An insurance industry supervision reporting shunting method and system based on exception types

Abstract

The application provides an insurance industry supervision reporting shunting method and system based on an exception type, the method comprising the following steps: obtaining insurance industry to-be-reported data, identifying the to-be-reported data for exceptions through a multi-agent system, and generating exception identification information of the to-be-reported data; quantitatively evaluating the severity of the exceptions of the to-be-reported data according to the exception identification information of the to-be-reported data, quantitatively evaluating each piece of to-be-reported data with exceptions from four dimensions of risk level, influence range, correction difficulty and historical frequency, and obtaining a comprehensive score of each piece of to-be-reported data with exceptions; dividing the exception data into different shunting levels according to the comprehensive score and a preset shunting threshold, and performing different shunting reporting processing according to the shunting levels, thereby changing the mode of uniformly processing the exception data in the prior art and improving the reliability of the insurance industry supervision reporting data processing through a four-level shunting mechanism.

Description

Technical Field

[0001] This invention relates to the field of data processing for regulatory reporting in the insurance industry, and in particular to a method and system for diverting regulatory reporting in the insurance industry based on anomaly types. Background Technology

[0002] In existing insurance regulatory data verification schemes, after a multi-agent system identifies abnormal data, it typically adopts a uniform approach—either returning the entire batch of data to the business system for resubmission, or simply marking it as abnormal and still submitting it according to the regular process. This "one-size-fits-all" approach has significant technical flaws: First, the risk levels and handling methods differ significantly among different anomaly types. Data format errors may only require simple correction and resubmission, while data integrity deficiencies may necessitate supplementary data collection by the business system, and abnormal trends may indicate problems within the business logic itself. A uniform approach cannot adapt to the differentiated needs of different anomaly types.

[0003] Second, the urgency of the abnormal data differs. Data anomalies involving significant risk indicators require immediate manual intervention, while general formatting flaws can be automatically corrected and the data can continue to be submitted. Returning the entire batch of data at once would delay urgent reporting tasks.

[0004] Third, existing solutions lack quantitative assessment of the severity of anomalies and a triage decision-making mechanism based on assessment results.

[0005] Therefore, there is an urgent need for a data processing method and system that can make intelligent triage decisions based on the type and severity of anomalies. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention innovatively proposes a method and system for diverting insurance industry regulatory reports based on anomaly types. This effectively solves the problem of low reliability in processing insurance industry regulatory reports due to existing technologies, and significantly improves the reliability of insurance industry regulatory reports processing.

[0007] The first aspect of this invention provides a method for diverting regulatory reports in the insurance industry based on anomaly types, comprising: The system acquires data to be reported from the insurance industry, performs anomaly identification on the data through a multi-agent system, and generates anomaly identification information for the data to be reported. The anomaly identification information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. Based on the anomaly identification information of the data to be reported, the severity of the anomalies of the data to be reported is quantitatively assessed. The anomalies are evaluated from four dimensions: risk level, scope of impact, difficulty of correction, and historical frequency, and a comprehensive score is obtained for each anomaly data to be reported. Based on the comprehensive score and the preset diversion threshold, abnormal data is divided into different diversion levels, and differentiated diversion reporting processing is performed according to the diversion level.

[0008] Optionally, the multi-agent system includes a rule verification agent, an anomaly detection agent, and a trend comparison agent; anomaly types include format anomalies, integrity anomalies, business logic anomalies, statistical anomalies, and trend anomalies; the multi-agent system identifies anomalies in the data to be reported, generating anomaly identification information for the data to be reported, including: The rule verification agent is used to verify the format compliance of the data to be submitted. Referring to regulatory documents, it transforms the regulatory system into executable verification rules to verify the field format, field integrity and logical consistency. The anomaly detection agent is used to identify anomalies at the business logic level in the data to be reported, including regulatory red line detection, statistical anomaly detection, and data quality detection. The trend comparison agent is used to perform vertical and horizontal trend analysis on the data to be reported. Through month-on-month change detection, year-on-year change detection, and peer comparison detection, it identifies abnormal changes in time series or among similar institutions. After the rule verification agent, anomaly detection agent, and trend comparison agent have completed their respective verifications, cross-validation is performed among the rule verification agent, anomaly detection agent, and trend comparison agent.

[0009] Furthermore, the cross-validation among the rule verification agent, the anomaly detection agent, and the trend comparison agent is as follows: For logical contradictions marked by the rule verification agent, the anomaly detection agent further verifies whether the data to be reported involves regulatory red line indicators; the trend comparison agent confirms whether the logical contradictions in the data to be reported are occasional errors or systemic trend anomalies. If all three agents point to the same problem, it is directly confirmed; if only a single agent marks an anomaly, the priority is reduced or it is submitted for manual review.

[0010] Optionally, based on the anomaly identification information of the data to be reported, the severity of the anomalies in the data to be reported is quantitatively assessed. Each piece of abnormal data is quantitatively assessed from four dimensions: risk level, scope of impact, difficulty of correction, and historical frequency, resulting in a comprehensive score for each piece of abnormal data to be reported. Obtain abnormal pending data, and obtain the risk level score, impact scope score, correction difficulty score and historical frequency score of the abnormal pending data; The comprehensive score for each abnormal data point to be reported is obtained by weighting and summing the risk level score, impact scope score, correction difficulty score, and historical frequency score. The weighting of different dimensions is dynamically determined based on the regulatory focus shift factor, the data quality characteristics of the batch to be reported, and the feedback of historical diversion effects.

[0011] Furthermore, the risk level score is determined based on whether the data items involved in the abnormal data to be reported are preset core regulatory indicators and the real-time weight attributes of the data item nodes in the knowledge graph where the data items are located; specifically, the risk level score R(e) is calculated as follows: R(e) = [W base (i) + ·C factor (i) + Dev factor (i)] Among them: W base (i) indicates whether data item i is a basic regulatory weight for a preset core regulatory indicator; C factor (i) represents the regulatory dynamic attention factor corresponding to data item i; Dev factor (i) represents the current deviation factor corresponding to data item i; , These are the weight coefficients of the regulatory dynamic attention factor and the current deviation factor, respectively. This is the sigmoid normalization function.

[0012] Optionally, the impact scope score is determined based on the proportion of records affected by the abnormal data to be reported and the number of reporting dimensions involved; specifically, the impact scope score I(e) is calculated as follows: I(e) = min{1, ·P ratio + ·ln(1 + Dim count ) + ·Layer depth} Where: P ratio Dim represents the number of affected records for abnormal pending data submissions divided by the total number of records in the batch. count This refers to the number of reporting dimensions involved in the abnormal pending data; Layer depth The depth of the hierarchy is affected by abnormal pending data. , , These are the weighting coefficients for the percentage of records, the number of reporting dimensions, and the influence on the depth of the hierarchy, respectively.

[0013] Optionally, the correction difficulty score is dynamically determined based on the correction complexity of the anomaly type, the current system resource status, and the reporting timeliness requirements; specifically, the correction difficulty score D(e) is calculated as follows: D(e) = [D] base (i) + Peak factor (i)+ ·Ext dep，factor (i) Where: D base (i) represents the normalized value of the basic modified complexity for the anomaly type corresponding to data item i; Peak factor (i) represents the peak reporting period factor corresponding to data item i; Ext dep，factor (i) represents the external dependency factor corresponding to data item i; , These are the weighting coefficients for the peak reporting period factor and the external dependence factor, respectively. This is the sigmoid normalization function.

