Financial account statement intelligent checking system, method, electronic device and medium

Abstract

The application provides a financial account intelligent checking system, method, electronic device and medium, wherein the system comprises: a data acquisition layer for acquiring and preprocessing financial multi-source heterogeneous data, wherein the financial multi-source heterogeneous data specifically comprises at least one of the following: business vouchers, accounting data, and tax declaration data; an intelligent processing layer for performing multi-mode heterogeneous feature extraction on the financial multi-source heterogeneous data, generating feature vectors of each document, performing three-single matching and industry-finance association based on the feature vectors through a graph matching model, and constructing an industry-finance fusion graph; checking the account, intelligently diagnosing the account difference through an abnormality detection model, obtaining a diagnosis result, automatically generating an executable adjustment rectification scheme and risk warning information based on the diagnosis result; and a result output layer for outputting the industry-finance fusion graph, the diagnosis result, the executable adjustment rectification scheme and the risk warning information through a visual interface or an API interface. The technical scheme of the application can effectively solve the problem that the business, financial and tax data are isolated in the prior art, a five-flow integrated closed loop cannot be constructed, and intelligent matching and difference root cause diagnosis are difficult to achieve.

Description

Technical Field

[0001] This document relates to the field of financial information technology, and in particular to a financial accounting intelligent reconciliation system, method, electronic device and medium. Background Technology

[0002] In the field of VAT verification, existing technologies mainly focus on single-dimensional invoice information identification and basic verification. For example, Chinese patent CN110781875A discloses a system based on OCR technology to identify VAT invoice images, extract structured fields (such as invoice code, amount, and date), and perform equivalence comparison with preset rules, realizing automated invoice scanning and formal compliance review. Another mainstream solution is integrated into enterprise ERP systems (such as SAP and Yonyou), using a "purchase order - receipt note - invoice" two-document matching mechanism to verify the consistency of quantity and amount, thereby controlling procurement payment risks. However, the above technologies are limited to isolated verification of local links, failing to connect the data links between business, finance, and taxation, and unable to construct a complete evidence loop integrating the "invoice flow - goods flow - capital flow - accounting flow - tax declaration flow".

[0003] Existing technologies, relying solely on precise field matching and fixed threshold rules, cannot handle complex scenarios commonly found in real-world business operations, such as "one-to-many" and "many-to-one" relationships, non-standardized names, reasonable discrepancies in amounts, and cross-period accounting. This leads to a high false rejection rate when rules are tightened, and a significant increase in missed detection risks when rules are relaxed, creating a fundamental contradiction between "fuzzy matching" and "high-precision verification." More seriously, existing systems can only passively detect surface discrepancies, unable to trace the root causes or automatically diagnose whether they are due to time-related differences, incorrect account usage, failure to transfer input tax, incorrect declaration entries, or hidden tax risks such as "genuine invoices, fake transactions." They also cannot generate executable rectification plans that include accounting entries, declaration correction guidelines, and business process optimization suggestions, and lack mechanisms for tracking and verifying the implementation status of rectification. Consequently, tax risks are "discovered but not addressed," requiring long manual intervention cycles, reliance on experienced tax personnel, and the inability to form a closed-loop management system for continuous optimization. This severely restricts the transformation and upgrading of corporate tax risk control from "post-event correction" to "intelligent pre-control." Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent financial statement reconciliation system, method, electronic device, and medium, aiming to solve the aforementioned problems in the prior art.

[0005] This invention provides an intelligent financial statement reconciliation system, comprising: The data acquisition layer is used to collect and preprocess multi-source heterogeneous financial data, which specifically includes at least one of the following: business vouchers, accounting data, and tax declaration data. The intelligent processing layer is used to extract multi-modal heterogeneous features from the multi-source heterogeneous financial data, generate feature vectors for each document, and perform three-document matching and business-finance association based on the feature vectors through a graph matching model to construct a business-finance fusion graph. It also performs account and statement verification and uses an anomaly detection model to intelligently diagnose account and statement discrepancies, obtain diagnostic results, and automatically generate executable adjustment and rectification plans and risk warning information based on the diagnostic results. The result output layer is used to output the business-finance fusion graph, diagnostic results, executable adjustment and rectification plans, and risk warning information through a visual interface or API interface.

