average error frequency in consecutive abnormal periods exceeds the historical baseline, it is judged as a trend deviation error; otherwise, it is judged as random high-frequency fluctuation and is continuously monitored.

[0134] The trend deviation index is calculated by weighting the number of consecutive abnormal periods and the year-on-year growth rate of the average error frequency. Set the trend deviation threshold. , , ,and ;like If the deviation is classified as severe; If the deviation level is determined to be more severe; If the deviation level is determined to be the general level; if If the deviation level is determined to be at the attention level, a preliminary error list containing trend deviation errors and deviation levels will be generated.

[0135] By combining the periodic pattern unit, errors that are determined to be non-periodic are filtered out from the initial error list, and only errors that exhibit periodic patterns are retained;

[0136] For errors that exhibit periodic patterns, theoretical occurrence points are constructed in the error frequency periodic sequence based on the period length. The proportion of actual abnormal periods overlapping with theoretical periods is statistically analyzed and recorded as the pattern matching degree. If the pattern matching degree exceeds the preset matching degree threshold, it is determined to be a repeating pattern error, and the matching degree value and period length are marked. Otherwise, it is determined to be a non-repeating pattern error and marked as a weak periodic interference.

[0137] Based on the repeated pattern error, the matching degree value, and the cycle length, the preliminary error list is updated to obtain the repeated error list;

[0138] Extract response feature unit data from the periodic error trend feature library, including the equipment maintenance status, response efficiency score, and post-maintenance frequency reduction rate of the corresponding error;

[0139] For errors in the recurring error list, extract the equipment maintenance status within the recurring pattern cycle and perform maintenance correlation analysis, including: if the response efficiency score within the corresponding cycle of delayed maintenance is lower than the preset score value and the frequency decrease rate after maintenance is negative, then mark it as a maintenance-related candidate cause; otherwise, mark it as a non-maintenance-related cycle error and monitor it continuously.

[0140] The system retrieves the power plant's operation and maintenance logs and queries the records of maintenance execution for errors in the duplicate error list within the corresponding extended maintenance period. If records of missing maintenance processes or insufficient maintenance resources are found, the system determines that the maintenance-related candidate cause is valid and marks it as a potential cause; otherwise, the system determines that the maintenance-related candidate cause is invalid and marks it as pending verification.

[0141] Finally, by combining information from three categories—trend deviation, repetition pattern, and potential triggers—a complete list of periodic recurring errors was determined.

[0142] The multi-level early warning processing module is used to construct a set of errors to be warned based on the periodic error trend feature library and the periodic recurring error list. It combines the comprehensive early warning level score of deviation level, recurrence pattern and potential causes to trigger multi-level early warning and perform multi-level early warning processing.

[0143] Specifically, the steps for handling multi-level early warnings include:

[0144] By analyzing the structured fields of each combination of equipment type, job position, and error type in the periodic error trend feature database, and combining them with key tags extracted from the periodic recurring error list, such as deviation level, cycle length, pattern matching degree, and potential causes, a set of errors to be warned is constructed.

[0145] Determining weights using the analytic hierarchy process , , The deviation level quantization value is calculated through the deviation level and numerical mapping rule. The intensity of the repeating pattern is quantified by calculating the ratio of the pattern matching degree to the period length. The quantification value of potential causal markers is calculated through the mapping rule between causal state and numerical value. ;Will , , Each with corresponding weight , , Multiply the results and sum them to obtain the comprehensive early warning level score. The mapping rules between deviation level and numerical value and between cause state and numerical value are determined based on expert experience. The comprehensive warning level scoring expression is as follows:

[0146]

[0147] in, To deviate from the weighting of the grades, Values ​​are assigned based on severity level, moderate severity level, general severity level, and concern level. , , , ; Weights for the intensity of the repetitive pattern; Assign weights to potential incentives. Values ​​are taken based on whether the status is established or pending verification. , ;

[0148] Set three threshold levels , , ,in The comprehensive early warning level is scored by combining the set of errors to be warned. Mapping of early warning levels; when When it is determined to be a Level III warning, an immediate response and emergency measures are required; when When it is determined to be a Level II warning, close monitoring and a response plan are required; when When it is determined to be a Level 1 warning, there is a potential risk and continuous monitoring is required; when When this occurs, it is determined that there is no warning and the current status is normal;

[0149] Differentiated handling strategies are generated based on the different warning levels, including: for Level 3 warnings, the system is immediately pushed to the emergency response system to generate and execute temporary inspection tasks; for Level 2 warnings, the system is pushed to some responsible persons to generate and execute special analysis tasks.