[0014] Optionally, the historical frequency score is dynamically determined based on the frequency of occurrence of similar anomalies in the current batch and historical reporting cycles; specifically, the historical frequency score H(e) is calculated as follows: H(e) = min{1, Batch ratio + ·Avg(Hist freq )·exp(- ·Count excessive )} Among them: Batch ratio Hist represents the percentage of similar anomalies in the total number of anomalies in the current batch; freq This is an array representing the monthly trigger counts of similar anomalies over the past 6 months; Avg(·) is used to calculate the average; Count excessive This refers to the number of months in which the historical trigger count exceeded the warning threshold; , These are the weighting coefficients for the current batch and the historical reporting period, respectively; ξ is the attenuation coefficient.

[0015] Optionally, the traffic triage levels are divided into four levels from mild to severe anomalies: Level 1, Level 2, Level 3, and Level 4. Based on the comprehensive score and preset triage thresholds, abnormal data is classified into different triage levels, and differentiated triage reporting processing is performed according to the triage level. When the exception type is a format error and the comprehensive score does not exceed the first preset triage threshold, the corresponding triage level is level one; after the data correction is completed, the verification is re-executed, and after the verification is passed, the report is submitted according to the normal process. When the anomaly type is data missing or logical error, and the comprehensive score is greater than the first preset triage threshold but less than the second preset triage threshold, the corresponding triage level is level two; a return work order is generated and pushed to the business system, and the work order is resubmitted after the business system corrects the error. When the data items involved in the abnormal data to be reported involve preset core regulatory indicators, and the comprehensive score is greater than the second preset diversion threshold but less than the third preset diversion threshold, the corresponding diversion level is level three; generate a manual approval task and push it to the approver, and continue the reporting process after waiting for approval; When the data items involved in the abnormal data to be reported are preset major risk indicators and the comprehensive score is greater than the third preset diversion threshold, the corresponding diversion level is level four; while performing manual approval, an emergency warning report is generated and reported to the regulatory department to achieve dual-track parallel processing; among them, the first preset diversion threshold is less than the second preset diversion threshold, and the second preset diversion threshold is less than the third diversion threshold; and the first preset diversion threshold, the second preset diversion threshold, and the third preset diversion threshold are dynamically determined according to the abnormal score distribution characteristics, reporting timeliness requirements, and the availability of corrective resources.

[0016] A second aspect of this invention provides an insurance industry regulatory reporting and triage system based on anomaly types, comprising: The data reporting access module acquires insurance industry data to be reported, and uses a multi-agent system to identify anomalies in the data to be reported, generating anomaly identification information for the data to be reported. The anomaly identification information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. The multi-agent verification and anomaly severity quantification assessment module quantifies the severity of anomalies in the data to be reported based on the anomaly identification information of the data to be reported. It quantifies each abnormal data to be reported from four dimensions: risk level, scope of impact, difficulty of correction, and historical frequency, and obtains a comprehensive score for each abnormal data to be reported. The intelligent triage decision module classifies abnormal data into different triage levels based on the comprehensive score and preset triage thresholds, and performs differentiated triage reporting processing according to the triage level.

[0017] The technical solution adopted in this invention has the following technical effects: 1. The technical solution of this invention acquires insurance industry data to be reported, and uses a multi-agent system to identify anomalies in the data, generating anomaly identification information for the data. This anomaly identification information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. Based on the anomaly identification information, the severity of the anomalies in the data is quantitatively assessed. Each piece of abnormal data is evaluated from four dimensions: risk level, scope of impact, correction difficulty, and historical frequency, resulting in a comprehensive score for each piece of abnormal data. Based on the comprehensive score and a preset diversion threshold, the abnormal data is divided into different diversion levels, and differentiated diversion reporting processing is performed according to the diversion level. This four-level diversion mechanism changes the existing unified processing model for abnormal data, avoiding the efficiency loss and delay risks caused by a "one-size-fits-all" approach. It effectively solves the problem of low reliability in processing insurance industry regulatory reporting data caused by existing technologies, and effectively improves the reliability of insurance industry regulatory reporting data processing.

[0018] 2. The multi-agent system of the present invention includes a rule verification agent, an anomaly detection agent, and a trend comparison agent; the anomaly types include format anomalies, integrity anomalies, business logic anomalies, statistical anomalies, and trend anomalies; the multi-agent system identifies anomalies in the data to be reported and generates anomaly identification information for the data to be reported; after the rule verification agent, anomaly detection agent, and trend comparison agent have completed their respective verifications, cross-validation is performed among the rule verification agent, anomaly detection agent, and trend comparison agent. This not only improves the reliability of anomaly identification and detection, but also allows the closed-loop mechanism of the feedback learning agent to continuously optimize the diversion decision parameters based on historical processing results. Compared with the static configuration scheme of the prior art, it has stronger adaptive capabilities.

[0019] 3. In the technical solution of this invention, abnormal data to be reported is obtained by acquiring risk level score, impact range score, correction difficulty score, and historical frequency score of the abnormal data to be reported. The comprehensive score of each abnormal data to be reported is obtained by weighted summation of the risk level score, impact range score, correction difficulty score, and historical frequency score. By introducing a multi-dimensional (risk level, impact range, correction difficulty, historical frequency) comprehensive quantitative evaluation model in the field of insurance regulatory reporting, a scientific quantitative basis is provided for diversion decision-making, overcoming the limitations of existing technologies that rely on experience-based judgment.

[0020] 4. In the technical solution of this invention, the diversion levels are divided into four levels in order of severity from minor to major anomalies: Level 1, Level 2, Level 3, and Level 4. Based on the comprehensive score and the preset diversion threshold, abnormal data is divided into different diversion levels, and differentiated diversion reporting processing is performed according to the diversion level. When the diversion level is Level 4, an emergency early warning channel is set up so that abnormal data involving major risk indicators can trigger manual intervention and early warning reporting in a timely manner, meeting the timeliness requirements of regulatory reporting.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the method of Embodiment 1 in the present invention; Figure 2 This is a schematic diagram of the system structure in Embodiment 2 of the present invention. Detailed Implementation

[0024] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0025] Example 1 This invention provides a method and system for reporting and diverting insurance industry regulatory data based on anomaly types. It aims to solve the problems of uniform processing of abnormal data and lack of differentiated diversion mechanisms in the existing technology, and to realize intelligent reporting and diversion decision-making based on anomaly types.

[0026] To achieve the above objectives, the specific technical solution adopted by the present invention is as follows: First, the reported data is collaboratively verified through a multi-agent system to identify abnormal data and generate structured abnormal information (including abnormality type, abnormality cause, and scope of impact).

[0027] Secondly, a quantitative assessment model for the severity of anomalies is established, with a comprehensive score based on four dimensions: risk level (whether it involves core regulatory indicators), scope of impact (percentage of anomaly records), difficulty of correction (complexity and time required for remediation), and historical frequency (frequency of similar anomalies). The comprehensive score is obtained by weighted summation of the four dimensions.