[0006] This invention provides a method for intelligent reconciliation of financial statements, used in the aforementioned intelligent reconciliation system. The method specifically includes: collecting multi-source heterogeneous financial data through a data acquisition layer and preprocessing it, wherein the multi-source heterogeneous financial data specifically includes at least one of the following: business vouchers, accounting data, and tax declaration data; extracting multi-modal heterogeneous features from the multi-source heterogeneous financial data through an intelligent processing layer to generate feature vectors for each document; based on the feature vectors, performing three-document matching and business-finance association through a graph matching model to construct a business-finance fusion graph; performing statement reconciliation and intelligently diagnosing discrepancies through an anomaly detection model to obtain diagnostic results; automatically generating executable adjustment and rectification plans and risk warning information based on the diagnostic results; and outputting the business-finance fusion graph, diagnostic results, executable adjustment and rectification plans, and risk warning information through a result output layer using a visual interface or API interface.

[0007] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-described intelligent reconciliation method for financial statements.

[0008] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described intelligent financial statement reconciliation method.

[0009] By adopting the embodiments of the present invention, the technical problems of isolated business, financial and tax data, inability to build a closed loop of five flows in one and difficult to achieve intelligent matching and root cause diagnosis of discrepancies in the prior art can be effectively solved. The invention achieves the significant effects of comprehensively improving the level of automation of financial verification, accurately identifying the source of risk, driving intelligent rectification and proactive early warning, and realizing collaborative governance of business, finance and taxation. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the intelligent financial statement reconciliation system according to an embodiment of the present invention; Figure 2 This is a flowchart of the intelligent reconciliation method for financial statements according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0012] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0013] As described in the background section, existing VAT verification technologies can only achieve partial matching of isolated documents such as invoices, purchase orders, and delivery receipts, or rely on fixed rules to perform coarse-grained comparisons of accounting records and tax returns. They cannot connect the data links between business vouchers, accounting records, and tax returns, resulting in the fragmentation of the five flows: invoice flow, goods flow, capital flow, accounting flow, and tax return flow. After discrepancies are discovered, it is impossible to trace the root cause, let alone generate an executable closed-loop rectification plan.

[0014] After long-term research, the inventors discovered that the most fundamental reason for the aforementioned predicament lies in the fact that existing technologies consistently treat financial statement reconciliation as a static and isolated "data comparison task," rather than a dynamic and interconnected "chain of evidence reasoning task." Its core limitations are twofold: firstly, the system only processes equivalence matching of structured fields, lacking the ability to model semantic relationships, temporal dependencies, and structural relationships between multi-source heterogeneous data; secondly, discrepancy diagnosis is simplified to a binary judgment of "whether they are consistent," completely lacking the ability to attribute the causes of discrepancies in a multidimensional way, and even more so, lacking a mechanism to transform diagnostic results into collaborative rectification instructions covering accounting, reporting, and business operations. This "point-based verification + manual judgment" model makes the system unable to cope with complex scenarios commonly found in real business operations, such as fuzzy matching, cross-period processing, account mismatch, and reporting discrepancies, ultimately resulting in a systemic failure that "can detect discrepancies, but doesn't understand why they occurred, how to correct them, or verify the closed loop."

[0015] Based on the above findings, the inventors proposed a concept to reconstruct financial statement reconciliation from "static comparison" to "dynamic graph reasoning and closed-loop decision-making." This involves uniformly collecting three types of heterogeneous data: business vouchers, accounting records, and tax returns. These data are first transformed into computable feature vectors. Then, based on a graph matching model, relationships between these data entities are established, constructing a dynamic fusion graph that integrates business and accounting flows. Subsequently, based on this graph, intelligent diagnosis is performed on numerical differences between the statements. Based on the diagnostic results, executable rectification plans, including accounting adjustments, declaration corrections, and business optimizations, are automatically generated, ultimately forming a complete processing chain from data collection, graph reasoning, diagnostic analysis to plan output. This concept breaks through the traditional system's limitation of merely "data verification," and for the first time, treats financial reconciliation as a systems engineering project with structural correlation, causal reasoning, and execution orientation. This transforms the system from a mere "alarm" reporting discrepancies into an "intelligent decision-making center" capable of reasoning causes, generating solutions, and driving execution.

