A civil affairs fund three-account consistency intelligent comparison and closed-loop supervision system and method

Abstract

The application discloses a civil affairs fund three-account consistency intelligent comparison and closed-loop supervision system and method, and belongs to the technical field of civil affairs business. Through multi-source heterogeneous data fusion and intelligent entity analysis comparison, the automatic cleaning, matching and consistency comparison of business, financial and statistical three-account data are realized. Traditional manual spot checks are changed into full-quantity real-time supervision, and the coverage and efficiency are greatly improved. A multi-dimensional abnormal behavior detection module identifies abnormalities from individual and group levels, automatically infers reasons and quantifies risk levels in combination with a knowledge graph and a graph neural network, and provides accurate decision support for supervision. An automatic closed-loop rectification module realizes automatic generation, intelligent distribution, real-time monitoring and supervision upgrading of task sheets, and forms a management closed loop in combination with credit score feedback. A blockchain technology chains key results, abnormal conclusions and rectification records for storage, ensures that electronic evidence cannot be tampered with, builds a transparent and reliable supervision environment, and eliminates information asymmetry risks.

Description

Technical Field

[0001] This invention belongs to the field of civil affairs business technology, specifically relating to a system and method for intelligent comparison and closed-loop supervision of the consistency of civil affairs funds' three accounts. Background Technology

[0002] Currently, the management process for civil affairs funds is typically divided into three independent stages: the application stage (such as eligibility verification and amount due), the disbursement stage (such as fund allocation and bank disbursement), and the statistical reporting stage (such as data aggregation and reporting to higher-level departments). These three stages operate on independent information systems developed by different vendors, and some grassroots units even rely on manual ledgers, resulting in three relatively independent accounts: the "business payable account," the "financial payable account," and the "statistical details account."

[0003] Existing regulatory methods primarily rely on manual spot checks by auditing departments after the fact. This approach is not only labor-intensive and inefficient, but also suffers from significant delays in risk detection, failing to achieve normalized, end-to-end supervision of fund operations. Therefore, how to solve the key technical problem urgently needing to be addressed in this field is how to intelligently compare multi-source heterogeneous data, accurately identify abnormal behavior, and implement closed-loop supervision. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, this application provides an intelligent comparison and closed-loop supervision system and method for ensuring consistency among the three accounts of civil affairs funds.

[0005] The first aspect of this application proposes an intelligent comparison and closed-loop supervision system for the consistency of the three accounts of civil affairs funds, including:

[0006] The multi-source heterogeneous data fusion module is used to connect to and acquire the original data from the business system, financial system and statistical system, which are of different sources and have different formats. The original data is cleaned through a three-level verification mechanism including basic field verification, business rule verification and cross-system consistency verification. The cleaned data is mapped into a unified three-dimensional data structure using a composite primary key strategy of ID number and timestamp.

[0007] The intelligent entity parsing and comparison module is connected to the multi-source heterogeneous data fusion module. It is used to parse and match entities in the three-dimensional data structure by adopting a hybrid matching strategy that combines precise matching and fuzzy matching. The fuzzy matching strategy integrates character edit distance and semantic vector similarity based on deep neural network, and solves the multi-candidate matching problem through entity disambiguation technology. Finally, the matching entities are compared in multiple dimensions using the ID number as the unique identifier.

[0008] The multidimensional abnormal behavior detection and reasoning module is connected to the intelligent entity parsing and comparison module. It is used to perform hierarchical detection on abnormal data in the consistency comparison results based on individual-level time series analysis and group-level clustering detection. The group-level clustering detection uses an improved density clustering algorithm to identify group abnormal patterns and combines it with the built-in knowledge graph in the civil affairs field to infer the cause of the abnormality and output the risk level using a graph neural network model.

[0009] The automated closed-loop rectification module, connected to the multi-dimensional abnormal behavior detection and reasoning module, is used to automatically generate rectification task orders based on the cause and risk level of the abnormality, using a combination of templates and dynamic content. It also intelligently distributes the rectification tasks according to the responsibilities, workload, and historical response speed of the personnel in charge, monitors the execution status of the rectification tasks and implements timeout alarms and supervision escalation, and automatically verifies the rectification results before updating the credit score of the personnel in charge, thus forming a closed-loop management system.

[0010] In some embodiments, the multi-source heterogeneous data fusion module includes:

[0011] The data access layer is configured with various data interface adapters for accessing data from different file formats and database sources.

[0012] The data cleaning layer is used to implement the three-level verification mechanism, which includes: Level 1 basic field verification, used to verify the ID number format according to national standards and verify whether the amount field is within a preset reasonable range; Level 2 business rule verification, used to verify whether the disbursement date is after the business application date and whether the subsidy type for the same service recipient is repeated in the same period; Level 3 cross-system consistency verification, used to compare whether the deviation between the total amount of financial disbursement and the total amount of business application is within a preset acceptable range.

[0013] The data standardization processing layer is used to map heterogeneous data from different systems into a unified three-dimensional data structure using a composite primary key strategy that combines ID card number and timestamp.

[0014] In some embodiments, the intelligent entity parsing and comparison module includes:

[0015] The precise matching unit is used to build a hash index for matching the standardized ID card numbers, and to establish a two-way verification mechanism for successfully matched records to verify whether the issuance date is later than the application date;

[0016] A fuzzy matching unit, connected to the precise matching unit, is used to calculate a comprehensive score for records that fail to be precisely matched, using a hybrid model that integrates edit distance and semantic vector similarity from a deep neural network, and then matching them using a regional dialect feature database.

[0017] The entity disambiguation unit, connected to the fuzzy matching unit, is used to construct a multi-dimensional feature space containing amount difference, time interval, region code, subsidy type and historical matching confidence when there are multiple matching candidates, and to use a classification algorithm to identify the most likely matching entity from the candidates.

[0018] The consistency comparison engine, connected to the entity disambiguation unit, is used to perform direct comparison of amounts and time series analysis of disbursement time on matched entities, and to use an incremental algorithm to build a cache index of historically successfully matched entities, so as to comprehensively determine the difference level based on the comparison results.

[0019] In some embodiments, the multidimensional abnormal behavior detection and reasoning module includes:

[0020] The individual-level time series analysis unit is used to construct a recurrent neural network time series prediction model with an attention mechanism based on the historical disbursement amount sequence of the service recipient, predict the expected amount for the current period, and set a dynamic anomaly judgment threshold that is automatically adjusted according to recent data fluctuations. When the deviation between the actual disbursement amount and the model prediction value exceeds the dynamic threshold, an individual anomaly alarm is triggered.

[0021] The group-level clustering detection unit is used to employ a density-based spatial clustering algorithm to cluster service recipients or operators with similar disbursement behaviors into clusters to identify abnormal group patterns by using the dispersion of amount, concentration of disbursement time, distribution of subsidy type, and standard deviation of operation time as feature vectors.

[0022] The intelligent reasoning engine, connected to the individual-level time series analysis unit and the group-level clustering detection unit, is used to automatically focus on knowledge nodes related to the anomaly detection task based on a knowledge graph that integrates subsidy policies, operating procedures, and regional characteristic entities, and infers the cause of the anomaly through a graph neural network model.

[0023] The risk level assessment unit, connected to the intelligent reasoning engine, is used to employ a multi-factor weighted scoring model to assign dynamic weights to the magnitude of the amount difference, the duration of the anomaly, the number of people involved, and historical credit factors, calculate a comprehensive score, and classify the risk into three levels: low, medium, and high.