[0028] Then, based on the comprehensive score, the anomalies are divided into four triage levels: Level 1 (formatting errors, automatically corrected), Level 2 (data missing or logical errors, returned to the business system with correction suggestions), Level 3 (involving core regulatory indicators, pushed for manual approval), and Level 4 (involving major risks, with emergency warnings and manual approval in parallel).

[0029] Next, differentiated processing is performed according to the triage level: Level 1 is automatically corrected and resubmitted after verification; Level 2 generates a work order and returns it to the source; Level 3 waits for manual approval before submission; Level 4 triggers an emergency warning, reports it, and pushes it to manual approval simultaneously.

[0030] Finally, a feedback learning closed-loop mechanism is established to record the decision parameters and actual processing results of each diversion, dynamically optimize the evaluation weights and diversion thresholds, and achieve adaptive evolution of the diversion strategy.

[0031] This solution solves the problem of "one-size-fits-all" handling of abnormal data in existing technologies by establishing a complete closed loop of "quantitative assessment → hierarchical diversion → differentiated processing → feedback optimization," thereby improving reporting efficiency and compliance.

[0032] Specifically, such as Figure 1 As shown, this invention provides a method for diverting regulatory reports in the insurance industry based on anomaly types, including: Step S1, multi-agent collaborative verification; that is, to obtain the data to be reported in the insurance industry, to identify anomalies in the data to be reported through the multi-agent system, and to generate anomaly identification information of the data to be reported. The anomaly identification information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. Step S2, quantitative assessment of the severity of anomalies; that is, based on the anomaly identification information of the data to be reported, the severity of the anomalies of the data to be reported that are marked as anomalies is quantitatively assessed. From four dimensions, risk level, scope of impact, difficulty of correction and historical frequency, each abnormal data to be reported is quantitatively assessed to obtain a comprehensive score for each abnormal data to be reported. Step S3: Determine the triage level based on the comprehensive score; that is, based on the comprehensive score and the preset triage threshold, the abnormal data is divided into different triage levels, and differentiated triage reporting processing is performed according to the triage level.

[0033] In step S1, the multi-agent system collaboratively verifies the reported data. The multi-agent system includes a rule verification agent, an anomaly detection agent, a trend comparison agent, and an anomaly summary module.

[0034] The rule verification agent is responsible for format compliance verification. Referring to regulatory documents such as the EAST data specification and the 1104 regulatory report system, it transforms the regulatory system into executable verification rules (up to 2,000 rules). It verifies the field format (e.g., the date must be in YYYY-MM-DD format and the month must be 01-12, and the certificate number must conform to the GB 11643 standard), field integrity (e.g., core fields such as the main diagnosis code and admission date cannot be empty), and logical consistency (e.g., the total cost equals the sum of the costs of each item, and the admission date must not be later than the discharge date).

[0035] The anomaly detection agent is responsible for identifying anomalies at the business logic level, including regulatory red line detection (such as whether the core solvency adequacy ratio is lower than 50%), statistical anomaly detection (such as the deviation of medical expenses of the same type of institution from the mean by more than ±3σ), and data quality detection (such as the inconsistency between the date of birth and the birthday field in the ID card number). Its detection rules are jointly maintained by the system administrator and the verification personnel and are continuously updated through feedback learning.

[0036] The trend comparison agent is responsible for vertical and horizontal trend analysis. It identifies abnormal changes in time series or among similar institutions by detecting month-on-month changes (such as when an institution's total expenses fluctuate by more than 50% compared to the previous month without a reasonable explanation), year-on-year changes (such as when the current period's compensation amount fluctuates abnormally compared to the same period last year), and peer comparisons (such as when an institution's average hospitalization expenses deviate significantly from the average of similar institutions). Its threshold parameters can be dynamically adjusted according to the distribution of historical data to avoid misjudgments caused by seasonal fluctuations.

[0037] The three intelligent agents work collaboratively according to the process of "parallel screening → cross-validation → comprehensive judgment" to build a complete system for identifying and verifying abnormal data. When a batch of reported data (such as the medical expense settlement data of an insurance company for the current month) enters the system, the three intelligent agents simultaneously start verification: the rule verification agent checks the format compliance and logical consistency of the batch of data field by field, identifying format errors and missing integrity; the anomaly detection agent detects the business compliance of the batch of data and identifies outliers by combining statistical models; the trend comparison agent calls the institution's historical data of the past 12 months to calculate the month-on-month change rate and retrieves peer data from similar institutions at the same level for horizontal comparison. The verification results of the three intelligent agents are aggregated in real time to the anomaly aggregation module, forming a preliminary anomaly list before entering the cross-validation stage. For "logical contradictions" marked by the rule verification intelligent agent (such as total cost not equal to the sum of its components), the anomaly detection intelligent agent further verifies whether the record involves regulatory red-line indicators, and the trend comparison intelligent agent confirms whether the contradiction is an occasional error or a precursor to a systemic trend anomaly. If all three intelligent agents point to the same problem, the anomaly confidence is high and it is directly confirmed. If only a single intelligent agent marks an anomaly, the priority is reduced or it is submitted for manual review. Taking "a certain institution's total cost increased by 55% month-on-month" as an example, the trend comparison intelligent agent first marks the trend as an anomaly. Then, the rule verification intelligent agent checks whether the institution's data for the month has format errors or logical contradictions (excluding data entry errors that lead to inflated figures). The anomaly detection intelligent agent simultaneously checks whether the institution has added any new large-amount claims or regulatory red-line indicator anomalies for the month (determining whether the growth stems from real business changes or potential risk exposure). Finally, the three intelligent agents aggregate their respective verification results to the anomaly aggregation module, which then performs a comprehensive score based on the anomaly severity quantification assessment module.

[0038] After verification, the anomaly summary module generates structured anomaly information in four stages: "anomaly discovery → attribution determination → information assembly → structured output". The anomaly information includes at least the anomaly identifier (whether it is anomaly), anomaly type, anomaly cause, scope of impact, and source agent.