[0016] Guided by this inventive concept, the inventors have proposed an intelligent financial statement reconciliation system, an intelligent financial statement reconciliation method, an electronic device, and a computer-readable storage medium.

[0017] System Implementation Examples According to embodiments of the present invention, an intelligent financial statement reconciliation system is provided. Figure 1 This is a schematic diagram of the intelligent financial statement reconciliation system according to an embodiment of the present invention, such as... Figure 1 As shown, the intelligent financial statement reconciliation system according to an embodiment of the present invention specifically includes: The data acquisition layer 10 is used to collect and preprocess multi-source heterogeneous financial data, wherein the multi-source heterogeneous financial data specifically includes at least one of the following: business vouchers, accounting data, and tax declaration data. In this embodiment, the data acquisition layer first accesses and preprocesses heterogeneous financial data from multiple sources, including at least one of business vouchers, accounting records, and tax return data. This process aims to transform raw financial information from different systems and in varying formats into structured or semi-structured input that can be recognized and used by the subsequent intelligent processing layer, providing foundational data support for subsequent verification, matching, and diagnosis. The data acquisition layer does not analyze or correlate the data content; it only acquires and preprocesses the data, ensuring that business vouchers, accounting records, and tax return data are completely and accurately incorporated into the system's processing flow. This lays the data input foundation for building a closed-loop verification system integrating business, finance, and taxation.

[0018] In this embodiment, the data acquisition layer 10 is specifically used for: recognizing VAT invoice images through OCR, extracting structured fields and unstructured clues, obtaining invoices and payment vouchers from the ERP system through API; obtaining detailed ledgers of VAT-related accounting subjects from the general ledger system through API; and obtaining column-level structured data of VAT and surcharge tax returns and their supplementary information for each period through API or RPA technology.

[0019] In this embodiment, OCR technology is used to recognize VAT invoice images, accurately extracting structured fields such as invoice code, number, amount, and tax amount, as well as unstructured clues such as seals and handwritten remarks. Combined with an API interface, corresponding delivery orders and payment vouchers are automatically retrieved from the ERP system. Simultaneously, detailed ledger data of VAT-related accounting subjects are retrieved from the general ledger system via API. Furthermore, structured data at the column level of VAT and surcharge tax returns and their supplementary information for each period are collected using API or RPA technology, realizing the integration of five different flows: invoice flow, goods flow, capital flow, accounting flow, and tax flow. The end-to-end, high-precision, and automated collection and structured reconstruction of data provides a complete, consistent, and associative input source for subsequent multi-modal heterogeneous feature vector generation and graph matching models. This supports the accurate construction of the three-item matching and business-finance fusion graph, ensuring that the accounting and financial statements have a sufficient data foundation. It enables the anomaly detection model to identify the root causes of discrepancies based on real, closed-loop five-flow data, and automatically generate executable rectification plans and risk warning information. This completely solves the problems of business-finance chain breakage, accounting and financial statement inaccuracies, and diagnostic failures caused by incomplete data collection, and realizes the closed-loop operation and proactive risk control of intelligent financial verification.

[0020] Intelligent processing layer 12 is used to extract multi-modal heterogeneous features from the multi-source heterogeneous financial data, generate feature vectors for each document, and perform three-document matching and business-finance association through a graph matching model based on the feature vectors to construct a business-finance fusion graph; perform account and statement verification, and perform intelligent diagnosis of account and statement differences through an anomaly detection model to obtain diagnostic results; and automatically generate executable adjustment and rectification plans and risk warning information based on the diagnostic results. In this embodiment, firstly, multi-modal heterogeneous feature extraction is performed on multi-source heterogeneous financial data. This involves extracting features from financial data from different sources that include multi-dimensional information such as semantics, numerical values, time series, and structure, forming a high-dimensional feature vector corresponding to each document. Subsequently, based on these feature vectors, a graph matching model is used to identify the relationships between the three business documents—invoices, delivery orders, and payment vouchers—enabling previously scattered business documents to establish connections based on their feature similarity. On this basis, the business-finance association is further completed, establishing a correspondence between the matched business documents and accounting vouchers, thereby constructing a business-finance fusion graph that reflects the mutual mapping between business and financial data, achieving a systematic integration from raw data to a structured relationship network.