[0024] In some embodiments, the automated closed-loop rectification module includes:

[0025] The intelligent task order generation unit is used to retrieve the corresponding template from the standardized task order template library based on the detected anomaly type and risk level, dynamically fill in the risk level, anomaly type, and evidence chain, and match and generate targeted rectification operation guidelines with the level of detail increasing with the risk level from the knowledge base.

[0026] A multi-channel task distribution unit, connected to the intelligent task order generation unit, is used to intelligently distribute tasks according to the responsibilities of the personnel, current load, and historical response speed, through internal business system messages, SMS, or email, and ensure that the task status is synchronized in real time across different systems.

[0027] The real-time monitoring dashboard unit is connected to the multi-channel task distribution unit and is used to dynamically display the overall status of rectification tasks from the dimensions of geographical distribution, time trend and responsible entity. When the proportion of unprocessed tasks exceeds the preset ratio or the delay of high-risk tasks exceeds the specified time limit, an alarm is automatically issued and the supervision escalation logic is executed to escalate the responsibility to the next higher level of management personnel.

[0028] The rectification verification unit, connected to the real-time monitoring dashboard unit, is used to automatically verify the validity of the electronic signature in the rectification voucher uploaded by the person in charge, compare the consistency between the recovered or reissued amount and the original abnormal amount, update the credit score of the person in charge based on the timeliness of task processing and the completion of rectification, and feed the rectification results back to the system to optimize the subsequent task distribution strategy and risk detection model.

[0029] In some embodiments, a trusted evidence storage module is also included. The trusted evidence storage module is connected to the intelligent entity parsing and comparison module, the multi-dimensional abnormal behavior detection and reasoning module, and the automated closed-loop rectification module, respectively. It is used to generate digital fingerprints of key comparison results, abnormal detection conclusions, and rectification records and store them on the blockchain to construct an immutable electronic evidence chain.

[0030] In some embodiments, the trusted evidence storage module includes:

[0031] The data fingerprint generation system is used to serialize structured data that needs to be stored for evidence, and then combine it with a timestamp accurate to milliseconds and a random number to generate a unique digital fingerprint through a high-strength cryptographic hash algorithm.

[0032] The smart contract notarization engine is connected to the data fingerprint generation system and is used to write the digital fingerprint into the distributed ledger by calling the smart contract deployed on the consortium blockchain and through a multi-party consensus mechanism.

[0033] A cross-chain verification gateway, connected to the smart contract notarization engine, is used to provide a standardized verification interface. It verifies the integrity and authenticity of the data by recalculating the hash value of the data to be verified and comparing it with the hash value stored on the chain.

[0034] The audit trail subsystem is connected to the data fingerprint generation system, the smart contract evidence storage engine, and the cross-chain verification gateway, respectively. It is used to record all evidence storage operations and leave an indelible record on the blockchain, thus constructing a complete traceable evidence chain.

[0035] In some embodiments, the data interface adapters configured in the data access layer of the multi-source heterogeneous data fusion module include an Excel parser, a CSV parser, a JSON parser, and connectors supporting MySQL and Oracle databases. Each adapter has a built-in automatic data format detection mechanism for intelligently identifying the encoding format and delimiter type of the source data.

[0036] In some embodiments, in the fuzzy matching unit of the intelligent entity parsing and comparison module, the hybrid model that integrates edit distance and semantic vector similarity of deep neural network weights the edit distance similarity and semantic vector similarity according to preset weights to form a comprehensive similarity score, and dynamically adjusts the matching threshold according to historical matching results;

[0037] In the risk level assessment unit of the multidimensional abnormal behavior detection and reasoning module, the multi-factor weighted scoring model assigns dynamically adjustable weights to different risk factors and classifies risks into three levels—low, medium, and high—based on the weighted comprehensive score. It also generates a detailed chain of evidence for high-risk events, including abnormal amounts, involved parties, and time points.

[0038] Secondly, this application proposes a method for intelligent comparison and closed-loop supervision of the consistency of the three accounts of civil affairs funds, including the following steps:

[0039] Multi-source heterogeneous data fusion steps: Connect and obtain the original data from the business system, financial system and statistical system, which are of different sources and have different formats. Clean the original data from the three accounts through a three-level verification mechanism including basic field verification, business rule verification and cross-system consistency verification. Then, use a composite primary key strategy of ID number combined with timestamp to map the cleaned data into a unified three-dimensional data structure.

[0040] Intelligent entity parsing and comparison steps: A hybrid matching strategy combining exact matching and fuzzy matching is adopted to parse and match entities in the three-dimensional data structure. The fuzzy matching strategy integrates character edit distance and semantic vector similarity based on deep neural network, and solves the multi-candidate matching problem through entity disambiguation technology. Finally, the matching entities are compared in multiple dimensions using the ID number as the unique identifier.

[0041] Multidimensional abnormal behavior detection and reasoning steps: Based on individual-level time series analysis and group-level clustering detection, the abnormal data in the consistency comparison results are subjected to hierarchical detection. The group-level clustering detection uses an improved density clustering algorithm to identify group abnormal patterns, and combines the built-in knowledge graph in the civil affairs field to infer the cause of the abnormality and output the risk level using a graph neural network model.

[0042] Automated closed-loop rectification steps: Based on the cause of the anomaly and the risk level, rectification task sheets are automatically generated by combining templates and dynamic content. They are intelligently distributed according to the responsibilities, workload, and historical response speed of the personnel in charge. The execution status of the rectification tasks is monitored and timeout alarms and supervision escalation are implemented. The rectification results are automatically verified and the credit score of the personnel in charge is updated, thus forming a closed-loop management system.

[0043] Thirdly, this application proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0044] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0045] The beneficial effects of this invention are:

[0046] By constructing a multi-source heterogeneous data fusion module and an intelligent entity parsing and comparison module, the system solves the problem of the inability to automatically compare business, financial, and statistical data due to their different sources and formats. The system can automatically complete data cleaning, entity matching, and multi-dimensional consistency comparison, transforming the traditional, slow, and inefficient regulatory model that relies on manual spot checks into a full-scale, real-time automated intelligent regulatory system, greatly improving regulatory coverage and efficiency. Through a multi-dimensional abnormal behavior detection and reasoning module, this invention can not only identify individual-level monetary differences but also discover hidden operational anomalies at the group level through clustering algorithms. More importantly, the intelligent reasoning engine, built by combining knowledge graphs and graph neural networks, can automatically infer the reasons behind anomalies and output quantified risk levels, providing precise and in-depth decision support for regulatory departments. Through an automated closed-loop rectification module, this invention seamlessly connects anomaly detection with problem handling. The system can automatically generate task orders, intelligently distribute them to responsible persons, monitor progress in real time, and implement supervision and escalation. The verification of rectification results and the feedback mechanism of credit scoring create positive or negative incentives for the personnel in charge, plugging risk loopholes in the management mechanism. It utilizes blockchain technology to store key comparison results, abnormal conclusions, and rectification records on the chain. The immutability of its hash value and timestamp provides a solid legal basis for subsequent auditing and accountability, effectively solving the problems of easy tampering and difficulty in identifying electronic evidence. This constructs a transparent and credible regulatory environment, fundamentally eliminating the space for illegal operations exploiting information asymmetry. Attached Figure Description

[0047] Figure 1 This is a system structure block diagram of the present invention. Detailed Implementation

[0048] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein; rather, these embodiments are provided so that a more thorough understanding of the invention can be achieved and that the full scope of the invention can be conveyed to those skilled in the art.

[0049] Firstly, this application proposes an intelligent comparison and closed-loop supervision system for the consistency of the three accounts of civil affairs funds, such as... Figure 1 As shown, it includes:

[0050] The multi-source heterogeneous data fusion module is used to connect to and acquire the original data from the business system, financial system and statistical system, which are of different sources and have different formats. The original data is cleaned through a three-level verification mechanism including basic field verification, business rule verification and cross-system consistency verification. The cleaned data is mapped into a unified three-dimensional data structure using a composite primary key strategy of ID number and timestamp.