[0150] A dynamic feedback mechanism is activated, and settings are configured for Level 3 and Level 2 early warnings. The periodic review mechanism counts the number of recurrences and the total number of monitoring sessions during the implementation of the differentiated treatment strategy, as well as the number of affected devices before and after the implementation of the differentiated treatment strategy.

[0151] The recurrence count is defined as the number of times the same type of error that has occurred before recurred within a specific monitoring period. The error recurrence rate is obtained by calculating the ratio of the recurrence count to the total number of monitoring sessions. Affected devices are defined as devices that experience direct malfunctions or functional failures due to the error before or after its occurrence or strategy implementation. Based on the number of affected devices before and after implementing the differentiated processing strategy, the difference between the number of affected devices before and after implementing the differentiated processing strategy is calculated. Dividing this difference by the number of affected devices before implementing the differentiated processing strategy yields the error impact reduction ratio. A weighted calculation of the error recurrence rate and the error impact reduction ratio yields the effectiveness index. Set the threshold for the indicator; if If the value exceeds the indicator threshold, the early warning is deemed effective; otherwise, it is deemed ineffective. Common characteristics of ineffective early warnings are statistically analyzed, such as job type, and weight parameters are dynamically adjusted accordingly. , , For example, adding weight parameters This improves the sensitivity of early warning systems and optimizes their accuracy.

[0152] In summary, this invention cleans and semantically analyzes raw power plant operation data using a data acquisition module, and labels the data with four dimensions: operator, equipment, error type, and cycle, based on a pre-defined regulatory rule base, to construct a compliance traceability dataset. A traceability graph construction module integrates the compliance traceability dataset, equipment association logic, and personnel operation association rules to construct a multi-dimensional weighted association table. Graph analysis generates a cross-cycle association graph, and sequence aggregation generates a preliminary error evolution sequence centered on the operator, equipment, and error type. A cycle error analysis module comprehensively utilizes the graph and the preliminary error evolution sequence, using the historical error incidence and fluctuation range of similar equipment and positions as the basis for analysis. Historical baselines are used to analyze the year-on-year growth rate and cyclical patterns of error frequency. Combined with equipment maintenance status and response efficiency, a periodic error trend feature library is constructed, incorporating frequency changes, cyclical patterns, and response characteristics. A list of periodically recurring errors is determined based on trend deviations, repetition patterns, and potential triggers. A multi-level early warning processing module is used, based on a comprehensive early warning level score considering deviation levels, repetition pattern intensity, and potential triggers. Multi-level early warning thresholds are set to trigger Level 1, Level 2, and Level 3 warnings, implementing differentiated processing strategies. A dynamic feedback mechanism is introduced, and effectiveness is verified through a review mechanism. The processing effect is evaluated based on effectiveness indicators, and weight parameters are dynamically adjusted to achieve closed-loop optimization and continuous iteration of the early warning process. This invention realizes cross-cycle tracing of power plant error data, periodic pattern recognition, and adaptive intelligent early warning, improving the proactive prevention and control capabilities of system operation risks.

[0153] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A cross-period backtracking analysis system for regulatory rule error data, characterized by, The method comprises the following steps: The data acquisition module is used for collecting, cleaning and analyzing the original data of power plant operation, labeling four-dimensional tags, and constructing a compliance traceability dataset; The traceability graph construction module is used for fusing the compliance traceability dataset, device association logic and personnel operation association rules, constructing a multi-dimensional weighted association relationship table, generating a cross-cycle association graph and a preliminary error evolution sequence; The cycle error analysis module is used for combining the cross-cycle association graph and the preliminary error evolution sequence, analyzing the frequency change, cycle law and response characteristics through dynamic trend fitting, constructing a periodic error trend feature library and determining a periodic repeated error list; The multi-level early warning processing module is used for constructing a to-be-early-warned error set, combining a comprehensive early warning level score, triggering multi-level early warning and processing. The specific steps of constructing the multi-dimensional weighted association relationship table comprise the following steps:

2. The cross-period backtracking analysis system of regulatory rule error data of claim 1, wherein, Obtain the device association logic, identify the device type and extract the association conditions to form a device association triple of device A, device B and association conditions; Based on the basic data, the time window is counted to obtain the time sequence co-occurrence frequency of error types, the association strength is calculated combined with the total number of operations, and the association confidence is normalized; Screening time period To The complete log, grouped by operator ID and equipment number, retains the basic data associated with each operator and equipment; According to the device-device, operator-device and operator-error type association types, a structured association rule library containing device association triples and association confidence is constructed. The specific steps of constructing the multi-dimensional weighted association relationship table further comprise the following steps:

3. The cross-period backtracking analysis system of regulatory rule error data of claim 2, wherein, Iterate through each compliance traceability record of the compliance traceability dataset; For the device-device association type, the operation time sequence and state sequence are extracted, the condition compliance degree is calculated by alignment analysis combined with the time sequence synchronization score and the state response matching score; For the operator-device and operator-error type association types, the matching field number of each compliance traceability record and the structured association rule library is counted respectively, and the association matching degree and the behavior matching degree are calculated combined with the total comparison field number; The comprehensive matching degree is calculated by weighting the condition compliance degree, the association matching degree and the behavior matching degree, and the association intermediate table is generated; The association weight is calculated for each intermediate matching record of the association intermediate table combined with the association confidence and the comprehensive matching degree; The core dimension of each intermediate matching record, the association type and the association weight are integrated to construct a multi-dimensional weighted association relationship table. The specific steps of generating the cross-cycle association graph comprise the following steps:

4. The cross-period backtracking analysis system of regulatory rule error data of claim 3, wherein, Based on the multi-dimensional weighted association relationship table, the operator, device and error type nodes are initialized; For the three types of association types, directed edges are constructed to form a basic association graph, and the basic association graph is split into cycle subgraphs; The directed edge weight distribution characteristics of each cycle subgraph are calculated to generate a subgraph set, and the same type node pairs are screened by iterating and comparing adjacent cycle subgraphs; If the same type node pair has an edge weight less than a preset weight threshold in the adjacent cycle subgraph, it is determined as a non-cross-cycle association candidate pair; otherwise, it is determined as a cross-cycle association candidate pair, the previous bridge weight and the next bridge weight are calculated, the sum of the cross-cycle association strength is calculated, the bridge edge is generated and the intermediate graph is formed; ​ The clusters are divided by inter-node label propagation, strong correlation edges with weights exceeding a preset maximum threshold are filtered out in the clusters, and the correlation paths are retained; and weak correlation edges with weights lower than a preset minimum threshold are filtered out, and accidental correlations are removed; A time dimension label is added to the intermediate graph to generate a cross-period correlation graph.

5. The cross-period backtracking analysis system of regulatory rule error data of claim 4, wherein, The specific steps of generating the preliminary error evolution sequence include: Extracting correlation edges and auxiliary features from the cross-period correlation graph to generate an original time series correlation dataset; The time series data in the operator ID, equipment, and operation error triplet are divided into sliding windows in ascending order of period; Based on the sliding windows, the numerical value features are normalized, the category features are one-hot encoded, and the fixed length feature vectors are concatenated to generate a time series feature sequence set; The self-attention mechanism is introduced to calculate the attention weight, and the top n windows are retained as key sequence fragments by descending order sorting; According to the operation error triplet, the key sequence fragments are traversed, and target fragments with the same operator ID, matching equipment fragments with the same equipment number, and matching error fragments with the same error type are filtered out, and a preliminary error evolution sequence is generated by concatenating in the order of period.

6. The cross-period backtracking analysis system of regulatory rule error data of claim 5, wherein, The specific steps of constructing the periodic error trend feature library include: A first grouping is constructed by device type, operator ID, period label, and error type, error frequencies are counted, and equipment maintenance status is extracted; According to the first grouping, sequence fragments in the preliminary error evolution sequence are located, effective error occurrence timestamps and effective repair completion timestamps are extracted, and response durations are calculated; A second grouping is constructed by device type, post, period label, and error type in combination with the first grouping; The error frequency, the equipment maintenance status, and the response duration are associated to form periodic error basic feature data; The historical error occurrence rate and fluctuation range of similar devices and the same post are used as historical baselines, which are associated to the second grouping to generate a periodic error basic dataset.