[0021] Subsequently, accounting data and tax return data are aligned and aggregated to make them comparable in terms of account scope and period, thus forming the basis for analyzing discrepancies between accounts and statements. Then, based on this discrepancy data, an anomaly detection model is used to perform intelligent diagnosis, automatically identifying the type and cause of the discrepancy. This model makes inferences solely based on the accounting data itself and its numerical relationships, without relying on external rule bases or manual intervention. Finally, based on the clear results of the diagnosis, an executable rectification plan covering accounting adjustments, tax return corrections, and business optimization is directly generated, and risk warning information associated with the discrepancy is output simultaneously. The entire process is completed automatically by the system without the need for manual intervention in diagnosis or plan design.

[0022] The intelligent processing layer 12 specifically includes: a multi-modal heterogeneous feature fusion and vectorization module, used to generate a high-dimensional dense feature vector that integrates semantic features, numerical features, temporal features, and structural features for each document collected by the data acquisition layer; and a business and financial invoice intelligent matching engine, used to determine candidate matching pairs based on the feature vectors and a graph matching model, according to the rigid features of the documents, construct an initial document relationship graph, and construct a dynamic graph attention network with all documents as nodes and candidate relationships as edges. Based on the multi-modal feature vectors carried by the nodes, the graph attention network aggregates neighbor information and performs inference, dynamically calculates the association probability between nodes, performs complex one-to-many and many-to-one matching deductions, and performs constrained secondary graph matching between matched invoices and accounting vouchers to form association edges between business documents and accounting vouchers. Based on a pre-established context-adaptive dynamic rule base, conflicts in the matching process are resolved, ultimately generating a business-finance integration graph. The dynamic rule base includes rigid rules and flexible rules. Rigid rules are used for hard blocking; flexible rules influence the matching score with weights, and the tolerance weights of flexible rules are automatically adjusted based on historical transaction patterns. An intelligent analysis and attribution engine for account discrepancies aligns the data scope and aggregates the periods of accounting entries and declaration forms, generating comparable accounting and declaration summaries for account reconciliation. An anomaly detection model using rule-based reasoning and case-based reasoning performs intelligent diagnosis of account discrepancies through penetrating root cause analysis, yielding diagnostic results. These results specifically include technical discrepancies, operational discrepancies, accounting processing discrepancies, and declaration discrepancies. An intelligent rectification suggestion generation engine automatically generates executable adjustment and rectification plans covering business, accounting, and tax levels, along with risk warning information, based on the diagnostic results. Executable adjustment and rectification plans specifically include at least one of the following: accounting adjustment suggestions, tax declaration rectification suggestions, and business-side optimization measures.

[0023] In this embodiment, a multi-modal heterogeneous feature fusion and vectorization module generates high-dimensional dense vectors that fuse semantic, numerical, temporal, and structural features for each business voucher, accounting data, and tax declaration data. This enables the intelligent matching engine for business invoices and accounting vouchers to construct a dynamic graph attention network based on a graph matching model, containing all documents as nodes and candidate relationships as edges. It dynamically calculates association probabilities using an aggregation and inference mechanism of multi-modal features between nodes, effectively achieving accurate deduction of complex matching relationships such as one-to-many and many-to-one relationships. Furthermore, by performing constrained secondary graph matching between matched invoices and accounting vouchers, closed-loop association edges are constructed between business documents and accounting vouchers. Combined with a context-adaptive rule base containing rigid rules (hard blocking) and flexible rules (dynamic weight adjustment), the tolerance threshold is automatically optimized based on historical transaction patterns, intelligently... It can resolve matching conflicts and ultimately generate a structured, highly reliable business and financial integration diagram. Simultaneously, the intelligent analysis and attribution engine for discrepancies between accounts and statements achieves accurate comparison between total accounting figures and total tax declaration figures through data caliber alignment and period aggregation. It also employs an anomaly detection model combining rule-based and case-based reasoning to conduct penetrating analysis of discrepancies, accurately identifying the root causes of four categories of discrepancies: technical, operational, accounting processing, and tax declaration. Furthermore, based on the above diagnostic results, the intelligent rectification suggestion generation engine automatically generates a three-dimensional executable plan covering business optimization, accounting adjustments, and tax declaration rectification, along with accompanying risk warning information. This systematically solves the core problems of traditional methods, such as low matching accuracy, rigid rules, inability to attribute discrepancies, and lack of rectification plans, achieving intelligent closed-loop management of the entire business, financial, and invoice data chain and proactive prevention and control of compliance risks.