[0051] In some embodiments, the multi-source heterogeneous data fusion module includes:

[0052] The data access layer is configured with multiple data interface adapters for accessing data from different file formats and database sources;

[0053] In some embodiments, the data interface adapters configured in the data access layer of the multi-source heterogeneous data fusion module include an Excel parser, a CSV parser, a JSON parser, and connectors supporting MySQL and Oracle databases. Each adapter has a built-in automatic data format detection mechanism for intelligently identifying the encoding format and delimiter type of the source data.

[0054] The data cleaning layer is used to implement the three-level verification mechanism, which includes: Level 1 basic field verification, used to verify the ID number format according to national standards and verify whether the amount field is within a preset reasonable range; Level 2 business rule verification, used to verify whether the disbursement date is after the business application date and whether the subsidy type for the same service recipient is repeated in the same period; Level 3 cross-system consistency verification, used to compare whether the deviation between the total amount of financial disbursement and the total amount of business application is within a preset acceptable range.

[0055] The data standardization processing layer is used to map heterogeneous data from different systems into a unified three-dimensional data structure using a composite primary key strategy that combines ID card number and timestamp.

[0056] This module uses a data import interface that supports multiple file formats and database connection methods to clean, verify, and standardize heterogeneous data from the three major business, financial, and statistical processes, and establishes a unified data structure mapping with the ID number of civil affairs service recipients as the core primary key.

[0057] The multi-source heterogeneous data fusion module adopts a layered architecture design to achieve fully automated processing of civil affairs fund data across the entire process. At the data access layer, the system is configured with various data interface adapters, including file parsers for Excel, CSV, and JSON, as well as connectors supporting mainstream databases such as MySQL and Oracle. Each adapter has a built-in automatic data format detection mechanism that can intelligently identify the encoding format and delimiter type of the source data.

[0058] At the data cleaning layer, the system implements a three-level verification mechanism: The first level performs basic field verification, such as verifying the ID number format according to national standards and checking whether the amount field is within a reasonable range; the second level performs business rule verification, such as ensuring that the disbursement date is after the business application date and that the subsidy type for the same service recipient is not repeated in the same period; the third level performs cross-system consistency verification, such as comparing the total amount of financial disbursement with the total amount of business application to confirm whether the deviation is within the preset acceptable range.

[0059] Data standardization employs a composite primary key strategy of "ID number + timestamp". The specific implementation process is as follows: First, ID number fields from different systems are uniformly converted to an 18-digit standard format; then, for each service object, a time partition key is generated according to a fixed business cycle (e.g., monthly); finally, a three-dimensional data mapping table is established with the standardized ID number and business cycle as the composite primary key. For records lacking key fields, the system will automatically process them according to pre-configured completion rules; for example, records lacking an issuance date will be assigned the default value of the last day of the month.

[0060] The intelligent entity parsing and comparison module is connected to the multi-source heterogeneous data fusion module. It is used to parse and match entities in the three-dimensional data structure by adopting a hybrid matching strategy that combines precise matching and fuzzy matching. The fuzzy matching strategy integrates character edit distance and semantic vector similarity based on deep neural network, and solves the multi-candidate matching problem through entity disambiguation technology. Finally, the matching entities are compared in multiple dimensions using the ID number as the unique identifier.

[0061] In some embodiments, the intelligent entity parsing and comparison module includes:

[0062] The precise matching unit is used to build a hash index for matching the standardized ID card numbers, and to establish a two-way verification mechanism for successfully matched records to verify whether the issuance date is later than the application date;

[0063] A fuzzy matching unit, connected to the precise matching unit, is used to calculate a comprehensive score for records that fail to be precisely matched, using a hybrid model that integrates edit distance and semantic vector similarity from a deep neural network, and then matching them using a regional dialect feature database.

[0064] The entity disambiguation unit, connected to the fuzzy matching unit, is used to construct a multi-dimensional feature space containing amount difference, time interval, region code, subsidy type and historical matching confidence when there are multiple matching candidates, and to use a classification algorithm to identify the most likely matching entity from the candidates.

[0065] The consistency comparison engine, connected to the entity disambiguation unit, is used to perform direct comparison of amounts and time series analysis of disbursement time on matched entities, and to use an incremental algorithm to build a cache index of historically successfully matched entities, so as to comprehensively determine the difference level based on the comparison results.

[0066] A hybrid model combining rule-based and machine learning is adopted. First, the ID card number is matched precisely. Then, a pre-trained name similarity model is used for fuzzy matching. Entity disambiguation technology is combined to determine the most likely matching entity. Finally, the three accounts are matched using the ID card number as the unique identifier.

[0067] The intelligent entity parsing and comparison module adopts a multi-level progressive hybrid matching strategy, and the specific implementation steps are as follows:

[0068] The precise matching layer first standardizes the ID card number field in the three accounts data, unifying it to the same format. Then, using improved hash indexing technology, an efficient lookup table is built for all ID card numbers to accelerate the precise matching process. The system also establishes a two-way verification mechanism; for example, after a successful match between the business account and the financial account, it further verifies whether the issuance date is later than the declaration date to ensure the correctness of the business logic.

[0069] Fuzzy Matching Layer: For records where exact matching fails, the system initiates a fuzzy matching process. This process employs a hybrid model that integrates text similarity and deep semantic understanding. Specifically, the system simultaneously calculates the edit distance of the name (measuring character-level similarity) and the similarity between the name's semantic vector extracted through a deep neural network (such as a bidirectional long short-term memory network), and combines the two with certain weights to form a comprehensive similarity score. To adapt to the language habits of different regions, the system also incorporates a regional dialect feature library when constructing features to address issues such as differences in the transliteration of names among ethnic minorities. The matching threshold is dynamically adjusted based on historical matching data to optimize matching performance.

[0070] Furthermore, in some embodiments, in the fuzzy matching unit of the intelligent entity parsing and comparison module, the hybrid model that integrates edit distance and semantic vector similarity of deep neural network weights the edit distance similarity and semantic vector similarity according to preset weights to form a comprehensive similarity score, and dynamically adjusts the matching threshold according to historical matching results;

[0071] Entity Disambiguation Layer: When multiple possible matching candidates exist, the system enters the entity disambiguation stage. This stage constructs a multi-dimensional feature space, comprehensively considering factors such as amount differences, time intervals, region codes, subsidy types, and historical match confidence. Subsequently, an optimized classification algorithm (such as extreme gradient boosting) is used to analyze these features to accurately identify the most likely correct matching entity from multiple candidates, effectively solving the confusion problem caused by data duplication or similarity.

[0072] Consistency Comparison Engine: After completing entity matching, the system performs multi-dimensional consistency comparison. The comparison process includes not only direct comparison of amounts but also time-series analysis, such as determining the reasonableness of disbursement timing. Based on the comparison results of amount differences and time series data, the system comprehensively determines the severity of the discrepancy and initiates subsequent anomaly cause reasoning processes. The entire comparison process employs an incremental algorithm, establishing a cache index for historically successfully matched entities to improve processing efficiency.

[0073] The multidimensional abnormal behavior detection and reasoning module is connected to the intelligent entity parsing and comparison module. It is used to perform hierarchical detection on abnormal data in the consistency comparison results based on individual-level time series analysis and group-level clustering detection. The group-level clustering detection uses an improved density clustering algorithm to identify group abnormal patterns and combines it with the built-in knowledge graph in the civil affairs field to infer the cause of the abnormality and output the risk level using a graph neural network model.