[0039] Taking a batch of medical expense settlement data as an example: The rule verification agent found that the admission date field of the data to be reported, record A (settlement order number JS202607001), was "2025 / 13 / 01" (illegal month), triggering a format anomaly; the total cost of record B (settlement order number JS202607015) was 5,000 yuan, but the sum of the various expense details was 4,800 yuan, triggering a logical contradiction; the main diagnosis code field of record C (settlement order number JS202607023) was empty, triggering an integrity anomaly. The anomaly detection agent found that the core solvency adequacy ratio of record D (institution code 3701) was 45%, which was lower than the regulatory red line of 50%, triggering a regulatory red line anomaly; the medical expenses of record E (settlement order number JS202607045) were 150,000 yuan, while the average of similar cases in this institution was 50,000 yuan and the standard deviation was 20,000 yuan (5 standard deviations from the mean), triggering a statistical anomaly (outlier). The trend comparison AI detected that Institution 3701's total expenses for the month were 8 million yuan, while the average for the past 6 months was 5 million yuan (a 60% increase month-on-month). After verification with the anomaly detection AI, it was confirmed that the institution had no new large-amount claims for the month, thus ruling out business growth factors and confirming an anomaly. Simultaneously, the institution's average hospitalization expenses for the month were 28,000 yuan, while the average for similar institutions of the same level was 16,000 yuan (a 75% deviation), triggering an industry deviation anomaly. In the attribution determination phase, the anomaly aggregation module categorized the anomalies according to their source and nature—report record A was categorized as a format anomaly, report record B as a business logic anomaly (logical contradiction), report record C as a completeness anomaly (data missing), report record D as a business logic anomaly (regulatory red line), and report record E as a statistical anomaly (outlier). Institution 3701's total expense anomaly was categorized as a trend anomaly (month-on-month anomaly), and its average hospitalization expense anomaly was categorized as a trend anomaly (industry deviation). During the information assembly stage, standardized abnormal information is generated for each record: taking report A as an example, the abnormal data identifier is "JS202607001", the abnormal type is "format abnormality", the abnormal reason is "the month of the admission date field is illegal (13 months)", and the scope of impact is "the admission date field of this record"; taking report D as an example, the abnormal data identifier is "3701", the abnormal type is "business logic abnormality", the abnormal reason is "the core solvency adequacy ratio of 45% is lower than the regulatory red line of 50%", and the scope of impact is "all solvency-related indicators involved in the data submitted by the institution".In the structured output stage, all anomaly information is output to the anomaly severity quantification assessment module in a unified data structure. Each anomaly information contains five fixed fields: anomaly data identifier (used to uniquely locate the source of the anomaly data), anomaly type (format anomaly / completeness anomaly / business logic anomaly / statistical anomaly / trend anomaly and its subclasses), anomaly cause (specific description of the violation content or deviation), scope of impact (specific to the field, record or organizational level), and source agent (rule verification / anomaly detection / trend comparison and cross-validation conclusion). This provides standardized input data for the subsequent severity quantification assessment in step S2, thereby completing the entire process from identification to verification to anomaly information generation.

[0040] After the verification results of the three intelligent agents are summarized, the abnormal data are divided into major categories and subcategories such as format abnormalities (including format errors), integrity abnormalities (including missing data), business logic abnormalities (including violations of regulatory red lines and logical contradictions), statistical abnormalities (including outliers) and trend abnormalities (including month-on-month abnormalities and deviations from peers), providing basic data support for subsequent quantitative assessment of the severity of abnormalities and triage decisions.

[0041] In step S2, the anomaly aggregation module transmits structured anomaly information to the anomaly severity quantification assessment module. This module comprehensively quantifies each anomaly data point from four dimensions: risk level, scope of impact, difficulty of correction, and historical frequency, providing a quantitative basis for subsequent triage decisions.

[0042] (a) Risk level assessment Risk level is used to measure whether the data items involved in abnormal data are core regulatory indicators. Its assessment is determined based on the real-time weight attributes of data item nodes in the knowledge graph. The knowledge graph pre-sets basic weights for all data items involved in insurance regulatory reporting according to their regulatory importance, and dynamically adjusts these weights based on the real-time status of the data items during operation. Factors influencing node weights include: regulatory attention factor (dynamically adjusted based on recent regulatory documents and risk warnings issued by regulatory authorities), data item deviation factor (determined based on the degree of deviation between the current value of the data item and the industry average), historical trigger frequency factor (determined based on the frequency with which the data item was marked as abnormal in past reporting periods), and time decay factor (to gradually bring historical weights back to the baseline level). Taking the "core solvency adequacy ratio" data item as an example, if this data item is included in the key monitoring list by regulatory authorities and the institution's current value deviates significantly from the industry average, its real-time weight is close to 1.0, indicating that the anomaly has an extremely high risk level.

[0043] Specifically, the risk level score R(e) is dynamically determined based on the real-time weight attributes of data item nodes in the knowledge graph, and its influencing factors include: Basic regulatory weight: A pre-set base score based on the importance of data items in regulatory policy documents; Regulatory Dynamics Attention Level: The level of real-time attention paid to this data item by regulatory authorities in recent normative documents, risk warnings, special inspection notices, etc. Current deviation of a data item: The degree of deviation between the actual value of this data item in the current batch of submitted data and the industry average. Let data item i be involved in an anomaly data e. Then, its real-time risk level score R(e) is determined by the following formula: R(e) = [W base (i) + ·C factor (i) + Dev factor (i)] Among them: W base (i) is the basic regulatory weight for whether data item i is a preset core regulatory indicator (range 0~0.7, preset according to the importance of the data item in the regulatory system); C factor (i) represents the regulatory dynamic attention factor (range 0~0.3, with the higher value taken when the regulatory authorities are focusing on monitoring) corresponding to data item i; Dev factor (i) is the current deviation factor corresponding to data item i (range 0~0.2, obtained by standardizing the deviation between the current value of the data item and the industry average); , These are the regulatory dynamic attention factor and the current deviation factor weight coefficient (default). =1.0, =0.8); The sigmoid normalization function maps the weighted sum to the interval [0,1].

[0044] Application example: Taking the "core solvency adequacy ratio" data item as an example, the basic regulatory weight W of this data item is... base =0.65. In January 2026, the regulatory authorities included it in the key monitoring list. C factor =0.3; The reported value of a certain institution in the current period is 45%, the industry average is 68%, and the deviation is (68%-45%) / 68%≈33.8%. After standardization, Dev factor =0.15. Therefore, R(e) = [0.65 + 1.0 × 0.3 + 0.8 × 0.15] = [1.07]≈0.92 indicates that the anomaly has an extremely high risk level.

[0045] (II) Impact Scope Assessment The impact assessment comprises two sub-dimensions: the percentage of affected records and the number of reporting dimensions involved. The percentage of affected records is calculated as "number of affected records / total number of records in the batch × 100%", reflecting the coverage ratio of a single anomaly in the batch of data. The number of reporting dimensions involved reflects the breadth of regulatory reporting forms or data dimensions affected by the anomaly data. Different anomaly types may affect a single dimension (e.g., only the "Medical Expense Settlement Details Table") or multiple dimensions (e.g., simultaneously affecting the "Details Table", "Summary Table", and "Institutional Risk Indicator Table"). The comprehensive quantification of the impact scope is based on the percentage of affected records, with the number of reporting dimensions involved as a multiplier factor (single dimension × 1, two dimensions × 1.5, three or more dimensions × 2), normalized to [0, 1] to obtain the impact scope score I(e). Insurance scenario example: A batch of 10,000 medical expense settlement records are submitted. If the "admission date" field of all 500 records submitted by an institution has a format error, the anomaly will affect 5% of the records, which is a relatively wide-ranging anomaly. If the "total cost" field of an institution has a logical contradiction, the anomaly will affect three reporting dimensions: "details table", "summary table" and "institutional risk indicator table", which is a multi-dimensional anomaly that significantly increases the triage level.