[0024] In this embodiment, the intelligent processing layer 12 further includes: a self-evolving anomaly detection and closed-loop management engine, used to identify risk patterns such as capital idleness graphs, long-term deviations between accounts and statements, and manipulation of accounting subjects based on the business-finance fusion graph, account-statement discrepancies, and historical data, utilizing a dual mechanism of graph anomaly detection and time-series pattern mining; providing a manual annotation interface, when an alert is confirmed or falsely reported, obtaining annotation signals through the manual annotation interface, using the annotation signals as reward signals for reinforcement learning or fine-tuning datasets, and automatically feeding them back to the graph matching model and anomaly detection model for online learning and continuous performance evolution; continuously tracking adopted executable adjustment and rectification plans, and automatically verifying whether the corresponding account-statement discrepancies have been eliminated in the next verification cycle.

[0025] In this embodiment, the intelligent processing layer, through a self-evolving anomaly detection and closed-loop management engine, integrates business-finance fusion graphs, accounting discrepancies, and historical data. Employing a dual mechanism of graph anomaly detection and time-series pattern mining, it accurately identifies hidden complex risks such as idle capital flow graphs, long-term accounting discrepancies, and manipulation of accounting subjects, overcoming the limitations of traditional static rules in identifying dynamic risks. Simultaneously, it collects user confirmations or false alarms regarding risk warnings through a manual annotation interface, transforming these annotation signals into reinforcement learning reward signals or fine-tuning datasets, which are automatically fed back to the graph matching model and anomaly detection model. This drives online parameter iteration and continuous performance evolution, achieving self-improvement in risk identification accuracy. Furthermore, the system continuously tracks the execution status of adopted rectification plans and automatically compares relevant accounting data in the next verification cycle to verify whether discrepancies have been eliminated, forming a complete closed loop of "risk identification—manual intervention—model optimization—rectification verification." This effectively solves the problems of a lack of feedback mechanisms in risk diagnosis and the system's inability to self-evolve, achieving a qualitative leap in financial verification from passive response to proactive intelligent evolution, significantly improving the accuracy, robustness, and long-term autonomy of business-finance fusion.

[0026] The intelligent processing layer 14, the result output layer, is used to output the business-finance integration diagram, the diagnostic results, the executable adjustment and rectification plan, and the risk warning information through a visual interface or API interface.

[0027] In this embodiment, the output layer presents the business-finance integration diagram, diagnostic results, executable adjustment and rectification plans, and risk warning information through a visual interface or API, achieving structured output and interactive delivery of information. The business-finance integration diagram, as a graph structure representation of the entire chain of relationships, is provided externally through an interface or API, allowing users to intuitively obtain the network of relationships between business, financial, and tax documents. The diagnostic results, based on the difference classification and attribution conclusions automatically analyzed by the system, are directly transmitted through the output layer, ensuring that risk positioning information is not interrupted or distorted by intermediate links. The executable adjustment and rectification plans include accounting adjustment entries, tax declaration correction guidelines, and business optimization measures. This plan, as a complete disposal instruction generated by the system, is delivered completely through the output layer without any deletions or summaries. The risk warning information, based on the potential risk patterns output by the anomaly detection engine, is uniformly pushed through the output layer, ensuring that the warning content and analysis conclusions are presented synchronously. All output content is generated directly by the system and provided externally through a visual interface or API without relying on manual secondary processing or interpretation, ensuring the integrity and real-time nature of information transmission.