[0074] In some embodiments, the multidimensional abnormal behavior detection and reasoning module includes:

[0075] The individual-level time series analysis unit is used to construct a recurrent neural network time series prediction model with an attention mechanism based on the historical disbursement amount sequence of the service recipient, predict the expected amount for the current period, and set a dynamic anomaly judgment threshold that is automatically adjusted according to recent data fluctuations. When the deviation between the actual disbursement amount and the model prediction value exceeds the dynamic threshold, an individual anomaly alarm is triggered.

[0076] The group-level clustering detection unit is used to employ a density-based spatial clustering algorithm to cluster service recipients or operators with similar disbursement behaviors into clusters to identify abnormal group patterns by using the dispersion of amount, concentration of disbursement time, distribution of subsidy type, and standard deviation of operation time as feature vectors.

[0077] The intelligent reasoning engine, connected to the individual-level time series analysis unit and the group-level clustering detection unit, is used to automatically focus on knowledge nodes related to the anomaly detection task based on a knowledge graph that integrates subsidy policies, operating procedures, and regional characteristic entities, and infers the cause of the anomaly through a graph neural network model.

[0078] The risk level assessment unit, connected to the intelligent reasoning engine, is used to employ a multi-factor weighted scoring model to assign dynamic weights to the magnitude of the amount difference, the duration of the anomaly, the number of people involved, and historical credit factors, calculate a comprehensive score, and classify the risk into three levels: low, medium, and high.

[0079] Furthermore, in the risk level assessment unit of the multi-dimensional abnormal behavior detection and reasoning module, the multi-factor weighted scoring model assigns dynamically adjustable weights to different risk factors and divides the risk into three levels—low, medium, and high—based on the weighted comprehensive score, generating a detailed chain of evidence for high-risk events, including abnormal amounts, involved parties, and time points.

[0080] It includes two levels: individual-level time series analysis and group-level cluster detection. Combined with an intelligent reasoning engine based on a knowledge base in the civil affairs field, it automatically infers the causes of anomalies and outputs the risk level.

[0081] The multidimensional abnormal behavior detection and reasoning module adopts a hierarchical detection architecture, and the specific implementation steps are as follows:

[0082] Individual-level time series analysis: For each civil affairs service recipient, the system constructs a time series prediction model (such as a recurrent neural network with an attention mechanism) based on their historical disbursement amount sequence. This model can learn the individual's normal disbursement pattern and predict the expected amount for the next period. Simultaneously, the system sets a dynamic anomaly detection threshold, which automatically adjusts with recent data fluctuations. When the deviation between the actual disbursement amount and the model's prediction exceeds this dynamic threshold, the system triggers an anomaly alert for that individual.

[0083] Group-level clustering detection: Beyond individual analysis, the system also uncovers potential risks at the group level. By employing an improved density clustering algorithm (such as density-based spatial clustering), the system groups service recipients or operators with similar disbursement behaviors into clusters. When calculating core distance, the algorithm comprehensively considers multiple dimensions of features, including the dispersion of monetary amounts, the concentration of disbursement time, the distribution of subsidy types, and the standard deviation of operation time, thereby effectively identifying groups of operational behaviors or beneficiaries with similar abnormal patterns.

[0084] Intelligent Reasoning Engine: The system incorporates a knowledge graph integrating knowledge from the civil affairs field, containing entities such as subsidy policies, operational procedures, and regional characteristics, as well as their interrelationships. Based on this knowledge graph, the system utilizes a graph neural network model for reasoning. This model can automatically focus on the knowledge nodes most relevant to the current anomaly detection task through a graph attention mechanism, thereby learning and inferring the potential causes behind abnormal behavior, such as policy implementation deviations, operational errors, or systemic violations.

[0085] Risk Level Assessment: Based on the combined results of individual and group anomaly detection, the system employs a multi-factor weighted scoring model to assess overall risk. This model assigns different weights to different risk factors (such as the magnitude of the amount difference, the duration of the anomaly, the number of people involved, and historical credit history), and these weights can be dynamically adjusted according to the actual application scenario and regulatory priorities. Ultimately, the system classifies risks into low, medium, and high levels based on the weighted comprehensive score, and generates detailed evidence chains for high-risk events.

[0086] The automated closed-loop rectification module, connected to the multi-dimensional abnormal behavior detection and reasoning module, is used to automatically generate rectification task orders based on the cause and risk level of the abnormality, using a combination of templates and dynamic content. It also intelligently distributes the rectification tasks according to the responsibilities, workload, and historical response speed of the personnel in charge, monitors the execution status of the rectification tasks and implements timeout alarms and supervision escalation, and automatically verifies the rectification results before updating the credit score of the personnel in charge, thus forming a closed-loop management system.

[0087] In some embodiments, the automated closed-loop rectification module includes:

[0088] The intelligent task order generation unit is used to retrieve the corresponding template from the standardized task order template library based on the detected anomaly type and risk level, dynamically fill in the risk level, anomaly type, and evidence chain, and match and generate targeted rectification operation guidelines with the level of detail increasing with the risk level from the knowledge base.

[0089] A multi-channel task distribution unit, connected to the intelligent task order generation unit, is used to intelligently distribute tasks according to the responsibilities of the personnel, current load, and historical response speed, through internal business system messages, SMS, or email, and ensure that the task status is synchronized in real time across different systems.

[0090] The real-time monitoring dashboard unit is connected to the multi-channel task distribution unit and is used to dynamically display the overall status of rectification tasks from the dimensions of geographical distribution, time trend and responsible entity. When the proportion of unprocessed tasks exceeds the preset ratio or the delay of high-risk tasks exceeds the specified time limit, an alarm is automatically issued and the supervision escalation logic is executed to escalate the responsibility to the next higher level of management personnel.

[0091] The rectification verification unit, connected to the real-time monitoring dashboard unit, is used to automatically verify the validity of the electronic signature in the rectification voucher uploaded by the person in charge, compare the consistency between the recovered or reissued amount and the original abnormal amount, update the credit score of the person in charge based on the timeliness of task processing and the completion of rectification, and feed the rectification results back to the system to optimize the subsequent task distribution strategy and risk detection model.

[0092] When financial risks are detected, a rectification task order with risk level and abnormal description is automatically generated, which is pushed to the responsible personnel through API connection to the business system, and provides real-time dashboard tracking of task status and timeout alarm function.

[0093] The steps for implementing the automated closed-loop rectification module are as follows:

[0094] Intelligent Task Order Generation Unit: The system automatically generates rectification task orders by combining templates with dynamic content. It intelligently populates the detected risk level, anomaly type, and relevant evidence chain (such as abnormal amount, involved parties, and time point) into a predefined standardized task order template. Simultaneously, based on the anomaly type, the system dynamically generates targeted rectification operation guidelines from the knowledge base, with the level of detail increasing as the risk level rises.

[0095] Multi-channel task distribution unit: After a task is generated, the system intelligently distributes it based on its priority. Priority is determined by both risk level and task processing timeliness. The system automatically selects the optimal distribution channel based on the responsibilities of the personnel in charge, current workload, and historical response speed, such as pushing notifications via internal business system messages, SMS, or email, and ensures that task status is synchronized in real time across different systems.

[0096] Real-time monitoring dashboard unit: The system provides a visual real-time monitoring dashboard that dynamically displays the overall status of rectification tasks from three dimensions: geographical distribution, time trend, and responsible party. The dashboard has built-in intelligent early warning rules. For example, when the proportion of overdue tasks exceeds a certain percentage, or when high-risk tasks are delayed beyond the specified time limit, the system will automatically issue an alarm. In addition, the system is designed with escalation logic, which automatically escalates the responsibility for overdue tasks to the next higher level of management.