[0046] Specifically, the impact scope score I(e) is dynamically determined based on the proportion of records affected by the anomaly and the number of reporting dimensions involved. Its influencing factors include: Record count percentage: The proportion of affected records to the total number of records in the batch; Number of reporting dimensions: The number of regulatory reporting forms or data dimensions affected by abnormal data; Impact level depth: The report level affected by abnormal data (basic detail table, summary table, derived analysis table, etc.). The dynamic determination expression for the scope of influence score I(e): Suppose that the impact range score of a certain outlier data e is determined by the following formula: I(e) = min{1, ·P ratio + ·ln(1 + Dim count ) + ·Layer depth} Where: P ratio Dim represents the number of affected records for abnormal pending data submissions divided by the total number of records in the batch. count This refers to the number of reporting dimensions involved in the abnormal pending data; Layer depth The impact level of abnormal pending data is determined by the hierarchy (basic detail table hierarchy = 1, summary table hierarchy = 2, derived analysis table hierarchy = 3). , , These are the percentage of records, the number of reported dimensions, and the weighting coefficients affecting the depth of the hierarchy (default). =2.0, =0.3, =0.15); min{1, ·} truncates the result to the interval [0,1].

[0047] Application Example: A batch of 10,000 medical expense settlement records was submitted. A logical contradiction / abnormality affected 500 records (P). ratio =0.05), which also affects three reporting dimensions: "Expense Settlement Details", "Settlement Summary", and "Institutional Risk Indicator Table" (Dim count =3), involving a basic detailed table + summary table (Layer) depth =2). Then I(e)=min{1, 2.0×0.05 + 0.3×ln(4) + 0.15×2} = min{1, 0.1 + 0.416 + 0.3} = 0.816, indicating that the anomaly has a wide range of influence.

[0048] (III) Revise the difficulty assessment The difficulty of correction is determined jointly by a "basic mapping table + real-time correction environment factors". The system first queries the pre-set basic correction difficulty mapping table according to the anomaly type to obtain a basic correction complexity score (1-10 points) and a basic expected time. Based on this, the system dynamically adjusts the difficulty by taking into account real-time environment factors such as whether the current period is a peak reporting period, whether correction resources are scarce, and whether the correction depends on external systems.

[0049] Specifically, the correction difficulty score D(e) is dynamically determined based on the correction complexity of the anomaly type, the current system resource status, and the reporting timeliness requirements. Its influencing factors include: Basic correction complexity: The basic correction difficulty obtained from the pre-defined mapping table based on the anomaly type (out of 1 to 10). Peak reporting period factor: Whether the current period is within a concentrated reporting window affects the available time for corrections; External dependency factor: whether cooperation from external systems is required and the current availability status of external systems; D(e) dynamically determines the expression: Suppose that the difficulty score for correcting an outlier data e is determined by the following formula: D(e) = [D] base (i) + Peak factor (i)+ ·Extdep，factor (i) Where: D base (i) represents the normalized value of the base modified complexity for the anomaly type corresponding to data item i (0~0.6); Peak factor (i) represents the peak reporting period factor corresponding to data item i (values ​​range from 0 to 0.2, with higher values ​​taken during the concentrated reporting period at the end of the quarter); Ext dep，factor (i) is the external dependency factor corresponding to data item i (values ​​range from 0 to 0.3, with higher values ​​taken when the system depends on the core business system and is currently busy). , These are the reporting peak period factor and external dependency factor weighting coefficients (default). =0.8, =0.6); This is the sigmoid normalization function.

[0050] Application Example: A logical contradiction anomaly exists where "total cost is not equal to the sum of itemized costs." The basic correction complexity is 6 / 10, and the normalized complexity is D. base =0.6. Currently, it's the peak reporting period at the end of the quarter. factor =0.2; The correction requires retrieving the expense details from the core business system, which is currently undergoing end-of-day batch processing. Ext dep，factor =0.25. Therefore, D(e) = [0.6 + 0.8 × 0.2 + 0.6 × 0.25] = [0.6 + 0.16 + 0.15] = [0.91] ≈ 0.78, indicating that the anomaly is difficult to correct in the current environment.

[0051] (iv) Historical frequency assessment Historical frequency is used to measure how often the same type of anomaly occurs in the current batch or in historical reports. This assessment is determined by statistically analyzing the number of times the same anomaly type occurs in the current batch, combined with historical trigger records of this type of anomaly over several past reporting periods (e.g., the past 6 months). The historical frequency score H(e) ranges from [0, 1], with higher frequencies resulting in higher scores. However, when the historical frequency is too high, the weight is appropriately reduced through a decay mechanism to avoid long-term repetitive anomalies excessively affecting triage decisions.

[0052] Specifically, the historical frequency score H(e) is dynamically determined based on the frequency of occurrence of similar anomalies in the current batch and historical reporting cycles, and its influencing factors include: Current batch percentage: The percentage of the same type of anomaly in the total number of anomalies in the current batch; Historical trigger frequency: The number of times the same type of anomaly has been triggered in the past few reporting cycles (e.g., the past 6 months); Frequency decay factor: When the historical frequency is too high, decay is introduced to avoid long-term repetition of abnormally high frequencies from affecting the diversion decision. H(e) dynamically determines the expression: Suppose that the historical frequency score of a certain outlier data e is determined by the following formula: H(e) = min{1, Batch ratio + ·Avg(Hist freq )·exp(- ·Count excessive )} Among them: Batch ratio Hist represents the percentage of similar anomalies in the total number of anomalies in the current batch; freq This is an array representing the monthly trigger counts of similar anomalies over the past 6 months; Avg(·) is used to calculate the average; Count excessive This refers to the number of months in which the historical trigger count exceeded the warning threshold; , These are the weighting coefficients for the current batch and the historical reporting period (default). =1.5, =0.8); ξ is the attenuation coefficient (default ξ=0.3).

[0053] Application Example: A batch identified 200 "Date format error" exceptions of the same type. The current batch has a total of 500 exceptions. ratio =0.4. The monthly trigger count for this type of anomaly over the past 6 months was [15, 18, 12, 20, 22, 25], with an average of 18.7 times / month. This exceeds the warning threshold of 10 times / month for six consecutive months. Count excessive =6. Then H(e)=min{1,1.5×0.4 + 0.8×18.7 / 30×exp(-0.3×6)} = min{1, 0.6 + 0.499×0.165} = min{1,0.682} = 0.682, indicating that this type of anomaly has a high historical frequency and is a systemic format problem that needs to be solved through source system transformation.

[0054] (V) Overall Score The anomaly severity quantification assessment module performs a weighted summation of the scores from the four dimensions mentioned above to obtain the comprehensive score S(e) for the anomaly data e, expressed as: S(e) = ·R(e) + ·I(e) + ·D(e) + ·H(e), where, , , , The weighting coefficients for each dimension are 0.35, 0.25, 0.25, and 0.15, respectively, with a sum of 1. The comprehensive score S(e) ranges from [0, 1]. A higher score indicates a higher degree of abnormality and requires priority processing. The scores for each dimension and the comprehensive score will serve as input data for determining the triage level in step S3.