[0028] Applying the technical solution of this application, the data acquisition layer uniformly accesses and preprocesses multi-source heterogeneous financial data such as business vouchers, accounting data, and tax declaration data, providing a structured foundation for subsequent intelligent analysis. The intelligent processing layer extracts multi-modal heterogeneous features from various documents, generating feature vectors that accurately represent the semantics and structure of the documents. Based on a graph matching model, it achieves automatic matching and business-finance association between business documents, accounting vouchers, and tax declarations, constructing a business-finance integration graph that integrates business flow, capital flow, invoice flow, information flow, and tax flow, fundamentally breaking down data silos and realizing... The system integrates five key flows (business, finance, and taxation) into one. Simultaneously, during the reconciliation process, the intelligent processing layer uses an anomaly detection model to perform in-depth intelligent diagnosis of discrepancies between accounts and statements. It automatically identifies the root causes of these discrepancies and links them to specific business processes or data anomalies. Based on the diagnostic results, it automatically generates actionable adjustment and rectification plans and precise risk warning information, forming a closed-loop governance mechanism of "detection—diagnosis—rectification—early warning." Finally, the results output layer integrates the business-finance integration diagram, diagnostic results, rectification plans, and risk warning information through a visual interface or API, providing full-chain support for financial control and decision-making. Therefore, it effectively solves the technical problems of isolated business, financial, and tax data, the inability to construct a closed-loop system integrating five flows, and the difficulty in achieving intelligent matching and root cause diagnosis of discrepancies in existing technologies. This achieves significant results in comprehensively improving the automation level of financial reconciliation, accurately identifying risk sources, driving intelligent rectification and proactive early warning, and realizing collaborative governance of business, finance, and taxation.

[0029] The technical solutions described above in the embodiments of the present invention will be explained in detail below.

[0030] (a) Data acquisition layer, used for the access and preprocessing of multi-source heterogeneous data, including the following acquisition objects: Business vouchers: Use OCR to recognize VAT invoice images and extract structured fields and unstructured clues (such as contract number in the remarks column); use API to obtain inbound / outbound orders and payment voucher data from the ERP system.

[0031] Accounting data: Detailed ledgers of VAT-related accounting subjects are obtained from the general ledger system via API, such as the voucher-level transaction amounts and summaries of subjects like "Taxes Payable - VAT Payable (Input Tax / Input Tax Transferred Out / Output Tax)".

[0032] Tax return data: Obtain the column-level structured data of the "Value-Added Tax and Surcharges Declaration Form" and its supplementary information for each period through API or RPA technology.

[0033] (ii) Intelligent processing layer (core): This layer is the core technology of the present invention, consisting of multiple intelligent engines working together to form a complete closed-loop processing chain.

[0034] Module 1: Multimodal Heterogeneous Feature Fusion and Vectorization Module: Generates a high-dimensional dense feature vector for each document, which integrates semantic features (text semantics based on BERT), numerical features (amount, tax rate and derivative ratio), temporal features (time intervals between each stage), and structural features, laying the data foundation for intelligent matching.

[0035] Module Two: Intelligent Matching Engine for Business and Finance Invoices: Performs "three-invoice matching" and "business-finance association" to construct a business-finance integration diagram. Specifically, this includes: Coarse screening: Utilizing rigid features such as tax ID and date, candidate matching pairs are quickly recalled to construct an initial document relationship graph. Fine ranking: All documents are treated as nodes, and candidate relationships are treated as edges, constructing a dynamic heterogeneous graph network. Nodes carry multi-modal feature vectors, and neighbor information is aggregated and inferred through a graph attention network. The association probability between nodes is dynamically calculated, naturally supporting complex matching inferences such as "one-to-many" and "many-to-one".

[0036] Business-Finance Relationship: Matched invoices are matched with accounting vouchers using a constrained secondary graph matching process, forming a "business document → accounting voucher" relationship. Dynamic Rule Base Conflict Resolution: A dynamic rule base is constructed, consisting of "rigid rules + flexible rules + context-adaptive mechanisms." Rigid rules (such as inconsistent tax identification numbers) act as hard barriers; flexible rules (such as a 0.5% fluctuation in amount) influence the matching score with weights; the system can automatically adjust the tolerance weights of flexible rules based on historical transaction patterns.

[0037] Module 3: Intelligent Analysis and Attribution Engine for Account-Report Discrepancies: This engine performs intelligent diagnosis of account-report discrepancies. Specifically, it includes: "Account-Report" Data Alignment and Aggregation: Align the data scope and period of the accounting entries with the columns of the declaration form to form comparable "account summary" and "declaration summary".

[0038] Multidimensional differential root cause automatic diagnosis and classification: Employing a hybrid diagnostic model combining "rule-based reasoning" and "case-based reasoning," it performs penetrating root cause analysis of discrepancies. The diagnostic classification system covers: Technical differences: time differences, caliber differences, and rounding errors.