[0097] Rectification Verification Unit: After the responsible personnel upload rectification vouchers, the system automatically verifies the evidence chain, such as verifying the validity of electronic signatures or comparing the recovered amount with the original abnormal amount. After verification, the system automatically updates the credit score of the relevant responsible personnel based on factors such as the timeliness of task processing and the thoroughness of rectification. Simultaneously, the final rectification results are fed back to the system for subsequent optimization of task distribution strategies and risk detection models, forming a closed loop of continuous improvement.

[0098] In some embodiments, a trusted evidence storage module is also included. The trusted evidence storage module is connected to the intelligent entity parsing and comparison module, the multi-dimensional abnormal behavior detection and reasoning module, and the automated closed-loop rectification module, respectively. It is used to generate digital fingerprints of key comparison results, abnormal detection conclusions, and rectification records and store them on the blockchain to construct an immutable electronic evidence chain.

[0099] In some embodiments, the trusted evidence storage module includes:

[0100] The data fingerprint generation system is used to serialize structured data that needs to be stored for evidence, and then combine it with a timestamp accurate to milliseconds and a random number to generate a unique digital fingerprint through a high-strength cryptographic hash algorithm.

[0101] The smart contract notarization engine is connected to the data fingerprint generation system and is used to write the digital fingerprint into the distributed ledger by calling the smart contract deployed on the consortium blockchain and through a multi-party consensus mechanism.

[0102] A cross-chain verification gateway, connected to the smart contract notarization engine, is used to provide a standardized verification interface. It verifies the integrity and authenticity of the data by recalculating the hash value of the data to be verified and comparing it with the hash value stored on the chain.

[0103] The audit trail subsystem is connected to the data fingerprint generation system, the smart contract evidence storage engine, and the cross-chain verification gateway, respectively. It is used to record all evidence storage operations and leave an indelible record on the blockchain, thus constructing a complete traceable evidence chain.

[0104] Blockchain technology is used to perform digital hash calculations on key comparison results, anomaly detection conclusions, and rectification records, and these are stored on the chain to ensure the immutability and auditability of the verification process and results. The trusted evidence storage module employs blockchain-based distributed ledger technology to construct an immutable electronic evidence chain. This module includes the following core technical components:

[0105] Data fingerprint generation system: For structured data requiring notarization, the system first serializes it, converting it into a uniform binary format. Then, it uses a high-strength cryptographic hash algorithm (such as SHA-256), combined with a millisecond-accurate timestamp and a random number, to generate a unique "digital fingerprint" (i.e., hash value) for each piece of data. This fingerprint is the unique identifier of the data; any tiny data alteration will completely change the fingerprint.

[0106] Smart Contract Evidence Storage Engine: The evidence storage engine is built on consortium blockchain technology. The system calls smart contracts deployed on the chain to "store" the data fingerprint generated in the previous step into the distributed ledger of the blockchain. This process uses a multi-party consensus mechanism (such as a practical Byzantine fault-tolerant algorithm) to ensure that all participating nodes reach a consensus on the evidence storage operation, thereby guaranteeing the authenticity and immutability of the evidence storage record.

[0107] Cross-chain verification gateway: To facilitate verification by external systems, the module provides a standardized verification interface. During verification, the system searches for the original notarization record on the blockchain based on the notarization transaction ID provided by the user. By recalculating the hash value of the data to be verified and comparing it with the hash value stored on the chain, the integrity and authenticity of the data can be verified. Simultaneously, the system also verifies the authority of the timestamp attached to the notarization.

[0108] The audit trail subsystem module provides a nationally compliant timestamp service for every evidence storage operation, ensuring the objectivity and non-repudiation of the evidence storage time. All evidence storage operations (including data storage, querying, and verification) leave an indelible record on the blockchain, thus constructing a complete and traceable chain of evidence, from the original anomaly record to the evidence storage transaction, and then to the block header and timestamp certificate, providing a solid legal basis for subsequent audits and liability determination.

[0109] The following section will provide a detailed explanation of the intelligent comparison and closed-loop supervision system and method for consistency of civil affairs funds' three accounts, using a specific scenario of full-process transparent supervision of urban minimum living allowance funds. This embodiment will fully demonstrate the entire process of the system, from data access, intelligent comparison, deep reasoning to closed-loop rectification and blockchain evidence storage.

[0110] Application Scenario: After the minimum living allowance payment cycle ended in May 2024, the Civil Affairs Bureau of District A under a certain city discovered a discrepancy between the number of people reported in the "Statistical Details" and the "Business Payable Account" submitted to higher authorities. To investigate the cause and prevent risks, the system of this invention was activated.

[0111] Step 1: End-to-end automated fusion of multi-source heterogeneous data;

[0112] After the system starts up, the multi-source heterogeneous data fusion module begins to work.

[0113] Data Access: The data access layer of this module automatically extracts the May "Business Payable Accounts" data from Area A during the low-load period of the business system at 2 AM via a configured MySQL connector. This data includes fields such as "employee ID number, name, payable amount, and business approval date," totaling 1200 records. Simultaneously, an Excel parser reads the "Financial Actual Payment Accounts" transaction file (CSV format) obtained by the finance department from the bank. This file contains fields such as "payee account number, actual payment amount, bank transaction number, and payment date," totaling 1198 records. Finally, the module retrieves "Statistical Details" data from the statistical reporting system via an HTTP API interface, including fields such as "summary reporting personnel, total reported amount, and administrative division code."

[0114] Level 3 cleaning and verification: Data enters the data cleaning layer.

[0115] Basic field validation: The system automatically detected a record in the financial ledger with an ID number that was only 15 digits long. It then activated the intelligent ID number completion algorithm, using the first 6 digits (region code) and 8 digits (birth date) along with the checksum rules to complete the number into a standard 18-digit format (510101199001011234). Simultaneously, the system validated the amount field and found a record in the business ledger with a payable amount of 3500 yuan, while the highest urban minimum living allowance standard in Area A is 1200 yuan. The system determined this record to be "amount exceeding the limit" and marked it as requiring verification.

[0116] Business rule validation: The system performs logical validation on the three accounts (business ledger, financial ledger, and accounting ledger). For example, the approval date for a service object in the business ledger is "2024-05-28", while the corresponding disbursement date in the financial ledger is "2024-05-25". The system triggers the timing logic validator, determines that "the disbursement date is earlier than the approval date" violates the business specification, and marks the record as "logical anomaly".

[0117] Cross-system consistency verification: The system performs a macro-level verification of the total amounts in the three accounts. The total amount payable for business operations is 960,000 yuan, the total amount actually paid by the finance department is 952,000 yuan, and the total amount reported by the statistics department is 960,000 yuan. The system's calculation deviation rate is 0.83%, which is within the preset 1% threshold. It is initially determined that there are no significant macro-level differences, but micro-level differences need to be further identified.

[0118] Standardized Mapping: After cleaning, the data standardization process adopts a composite primary key strategy of "ID number + timestamp". The system generates a unique primary key for each record, such as 510101199001011234_202405, and maps all relevant fields (amount, date, type, etc.) in the three accounts data to a unified data structure table, forming a time-comparable fused dataset centered on the service object.

[0119] Step 2: High-precision entity parsing and comparison based on a hybrid model;

[0120] After data fusion, the intelligent entity analysis and comparison module starts the core comparison engine.