7. The cross-period backtracking analysis system of regulatory rule error data of claim 6, wherein, The specific steps of constructing the periodic error trend feature library also include: For each device type, post, and error type in the periodic error basic dataset, the adjacent period error frequency year-on-year growth rate is calculated in ascending order of period, and the dynamic trend slope is obtained by linear fitting; If the error frequency year-on-year growth rate exceeds the fluctuation range, the corresponding period is marked as an abnormal period; otherwise, it is marked as a normal period; The error frequency period sequence is obtained, the frequency spectrum is obtained by signal decomposition, the main frequency frequency with the largest amplitude is filtered out, and the period length is calculated in combination with the sampling period; If the amplitude of the main frequency frequency exceeds a preset frequency threshold, it is determined that there is a periodic rule; otherwise, it is determined that it is non-periodic; The maintenance execution period is obtained, the post-maintenance frequency drop rate is calculated in combination with the pre-maintenance frequency and the post-maintenance frequency, the number of errors with a response duration lower than a preset response duration threshold is counted, the response duration compliance rate is calculated, and the response efficiency score is calculated by weighting. The frequency change unit, the periodic rule unit, and the response feature unit are constructed to store the calculated various indicators respectively, and the periodic error trend feature library is constructed.

8. The cross-period backtracking analysis system of regulatory rule error data of claim 7, wherein, The specific steps of determining the periodic repeated error list include: Obtain frequency change unit data, and screen high-risk candidate errors that exceed m abnormal periods in succession; If the dynamic trend slope is a positive value and the average error frequency year-on-year growth rate exceeds the historical baseline, determine that the trend deviates from the error; otherwise, determine that it is accidental fluctuation; Count the number of consecutive abnormal periods, calculate the trend deviation index by weighting, determine the deviation level in combination with a preset trend deviation index threshold, and form a preliminary error list; In combination with the period rule unit, retain errors with period rules in the preliminary error list, calculate the rule matching degree based on the period length, determine the repetition mode, and update it to a repeated error list; Obtain response feature unit data, and if the response efficiency score in the delayed maintenance corresponding period is lower than a preset score value and the frequency decline rate after maintenance is a negative value, mark the maintenance associated candidate cause; otherwise, mark the non-maintenance associated type period error; Call the power plant operation and maintenance log, query the error maintenance record in the repeated error list, and if maintenance is missing or resources are insufficient, mark the potential cause; otherwise, mark to be checked; and finally determine the periodic repeated error list.

9. The cross-period backtracking analysis system of regulatory rule error data of claim 8, wherein, The specific steps for processing multi-level early warning include: Based on the periodic error trend feature library and the periodic repeated error list, construct a to-be-early-warned error set; The deviation level quantization value, the repetition mode intensity quantization value and the potential inducement sign quantization value are calculated respectively, and the comprehensive early warning level score is calculated by weighting ; Setting a three-level threshold , , wherein , a warning level mapping is performed in combination with the set of errors to be warned. When a tertiary warning is determined, an immediate response is required. When a secondary warning is determined, close attention is required. When a first level warning is determined, there is a potential risk and continuous monitoring is required; When no warning is determined.

10. The cross-period backtracking analysis system of regulatory rule error data of claim 9, wherein, The specific steps for processing multi-level early warning also include: According to different early warning levels, implement differentiated processing strategies; Start a dynamic feedback mechanism, set a back-checking mechanism, count the number of recurrences in the process of executing the differentiated processing strategies and the total number of monitoring times, and calculate the error recurrence rate; And count the number of affected devices before and after the execution of the differentiated processing strategies, and calculate the error impact range reduction ratio; The effectiveness index is obtained by weighting calculation in combination with the error recurrence rate and the error influence range reduction ratio ; If If the index is greater than the preset index threshold, it is determined that the early warning processing is effective; otherwise, it is determined that the early warning processing is ineffective, the common features of the ineffective early warning processing are counted, and the weight parameters are dynamically adjusted.

Images