[0055] Among them, the weighting coefficient , , , It is not fixed, but dynamically adjusted based on the characteristics of the data submitted in each batch and the regulatory environment, allowing the comprehensive score to adapt to different business scenarios. Its influencing factors include: Regulatory focus shift factor: The weighting of each dimension is dynamically adjusted based on the current key monitoring directions of the regulatory authorities. Batch data quality characteristics: Automatically adjust the weights of each dimension based on the overall quality of the current batch of data; Historical traffic allocation feedback: Based on the accuracy of historical traffic allocation decisions, the weight configuration is dynamically optimized through feedback learning; Dynamically determine the expression: Suppose that the dynamic values ​​of each weight coefficient are determined by the following formula: = [ base + ·Reg shift + Quality adj + ·Feed learn ] = [ base + ·Reg shift + Quality adj + ·Feed learn ] = [ base + ·Reg shift + Quality adj + ·Feedlearn ] = 1 - - -

[0056] in: base , base , base These are the base weighting coefficients (default 0.35, 0.25, 0.25) for the risk level score, impact scope score, and difficulty of correction score, respectively; Reg shift As a regulatory focus offset factor, the offset amount for each dimension is determined based on the recent focus of regulatory authorities; Quality adj This is a batch data quality adjustment factor; when the overall quality of the batch data is poor, the weight of the correction difficulty is increased. (Feed) learn To provide feedback for learning and adjustment factors, the system is optimized in reverse based on historical triage accuracy. ~ For the corresponding adjustment coefficient; (·) is the sigmoid normalization function, which ensures that each weight is within a reasonable range; From 1 - - - Confirm that the sum of the four is 1.

[0057] For example, if the regulatory authorities are currently focusing on monitoring solvency-related data (Reg... shift (biased towards risk level dimension), and the overall quality of this batch of data is relatively good. adj The weighting of correction difficulty was reduced, and historical feedback showed that the weighting of the scope of impact dimension was too low, leading to inaccurate routing in some feeds. learn (Appropriately increase the weight of the scope of influence). After dynamic adjustment, =0.40, =0.28, =0.20, =0.12, which is more suitable for the current scenario than the default configuration.

[0058] In step S3, based on the comprehensive quantitative evaluation results, the abnormal data is divided into the following triage levels: After the anomaly severity quantification assessment module completes the scoring of each dimension in step S2, it transmits the comprehensive scoring result to the intelligent triage decision module. Based on the comprehensive score S(e) and the preset triage threshold, this module divides the abnormal data into four triage levels: Level 1 (automatic correction), Level 2 (return for resubmission), Level 3 (manual approval), and Level 4 (emergency warning), and performs differentiated triage reporting processing.

[0059] Based on the comprehensive score and the preset triage threshold, abnormal data is divided into different triage levels, and differentiated triage reporting processing is performed according to the triage level. When the exception type is a format error and the comprehensive score does not exceed the first preset triage threshold T1, the corresponding triage level is level one; after the data correction is completed, the verification is re-executed, and after the verification is passed, the report is submitted according to the normal process; that is, the correction suggestion agent is called to automatically generate a correction plan and execute it. After the correction is completed, the process re-enters step S1 for verification, and after the verification is passed, the report is submitted according to the normal process. When the anomaly type is data missing or logical error, and the comprehensive score is greater than the first preset triage threshold T1 but less than the second preset triage threshold T2, the corresponding triage level is level two; a return work order is generated and pushed to the business system, and it is resubmitted after the business system corrects it; that is, a return work order (including the reason for the anomaly and the correction instructions) is generated and pushed to the business system, and it is resubmitted after the business system corrects it. When the data items involved in the abnormal data to be reported involve preset core regulatory indicators, and the comprehensive score is greater than the second preset diversion threshold T2 but less than the third preset diversion threshold T3, the corresponding diversion level is level three; generate a manual approval task and push it to the approver, and continue the reporting process after waiting for approval; When the data items involved in the abnormal data to be reported are preset major risk indicators and the comprehensive score is greater than the third preset diversion threshold T3, the corresponding diversion level is level four; while performing manual approval, an emergency warning report is generated first and reported to the regulatory department to achieve dual-track parallel processing; among them, the first preset diversion threshold T1 is less than the second preset diversion threshold T2, and the second preset diversion threshold T2 is less than the third diversion threshold T3; and the first preset diversion threshold T1, the second preset diversion threshold T2, and the third preset diversion threshold T3 are dynamically determined according to the abnormal score distribution characteristics, reporting timeliness requirements, and the availability of corrective resources.

[0060] The triage thresholds T1, T2, and T3 determine the grading boundaries for abnormal data, and are dynamically adjusted based on the current batch data distribution characteristics and business scenario. Influencing factors: Anomaly score distribution characteristics: Based on the comprehensive score distribution of the current batch of anomaly data, the threshold position is automatically adjusted to optimize the diversion ratio; Reporting timeliness requirements: The urgency of the current batch of reports affects the tightness of thresholds at each level; Adjustment resource availability: The currently available adjustment manpower and automatic adjustment resources affect the setting of the first-level threshold; Dynamically determine the expression: Suppose that the dynamic values ​​of each shunt threshold are determined by the following formula: T1= [T] 1，base +ξ1·Score dist +ξ2·Resource adj ξ3·Urgency factor ] T2= [T] 2，base +ξ4·Score dist + ξ5·Urgency factor ] T3= [T] 3，base +ξ6·Score dist ] Wherein: T 1，base T 2，base T 3，base These are the base thresholds for the first preset traffic splitting threshold, the second preset traffic splitting threshold, and the third preset traffic splitting threshold (default values ​​are 0.25, 0.50, and 0.75, respectively); Score dist Adjust the distribution characteristics of the current batch of abnormal scores using a factor (appropriately raise the threshold when the overall scores are too high); Resource adj To adjust the resource availability adjustment factor (appropriately reduce T1 when resources are sufficient, allowing more anomalies to enter the automatic correction channel); Urgency factor The reporting timeliness requirement factor (when timeliness is tight, the thresholds at each level should be appropriately lowered so that more anomalies can be processed with higher priority); ξ1~ξ6 are the corresponding adjustment coefficients; (·) is the sigmoid normalization function, ensuring that the threshold is within the interval (0,1) and satisfies T1. <T2<T3。

[0061] For example: The overall anomaly score for the current batch is low (Score) dist Bias to adjust threshold downwards), and correct for insufficient resources (Resource) adj =0.2), but the reporting timeframe is tight (Urgency) factor =0.3). After dynamic adjustment, T1=0.20, T2=0.45, T3=0.72, allowing more anomalies to enter the automatic correction channel (expanding the first-level coverage), while anomalies with higher urgency are more likely to trigger higher-level diversion.

[0062] Preferably, step S4 can also be used to establish an anomaly diversion feedback learning mechanism; Record the decision parameters and actual processing results for each triage; optimize the weight configuration for severity quantification assessment through feedback learning agent; and dynamically adjust the judgment threshold for each triage level.

[0063] This invention acquires insurance industry data to be reported, and uses a multi-agent system to identify anomalies in the data, generating anomaly identification information. This information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. Based on this information, the severity of the anomalies is quantitatively assessed for each data point, considering four dimensions: risk level, scope of impact, correction difficulty, and historical frequency. This yields a comprehensive score for each data point. Based on the comprehensive score and a preset triage threshold, the data is categorized into different triage levels, and differentiated triage reporting processing is performed according to each level. This four-level triage mechanism changes the existing unified processing model for anomaly data, avoiding efficiency losses and delays caused by a "one-size-fits-all" approach. It effectively solves the problem of low reliability in processing insurance industry regulatory reporting data due to existing technologies, thus significantly improving the reliability of such processing.