[0121] Precise Matching Layer: The system constructs an improved hash index table for all standardized ID numbers. Through precise matching, 1195 records were successfully associated. For these 1195 records, the system initiates a two-way verification mechanism: when comparing amounts, it finds that 15 records have actual financial payments (800 yuan) lower than business payable amounts (1000 yuan), with a difference of 200 yuan; when comparing times, it finds that the aforementioned 3 records have time sequence anomalies. These 18 records are marked as "candidate anomalies" and proceed to the next stage.

[0122] Fuzzy Matching Layer: For records where exact matching fails, the system initiates fuzzy matching. Five additional records in the business ledger (containing ID numbers) and three additional records in the financial ledger (containing only names, not ID numbers) need to be cross-matched. For example, the business ledger contains a record with ID number 510101197505151111 and name "Ayiguli Maimaiti"; the financial ledger contains a record with the name "Ayiguli" but lacks an ID number. The system then calls the multimodal name matching model:

[0123] First, the character edit distance was calculated, and it was found that the edit distance between "Aygul·Maimaiti" and "Aygul" was 5 (corresponding to the deletion of the three characters "·Maimaiti"), and the similarity was 0.7.

[0124] The system further utilizes a deep semantic model (based on a bidirectional long short-term memory network and an attention mechanism) to convert the two names into semantic vectors. Simultaneously, the system loads name features from the Xinjiang region into a regional dialect feature database, identifying "Aygul Maimaiti" as the full Uyghur name, while "Aygul" is its common abbreviation.

[0125] Finally, the model weighted and fused the edit distance similarity (weight 0.4) and semantic vector similarity (weight 0.6) to obtain a comprehensive similarity score of 0.94, which exceeded the dynamic threshold of 0.85, and successfully matched them as the same entity.

[0126] Entity Disambiguation Layer: During the matching process, ambiguity arose. The financial ledger contains two records, both with the name "Zhang San," and ID numbers 510101199001011235 and 510101199001011236 respectively. However, the business ledger only contains one record with the name "Zhang San." The system enters the entity disambiguation stage, constructing a multi-dimensional feature space: comparing the amounts (both 800 yuan), disbursement dates (both May 20th), and subsidy types (both low-income assistance) of the two financial records, and combining this with historical matching records for the ID number. The system uses an extreme gradient boosting classifier and finds that ID number 510101199001011235 has a higher match degree with historical data of the service recipient in the business ledger (such as home address and contact information), thus accurately identifying the correct matching entity.

[0127] Consistency Comparison Engine: Ultimately, all 1198 successfully matched records are entered into the comparison engine. The system not only compares amounts but also performs time-series analysis, such as determining whether fluctuations in the disbursement amount for the same service recipient within the normal range for the current year. The comparison engine uses an incremental algorithm, storing the 1180 successfully matched records without anomalies in a cache index for direct matching in subsequent comparisons, greatly improving efficiency. Finally, the system outputs a detailed discrepancy report, including 23 warning messages related to amount differences and time-related logical anomalies.

[0128] Step 3: Deep abnormal behavior detection and reasoning based on the dual perspectives of "individual + group";

[0129] The multidimensional abnormal behavior detection and reasoning module conducted in-depth analysis on the 23 early warning messages identified through comparison.

[0130] Individual-level time series analysis: The system analyzes each case in the discrepancy report. Taking the service recipient "Li Si" (ID: 510101198512121234) as an example, the system retrieves its historical payment sequence for the past 12 months [800, 800, 800, 800, 800, 800, 800, 800, 800, 800, 1000]. After analysis by the time series prediction model (recurrent neural network with attention mechanism), it predicts that the payment amount in May 2024 should be 800 yuan, but the actual payment is 400 yuan. The system's dynamic anomaly threshold is set at ±20% of the historical average (i.e., 640-960 yuan). The actual value of 400 yuan is far below the lower limit of the threshold, and the system immediately triggers an individual-level anomaly alarm, marked as "sudden drop in amount".

[0131] Group-level clustering detection: In addition to individual analysis, the system performs unsupervised clustering of the operational behaviors of all personnel. Using an improved density clustering algorithm, the system uses the personnel, operation time, and amount variation as feature vectors. The clustering results show that among the 20 service recipients under the responsibility of "Wang Wu," a personnel officer in a certain street office in District A, 18 households exhibited an abnormal pattern of "actual financial payment being 200 yuan lower than the expected business payment." Furthermore, the operation times for these 18 payment records were all concentrated within the 5 minutes between "2024-05-20 14:30:00" and "2024-05-20 14:35:00." Abnormal records handled by other personnel were scattered. The system clustered "Wang Wu" and the 20 households under his responsibility into a high-risk cluster, marking it as "group-based operational anomalies."

[0132] Intelligent Inference Engine: The system inputs individual anomalies (sudden drop in amount) and group anomalies (same handler, concentrated time, reduced amount) into a built-in knowledge graph. This graph contains entities and relationships such as "low-income allowance disbursement process," "batch import operation specifications," and "handler's authority and responsibilities." The graph neural network model automatically focuses on the three key nodes of "batch import," "reduced amount," and "concentrated time" through a graph attention mechanism. Combining the historical credit score (good) of "handler Wang Wu" and the anomaly pattern (amount decreased rather than increased), the model ultimately infers the most likely cause of the anomaly as: "When importing disbursement data in batches, the handler used an old template from the previous month that contained incorrect amounts, causing the disbursement amounts for most service recipients to be incorrectly overwritten, which is an operational error," and gives a confidence level of 92%.

[0133] Risk Level Assessment: A multi-factor weighted scoring model was used for comprehensive assessment, including: amount difference factor (200 yuan per item, 3600 yuan in total, weight 0.3), duration of abnormality (only this month, weight 0.2), number of people involved (20 households, weight 0.3), operator's historical credit (good, weight -0.1), and abnormal pattern (operational error, weight 0.1). The weighted comprehensive score was 68 points, falling into the "medium risk" range (60-80 points). The system automatically generated a chain of evidence, including key information such as "operator Wang Wu," "operation time window of 5 minutes," "list of 20 households involved," and "details of amount difference," and outputted the final conclusion of "medium risk - operation error."

[0134] Step 4: "Automated closed-loop rectification" based on dynamic strategies;

[0135] Once the risk level is determined, the automated closed-loop rectification module immediately initiates a complete PDCA cycle.

[0136] Intelligent Task Order Generation: The system retrieves the "Medium Risk - Operational Error" template from the task order template library. It intelligently fills in the anomaly description ("Batch Import Using Error Template"), evidence chain (list of 18 households and amount difference table), and risk level (Medium). Simultaneously, it matches rectification guidelines from the knowledge base to generate dynamic, tiered operation instructions: "1. Please immediately verify the May minimum living allowance payment template file; 2. Provide supplementary payments (200 yuan / household) to the 18 affected minimum living allowance households; 3. Upload the bank receipt after the supplementary payment is completed. This task must be completed within 48 hours."

[0137] Multi-channel task distribution: The task distribution engine automatically selects the distribution channel based on the responsibilities of the handler "Wang Wu" (the street's low-income assistance specialist). Task details are pushed through internal business system messages, and a reminder SMS containing the task link is sent to "Wang Wu's" mobile phone via the SMS gateway. Simultaneously, according to upgrade rules, the system pushes the task's "CC" information to his direct superior "Section Chief Zhao" and the district civil affairs bureau's real-time monitoring dashboard.

[0138] Real-time monitoring dashboard: On the large screen of the Civil Affairs Bureau in Area A, the map module of the real-time monitoring dashboard shows a yellow warning light at the location of "Area A - a certain subdistrict office". The trend chart shows "1 medium-risk task, pending processing". When the task has been in progress for 36 hours (12 hours left before the 48-hour deadline), "Wang Wu" has still not uploaded the rectification certificate. The smart early warning rule built into the dashboard triggers the "timeout warning", changes the task status to "timeout risk", and automatically executes the supervision upgrade logic, sending supervision information to the deputy director in charge.

[0139] Rectification Verification: Upon receiving the supervision, "Wang Wu" completed the reimbursement of the difference within 46 hours and uploaded the electronic receipts (including electronic signatures) for the batch reimbursement from the bank. The rectification verification subsystem automatically parsed the receipts, extracted the total reimbursement amount (3600 yuan) and the number of people involved (18 people), and automatically compared them with the original abnormal data to confirm consistency. Simultaneously, the system verified the validity of the electronic signatures, and automatically updated the task status to "Completed" upon successful verification. Based on "Wang Wu's" timeliness (close to timeout) and rectification thoroughness (complete rectification), the system automatically updated his credit score (downgraded from 85 to 80 points). The complete data for this rectification (abnormality-task-processing-verification) was fed back to the system for optimizing future task allocation priorities and dynamic threshold adjustments for "Wang Wu."

[0140] Step 5: "Penetrative" trusted evidence storage based on consortium blockchain;

[0141] Throughout the regulatory process, the trusted evidence storage module operates synchronously as the underlying trust foundation.

[0142] Data fingerprint generation: After the "anomaly detection conclusion" is generated in step 3, the system automatically serializes it into JSON format. Subsequently, the data fingerprint generation system combines a timestamp accurate to milliseconds (2024-05-31 23:59:59.123) and a random number, and calls the SHA-256 high-strength hash algorithm to generate a unique 256-bit hash value for this evidence: 7f83b1657ff1fc53b92dc18148a1d65d...

[0143] Smart Contract Notification: The system calls a smart contract deployed on the Hyperledger Fabric consortium blockchain. This contract uses a practical Byzantine fault-tolerant consensus mechanism to write the aforementioned hash value into the blockchain's distributed ledger. This notification operation is witnessed and confirmed by all participating nodes (Audit Bureau, Finance Bureau, Civil Affairs Bureau) to ensure its immutability.

[0144] Cross-chain verification and audit trail: Three months later, the audit department conducted a routine audit. Auditors entered the notarized transaction ID through the cross-chain verification gateway. The system retrieved the original hash value from the chain and recalculated the hash value against the "anomaly detection conclusion" file provided by the auditors. The two hash values ​​matched, proving that the file had not been modified since the notarization date. Simultaneously, the audit trail subsystem showed that from the generation of the original anomaly record and the submission of the notarized transaction to this verification query, all operations left indelible records on the chain, forming a complete, credible, and traceable chain of evidence, providing a solid legal basis for subsequent audit accountability.

[0145] Secondly, this application proposes a method for intelligent comparison and closed-loop supervision of the consistency of the three accounts of civil affairs funds, including the following steps:

[0146] Multi-source heterogeneous data fusion steps: Connect and obtain the original data from the business system, financial system and statistical system, which are of different sources and have different formats. Clean the original data from the three accounts through a three-level verification mechanism including basic field verification, business rule verification and cross-system consistency verification. Then, use a composite primary key strategy of ID number combined with timestamp to map the cleaned data into a unified three-dimensional data structure.

[0147] Intelligent entity parsing and comparison steps: A hybrid matching strategy combining exact matching and fuzzy matching is adopted to parse and match entities in the three-dimensional data structure. The fuzzy matching strategy integrates character edit distance and semantic vector similarity based on deep neural network, and solves the multi-candidate matching problem through entity disambiguation technology. Finally, the matching entities are compared in multiple dimensions using the ID number as the unique identifier.

[0148] Multidimensional abnormal behavior detection and reasoning steps: Based on individual-level time series analysis and group-level clustering detection, the abnormal data in the consistency comparison results are subjected to hierarchical detection. The group-level clustering detection uses an improved density clustering algorithm to identify group abnormal patterns, and combines the built-in knowledge graph in the civil affairs field to infer the cause of the abnormality and output the risk level using a graph neural network model.

[0149] Automated closed-loop rectification steps: Based on the cause of the anomaly and the risk level, rectification task sheets are automatically generated by combining templates and dynamic content. They are intelligently distributed according to the responsibilities, workload, and historical response speed of the personnel in charge. The execution status of the rectification tasks is monitored and timeout alarms and supervision escalation are implemented. The rectification results are automatically verified and the credit score of the personnel in charge is updated, thus forming a closed-loop management system.

[0150] Thirdly, this application proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0151] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0152] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0153] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0154] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0155] In the embodiments provided in this disclosure, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0156] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0157] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0158] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0159] The above are merely preferred embodiments of the present invention. It should be noted that any modifications and improvements made by those skilled in the art without departing from the present technical solution should also be considered to fall within the scope of protection claimed by the present solution.

Claims

1. A system for intelligent comparison and closed-loop supervision of the consistency of civil affairs funds' three accounts, characterized in that, include: The multi-source heterogeneous data fusion module is used to connect to and acquire the original data from the business system, financial system and statistical system, which are of different sources and have different formats. The original data is cleaned through a three-level verification mechanism including basic field verification, business rule verification and cross-system consistency verification. The cleaned data is mapped into a unified three-dimensional data structure using a composite primary key strategy of ID number and timestamp. The intelligent entity parsing and comparison module is connected to the multi-source heterogeneous data fusion module. It is used to parse and match entities in the three-dimensional data structure by adopting a hybrid matching strategy that combines precise matching and fuzzy matching. The fuzzy matching strategy integrates character edit distance and semantic vector similarity based on deep neural network, and solves the multi-candidate matching problem through entity disambiguation technology. Finally, the matching entities are compared in multiple dimensions using the ID number as the unique identifier. The multidimensional abnormal behavior detection and reasoning module is connected to the intelligent entity parsing and comparison module. It is used to perform hierarchical detection on abnormal data in the consistency comparison results based on individual-level time series analysis and group-level clustering detection. The group-level clustering detection uses an improved density clustering algorithm to identify group abnormal patterns and combines it with the built-in knowledge graph in the civil affairs field to infer the cause of the abnormality and output the risk level using a graph neural network model. The automated closed-loop rectification module, connected to the multi-dimensional abnormal behavior detection and reasoning module, is used to automatically generate rectification task orders based on the cause and risk level of the abnormality, using a combination of templates and dynamic content. It also intelligently distributes the rectification tasks according to the responsibilities, workload, and historical response speed of the personnel in charge, monitors the execution status of the rectification tasks and implements timeout alarms and supervision escalation, and automatically verifies the rectification results before updating the credit score of the personnel in charge, thus forming a closed-loop management system.

2. The system according to claim 1, characterized in that: The multi-source heterogeneous data fusion module includes: The data access layer is configured with multiple data interface adapters for accessing data from different file formats and database sources; The data cleaning layer is used to implement the three-level verification mechanism, which includes: Level 1 basic field verification, used to verify the ID number format according to national standards and verify whether the amount field is within a preset reasonable range; Level 2 business rule verification, used to verify whether the disbursement date is after the business application date and whether the subsidy type for the same service recipient is repeated in the same period; Level 3 cross-system consistency verification, used to compare whether the deviation between the total amount of financial disbursement and the total amount of business application is within a preset acceptable range. The data standardization processing layer is used to map heterogeneous data from different systems into a unified three-dimensional data structure using a composite primary key strategy that combines ID card number and timestamp.