[0064] The multi-agent system of this invention includes a rule verification agent, an anomaly detection agent, and a trend comparison agent. Anomaly types include format anomalies, integrity anomalies, business logic anomalies, statistical anomalies, and trend anomalies. The multi-agent system identifies anomalies in the data to be reported, generating anomaly identification information. After the rule verification agent, anomaly detection agent, and trend comparison agent have completed their respective verifications, cross-validation is performed among them. This not only improves the reliability of anomaly identification and detection but also allows for continuous optimization of the diversion decision parameters based on historical processing results through a feedback learning mechanism. Compared to the static configuration schemes of existing technologies, this system has stronger adaptive capabilities.

[0065] The technical solution of this invention acquires abnormal data to be reported, and obtains the risk level score, impact range score, correction difficulty score, and historical frequency score of the abnormal data to be reported. The comprehensive score of each abnormal data to be reported is obtained by weighted summation of the risk level score, impact range score, correction difficulty score, and historical frequency score. By introducing a multi-dimensional (risk level, impact range, correction difficulty, historical frequency) comprehensive quantitative evaluation model in the field of insurance regulatory reporting, a scientific quantitative basis is provided for diversion decision-making, overcoming the limitations of existing technologies that rely on experience-based judgment.

[0066] In this invention, the triage levels are divided into four levels from mild to severe anomalies: Level 1, Level 2, Level 3, and Level 4. Based on the comprehensive score and the preset triage threshold, abnormal data is divided into different triage levels, and differentiated triage reporting processing is performed according to the triage level. When the triage level is Level 4, an emergency early warning channel is set up so that abnormal data involving major risk indicators can trigger manual intervention and early warning reporting in a timely manner, meeting the timeliness requirements of regulatory reporting.

[0067] Example 2 like Figure 2 As shown, the present invention also provides an insurance industry regulatory reporting and diversion system based on anomaly types, comprising: The data reporting access module acquires insurance industry data to be reported, and uses a multi-agent system to identify anomalies in the data to be reported, generating anomaly identification information for the data to be reported. The anomaly identification information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. The multi-agent verification and anomaly severity quantification assessment module quantifies the severity of anomalies in the data to be reported based on the anomaly identification information of the data to be reported. It quantifies each abnormal data to be reported from four dimensions: risk level, scope of impact, difficulty of correction, and historical frequency, and obtains a comprehensive score for each abnormal data to be reported. The intelligent triage decision module classifies abnormal data into different triage levels based on the comprehensive score and preset triage thresholds, and performs differentiated triage reporting processing according to the triage level.

[0068] The data reporting and access module is located in the data input layer, while the multi-agent verification and anomaly severity quantification assessment module (i.e., the multi-agent verification module and the anomaly severity quantification assessment module) is located in the core processing layer, and the intelligent triage decision module is also located in the core processing layer. Preferably, it may also include an output processing layer and a feedback learning layer. The output processing layer includes an automatic correction channel for triage level 1, a return and re-reporting channel for triage level 2, a manual approval channel for triage level 3, and an emergency warning channel for triage level 4. The feedback learning layer is mainly used to learn from the evaluation parameters and triage thresholds in the anomaly severity quantification assessment module to continuously optimize the parameters.

[0069] This invention acquires insurance industry data to be reported, and uses a multi-agent system to identify anomalies in the data, generating anomaly identification information. This information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. Based on this information, the severity of the anomalies is quantitatively assessed for each data point, considering four dimensions: risk level, scope of impact, correction difficulty, and historical frequency. This yields a comprehensive score for each data point. Based on the comprehensive score and a preset triage threshold, the data is categorized into different triage levels, and differentiated triage reporting processing is performed according to each level. This four-level triage mechanism changes the existing unified processing model for anomaly data, avoiding efficiency losses and delays caused by a "one-size-fits-all" approach. It effectively solves the problem of low reliability in processing insurance industry regulatory reporting data due to existing technologies, thus significantly improving the reliability of such processing.

[0070] The multi-agent system of this invention includes a rule verification agent, an anomaly detection agent, and a trend comparison agent. Anomaly types include format anomalies, integrity anomalies, business logic anomalies, statistical anomalies, and trend anomalies. The multi-agent system identifies anomalies in the data to be reported, generating anomaly identification information. After the rule verification agent, anomaly detection agent, and trend comparison agent have completed their respective verifications, cross-validation is performed among them. This not only improves the reliability of anomaly identification and detection but also allows for continuous optimization of the diversion decision parameters based on historical processing results through a feedback learning mechanism. Compared to the static configuration schemes of existing technologies, this system has stronger adaptive capabilities.

[0071] The technical solution of this invention acquires abnormal data to be reported, and obtains the risk level score, impact range score, correction difficulty score, and historical frequency score of the abnormal data to be reported. The comprehensive score of each abnormal data to be reported is obtained by weighted summation of the risk level score, impact range score, correction difficulty score, and historical frequency score. By introducing a multi-dimensional (risk level, impact range, correction difficulty, historical frequency) comprehensive quantitative evaluation model in the field of insurance regulatory reporting, a scientific quantitative basis is provided for diversion decision-making, overcoming the limitations of existing technologies that rely on experience-based judgment.

[0072] In this invention, the triage levels are divided into four levels from mild to severe anomalies: Level 1, Level 2, Level 3, and Level 4. Based on the comprehensive score and the preset triage threshold, abnormal data is divided into different triage levels, and differentiated triage reporting processing is performed according to the triage level. When the triage level is Level 4, an emergency early warning channel is set up so that abnormal data involving major risk indicators can trigger manual intervention and early warning reporting in a timely manner, meeting the timeliness requirements of regulatory reporting.

[0073] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for diverting regulatory reports in the insurance industry based on anomaly types, characterized in that, include: The system acquires data to be reported from the insurance industry, performs anomaly identification on the data through a multi-agent system, and generates anomaly identification information for the data to be reported. The anomaly identification information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. Based on the anomaly identification information of the data to be reported, the severity of the anomalies of the data to be reported is quantitatively assessed. The anomalies are evaluated from four dimensions: risk level, scope of impact, difficulty of correction, and historical frequency, and a comprehensive score is obtained for each anomaly data to be reported. Based on the comprehensive score and the preset diversion threshold, abnormal data is divided into different diversion levels, and differentiated diversion reporting processing is performed according to the diversion level.

2. The method for diverting insurance industry regulatory reports based on anomaly types according to claim 1, characterized in that, The multi-agent system includes a rule verification agent, an anomaly detection agent, and a trend comparison agent; the anomaly types include format anomalies, integrity anomalies, business logic anomalies, statistical anomalies, and trend anomalies. Anomaly identification is performed on the data to be reported using a multi-agent system, generating anomaly identification information for the data to be reported, including: The rule verification agent is used to verify the format compliance of the data to be submitted. Referring to regulatory documents, it transforms the regulatory system into executable verification rules to verify the field format, field integrity and logical consistency. The anomaly detection agent is used to identify anomalies at the business logic level in the data to be reported, including regulatory red line detection, statistical anomaly detection, and data quality detection. The trend comparison agent is used to perform vertical and horizontal trend analysis on the data to be reported. Through month-on-month change detection, year-on-year change detection, and peer comparison detection, it identifies abnormal changes in time series or among similar institutions. After the rule verification agent, anomaly detection agent, and trend comparison agent have completed their respective verifications, cross-validation is performed among the rule verification agent, anomaly detection agent, and trend comparison agent.