3. The system according to claim 2, characterized in that: The intelligent entity parsing and comparison module includes: The precise matching unit is used to build a hash index for matching the standardized ID card numbers, and to establish a two-way verification mechanism for successfully matched records to verify whether the issuance date is later than the application date. A fuzzy matching unit, connected to the precise matching unit, is used to calculate a comprehensive score for records that fail to be precisely matched, using a hybrid model that integrates edit distance and semantic vector similarity from a deep neural network, and then matching them using a regional dialect feature database. The entity disambiguation unit, connected to the fuzzy matching unit, is used to construct a multi-dimensional feature space containing amount difference, time interval, region code, subsidy type and historical matching confidence when there are multiple matching candidates, and to use a classification algorithm to identify the most likely matching entity from the candidates. The consistency comparison engine, connected to the entity disambiguation unit, is used to perform direct comparison of amounts and time series analysis of disbursement time on matched entities, and to use an incremental algorithm to build a cache index of historically successfully matched entities, so as to comprehensively determine the difference level based on the comparison results.

4. The system according to claim 3, characterized in that: The multidimensional abnormal behavior detection and reasoning module includes: The individual-level time series analysis unit is used to construct a recurrent neural network time series prediction model with an attention mechanism based on the historical disbursement amount sequence of the service recipient, predict the expected amount for the current period, and set a dynamic anomaly judgment threshold that is automatically adjusted according to recent data fluctuations. When the deviation between the actual disbursement amount and the model prediction value exceeds the dynamic threshold, an individual anomaly alarm is triggered. The group-level clustering detection unit is used to employ a density-based spatial clustering algorithm to cluster service recipients or operators with similar disbursement behaviors into clusters to identify abnormal group patterns by using the dispersion of amount, concentration of disbursement time, distribution of subsidy type, and standard deviation of operation time as feature vectors. The intelligent reasoning engine, connected to the individual-level time series analysis unit and the group-level clustering detection unit, is used to automatically focus on knowledge nodes related to the anomaly detection task based on a knowledge graph that integrates subsidy policies, operating procedures, and regional characteristic entities, and infers the cause of the anomaly through a graph neural network model. The risk level assessment unit, connected to the intelligent reasoning engine, is used to employ a multi-factor weighted scoring model to assign dynamic weights to the magnitude of the amount difference, the duration of the anomaly, the number of people involved, and historical credit factors, calculate a comprehensive score, and classify the risk into three levels: low, medium, and high.

5. The system according to claim 4, characterized in that: The automated closed-loop rectification module includes: The intelligent task order generation unit is used to retrieve the corresponding template from the standardized task order template library based on the detected anomaly type and risk level, dynamically fill in the risk level, anomaly type, and evidence chain, and match and generate targeted rectification operation guidelines with the level of detail increasing with the risk level from the knowledge base. A multi-channel task distribution unit, connected to the intelligent task order generation unit, is used to intelligently distribute tasks according to the responsibilities of the personnel, current load, and historical response speed, through internal business system messages, SMS, or email, and ensure that the task status is synchronized in real time across different systems. The real-time monitoring dashboard unit is connected to the multi-channel task distribution unit and is used to dynamically display the overall status of rectification tasks from the dimensions of geographical distribution, time trend and responsible entity. When the proportion of unprocessed tasks exceeds the preset ratio or the delay of high-risk tasks exceeds the specified time limit, an alarm is automatically issued and the supervision escalation logic is executed to escalate the responsibility to the next higher level of management personnel. The rectification verification unit, connected to the real-time monitoring dashboard unit, is used to automatically verify the validity of the electronic signature in the rectification voucher uploaded by the person in charge, compare the consistency between the recovered or reissued amount and the original abnormal amount, update the credit score of the person in charge based on the timeliness of task processing and the completion of rectification, and feed the rectification results back to the system to optimize the subsequent task distribution strategy and risk detection model.

6. The system according to claim 5, characterized in that: It also includes a trusted evidence storage module, which is connected to the intelligent entity parsing and comparison module, the multi-dimensional abnormal behavior detection and reasoning module, and the automated closed-loop rectification module, respectively. It is used to generate digital fingerprints of key comparison results, abnormal detection conclusions and rectification records and store them on the blockchain to build an immutable electronic evidence chain.

7. The system according to claim 6, characterized in that: The trusted evidence storage module includes: The data fingerprint generation system is used to serialize structured data that needs to be stored for evidence, and then combine it with a timestamp accurate to milliseconds and a random number to generate a unique digital fingerprint through a high-strength cryptographic hash algorithm. The smart contract notarization engine is connected to the data fingerprint generation system and is used to write the digital fingerprint into the distributed ledger by calling the smart contract deployed on the consortium blockchain and through a multi-party consensus mechanism. A cross-chain verification gateway, connected to the smart contract notarization engine, is used to provide a standardized verification interface. It verifies the integrity and authenticity of the data by recalculating the hash value of the data to be verified and comparing it with the hash value stored on the chain. The audit trail subsystem is connected to the data fingerprint generation system, the smart contract evidence storage engine, and the cross-chain verification gateway, respectively. It is used to record all evidence storage operations and leave an indelible record on the blockchain, thus constructing a complete traceable evidence chain.

8. The system according to claim 7, characterized in that: The data interface adapters configured in the data access layer of the multi-source heterogeneous data fusion module include an Excel parser, a CSV parser, a JSON parser, and connectors that support MySQL and Oracle databases. Each adapter has a built-in automatic data format detection mechanism to intelligently identify the encoding format and delimiter type of the source data.

9. The system according to claim 8, characterized in that: In the fuzzy matching unit of the intelligent entity parsing and comparison module, the hybrid model that integrates edit distance and semantic vector similarity of deep neural network weights the edit distance similarity and semantic vector similarity according to preset weights to form a comprehensive similarity score, and dynamically adjusts the matching threshold according to historical matching results; In the risk level assessment unit of the multidimensional abnormal behavior detection and reasoning module, the multi-factor weighted scoring model assigns dynamically adjustable weights to different risk factors and classifies risks into three levels—low, medium, and high—based on the weighted comprehensive score. It also generates a detailed chain of evidence for high-risk events, including abnormal amounts, involved parties, and time points.

10. A method for intelligent comparison and closed-loop supervision of the consistency of civil affairs funds' three accounts, characterized in that, Includes the following steps: Multi-source heterogeneous data fusion steps: Connect and obtain the original data from the business system, financial system and statistical system, which are of different sources and have different formats. Clean the original data from the three accounts through a three-level verification mechanism including basic field verification, business rule verification and cross-system consistency verification. Then, use a composite primary key strategy of ID number combined with timestamp to map the cleaned data into a unified three-dimensional data structure. Intelligent entity parsing and comparison steps: A hybrid matching strategy combining exact matching and fuzzy matching is adopted to parse and match entities in the three-dimensional data structure. The fuzzy matching strategy integrates character edit distance and semantic vector similarity based on deep neural network, and solves the multi-candidate matching problem through entity disambiguation technology. Finally, the matching entities are compared in multiple dimensions using the ID number as the unique identifier. Multidimensional abnormal behavior detection and reasoning steps: Based on individual-level time series analysis and group-level clustering detection, the abnormal data in the consistency comparison results are subjected to hierarchical detection. The group-level clustering detection uses an improved density clustering algorithm to identify group abnormal patterns, and combines the built-in knowledge graph in the civil affairs field to infer the cause of the abnormality and output the risk level using a graph neural network model. Automated closed-loop rectification steps: Based on the cause of the anomaly and the risk level, rectification task sheets are automatically generated by combining templates and dynamic content. They are intelligently distributed according to the responsibilities, workload, and historical response speed of the personnel in charge. The execution status of the rectification tasks is monitored and timeout alarms and supervision escalation are implemented. The rectification results are automatically verified and the credit score of the personnel in charge is updated, thus forming a closed-loop management system.

Images