3. The method for diverting insurance industry regulatory reports based on anomaly types according to claim 2, characterized in that, The cross-validation among the rule verification agent, anomaly detection agent, and trend comparison agent is as follows: For logical contradictions marked by the rule verification agent, the anomaly detection agent further verifies whether the data to be reported involves regulatory red line indicators; the trend comparison agent confirms whether the logical contradictions in the data to be reported are occasional errors or systemic trend anomalies. If all three agents point to the same problem, it is directly confirmed; if only a single agent marks an anomaly, the priority is reduced or it is submitted for manual review.

4. The method for diverting insurance industry regulatory reports based on anomaly types according to claim 1, characterized in that, Based on the anomaly identification information of the data to be reported, the severity of the anomalies in the data to be reported is quantitatively assessed. Each piece of abnormal data is quantitatively assessed from four dimensions: risk level, scope of impact, difficulty of correction, and historical frequency, resulting in a comprehensive score for each piece of abnormal data to be reported. Obtain abnormal pending data, and obtain the risk level score, impact scope score, correction difficulty score and historical frequency score of the abnormal pending data; The comprehensive score for each abnormal data point to be reported is obtained by weighting and summing the risk level score, impact scope score, correction difficulty score, and historical frequency score. The weighting of different dimensions is dynamically determined based on the regulatory focus shift factor, the data quality characteristics of the batch to be reported, and the feedback of historical diversion effects.

5. A method for diverting insurance industry regulatory reports based on anomaly types according to claim 4, characterized in that, The risk level score is determined based on whether the data items involved in the abnormal data to be reported are preset core regulatory indicators and the real-time weight attributes of the data item nodes in the knowledge graph where the data items are located; the specific calculation method of the risk level score R(e) is as follows: R(e) = [W base (i) + ·C factor (i) + ·Dev factor (i)] Among them: W base (i) indicates whether data item i is a basic regulatory weight for a preset core regulatory indicator; C factor (i) represents the regulatory dynamic attention factor corresponding to data item i; Dev factor (i) represents the current deviation factor corresponding to data item i; , These are the weight coefficients of the regulatory dynamic attention factor and the current deviation factor, respectively. This is the sigmoid normalization function.

6. A method for diverting insurance industry regulatory reports based on anomaly types according to claim 4, characterized in that, The impact scope score is determined based on the percentage of records affected by the abnormal data to be reported and the number of reporting dimensions involved; specifically, the impact scope score I(e) is calculated as follows: I(e) =min{1, ·P ratio + ·ln(1 + Dim count ) + ·Layer depth } Where: P ratio Dim represents the number of affected records for abnormal pending data submissions divided by the total number of records in the batch. count This refers to the number of reporting dimensions involved in the abnormal pending data; Layer depth The depth of the hierarchy is affected by abnormal pending data. , , These are the weighting coefficients for the percentage of records, the number of reporting dimensions, and the influence on the depth of the hierarchy, respectively.

7. A method for diverting insurance industry regulatory reports based on anomaly types according to claim 4, characterized in that, The correction difficulty score is dynamically determined based on the correction complexity of the anomaly type, the current system resource status, and the reporting timeliness requirements; specifically, the correction difficulty score D(e) is calculated as follows: D(e) = [ D base (i) + ·Peak factor (i)+ ·Ext dep，factor (i) ] Where: D base (i) represents the normalized value of the basic modified complexity for the anomaly type corresponding to data item i; Peak factor (i) represents the peak reporting period factor corresponding to data item i; Ext dep，factor (i) represents the external dependency factor corresponding to data item i; , These are the weighting coefficients for the peak reporting period factor and the external dependence factor, respectively. This is the sigmoid normalization function.

8. A method for diverting insurance industry regulatory reports based on anomaly types according to claim 4, characterized in that, The historical frequency score is dynamically determined based on the frequency of occurrence of similar anomalies in the current batch and historical reporting cycles; specifically, the historical frequency score H(e) is calculated as follows: H(e) =min{1, ·Batch ratio + ·Avg(Hist freq )·exp(- ·Count excessive )} Among them: Batch ratio Hist represents the percentage of similar anomalies in the total number of anomalies in the current batch; freq This is an array representing the monthly trigger counts of similar anomalies over the past 6 months; Avg(·) is used to calculate the average; Count excessive This refers to the number of months in which the historical trigger count exceeded the warning threshold; , These are the weighting coefficients for the current batch and the historical reporting period, respectively; ξ is the attenuation coefficient.

9. A method for diverting insurance industry regulatory reports based on anomaly types according to claim 1, characterized in that, The traffic triage levels are divided into four levels, from mild to severe, based on the severity of the anomaly. According to the comprehensive score and preset triage thresholds, abnormal data is classified into different triage levels, and differentiated triage reporting processing is performed according to the corresponding triage level. When the exception type is a format error and the comprehensive score does not exceed the first preset triage threshold, the corresponding triage level is level one; after the data correction is completed, the verification is re-executed, and after the verification is passed, the report is submitted according to the normal process. When the anomaly type is missing data or logical error, and the comprehensive score is greater than the first preset triage threshold but less than the second preset triage threshold, the corresponding triage level is level two. A return work order is generated and pushed to the business system. The work order is then resubmitted after the business system makes corrections. When the data items involved in the abnormal data to be reported involve preset core regulatory indicators, and the comprehensive score is greater than the second preset diversion threshold but less than the third preset diversion threshold, the corresponding diversion level is level three; Generate manual approval tasks and push them to the approvers. After approval, continue the submission process. When the data items involved in the abnormal data to be reported are preset major risk indicators and the comprehensive score is greater than the third preset diversion threshold, the corresponding diversion level is level four; while performing manual approval, an emergency warning report is generated first and reported to the regulatory department to achieve dual-track parallel processing; among them, the first preset diversion threshold is less than the second preset diversion threshold, and the second preset diversion threshold is less than the third diversion threshold; and the first preset diversion threshold, the second preset diversion threshold, and the third preset diversion threshold are dynamically determined according to the abnormal score distribution characteristics, reporting timeliness requirements, and the availability of corrective resources.

10. An insurance industry regulatory reporting and triage system based on anomaly types, characterized in that, include: The data reporting access module acquires insurance industry data to be reported, and uses a multi-agent system to identify anomalies in the data to be reported, generating anomaly identification information for the data to be reported. The anomaly identification information includes anomaly identifier, anomaly type, anomaly cause, scope of impact, and source agent. The multi-agent verification and anomaly severity quantification assessment module quantifies the severity of anomalies in the data to be reported based on the anomaly identification information of the data to be reported. It quantifies each abnormal data to be reported from four dimensions: risk level, scope of impact, difficulty of correction, and historical frequency, and obtains a comprehensive score for each abnormal data to be reported. The intelligent triage decision module classifies abnormal data into different triage levels based on the comprehensive score and preset triage thresholds, and performs differentiated triage reporting processing according to the triage level.

Images