[0039] Operational differences: missing / incorrect documents, duplicate processing, and error association.

[0040] Accounting discrepancies: incorrect use of accounts, incorrect amounts, omissions in accounting entries, and failure to promptly transfer input tax.

[0041] Reporting discrepancies: omissions / over-reporting, errors in column entry, and inconsistencies in the interrelationships between appendices.

[0042] Module Four: Intelligent Rectification Suggestion Generation Engine: After the root cause is accurately diagnosed, this engine automatically generates actionable adjustment and rectification plans covering three levels: business operations, accounting, and taxation. These include: Accounting Adjustment Suggestions: Generating standard corrected accounting entries with automatic calculation of debit and credit amounts. Tax Filing Rectification Suggestions: Clearly indicating the supplementary schedules to be corrected, specific columns, and adjustment amounts, along with operational instructions. Business Optimization Measures: Proposing front-end optimization suggestions such as freezing payments, improving approval processes, and strengthening inventory management. This solution uses natural language generation technology to translate into clear operational instructions explaining "why, how, and where."

[0043] Module 5: Self-evolving anomaly detection and closed-loop management engine 1. Anomaly Detection: By inputting the matching relationship graph, account-statement discrepancies, and historical data, and using a dual mechanism of "graph anomaly detection" and "time series pattern mining" (such as LSTM), new and hidden risk patterns such as "empty capital rotation graph", "long-term account-statement deviation pattern", and "accounting subject manipulation" are identified.

[0044] 2. Self-evolutionary learning: Provides a manual annotation interface. When an alert is confirmed or a false alarm is detected, the annotation signal is used as a reward signal for reinforcement learning or as a fine-tuning dataset, automatically fed back to the graph matching model and the anomaly detection model, enabling online learning and continuous performance evolution.

[0045] 3. Rectification closed-loop tracking: The system continuously tracks the adopted rectification suggestions and automatically verifies whether the difference has been eliminated in the next verification cycle, ensuring that the risk is resolved in a closed loop.

[0046] (III) Results Output Layer: The matching results, difference diagnosis reports, rectification suggestions, and risk warning information will be output through a visual interface or API interface. A full-link traceability view is provided, supporting drilling down from macro differences to the original invoice images, and providing an interactive entry point for "one-click adoption" of rectification solutions.

[0047] In summary, this invention achieves unified access to cross-system data by collecting and preprocessing multi-source heterogeneous financial information such as business vouchers, accounting data, and tax declaration data. It constructs a business-finance fusion graph based on multi-modal heterogeneous feature vectors and uses a graph matching model to associate invoices, invoices, and payment vouchers with accounting vouchers, overcoming the limitations of traditional static field comparison in handling fuzzy matching and complex business relationships. Furthermore, it uses an anomaly detection model to perform multi-dimensional root cause analysis of discrepancies between accounts and statements, automatically distinguishing between technical, operational, accounting, and declaration-related differences. Based on this, it generates an executable rectification plan including accounting entries, declaration correction guidelines, and business process optimization suggestions, while also outputting risk warning information. Finally, it provides a structured output of diagnostic results and rectification suggestions through a visual interface or API, transforming financial reconciliation from isolated single-point verification into a closed-loop processing mechanism covering data collection, intelligent association, accurate diagnosis, and collaborative rectification, thus improving the completeness, accuracy, and automation level of account and statement reconciliation.

[0048] Method Implementation Examples According to embodiments of the present invention, a method for intelligent reconciliation of financial statements is provided for use in the aforementioned intelligent reconciliation system for financial statements. Figure 2 This is a flowchart of the intelligent reconciliation method for financial statements according to an embodiment of the present invention, as follows: Figure 2 As shown, the intelligent reconciliation method for financial statements according to an embodiment of the present invention specifically includes: Step S201 involves collecting and preprocessing multi-source heterogeneous financial data through a data acquisition layer. This multi-source heterogeneous financial data specifically includes at least one of the following: business vouchers, accounting data, and tax declaration data. Specifically, this includes: recognizing VAT invoice images using OCR to extract structured fields and unstructured clues; obtaining invoices and payment vouchers from the ERP system via API; obtaining detailed ledgers of VAT-related accounting subjects from the general ledger system via API; and obtaining column-level structured data of VAT and surcharge tax returns and their supplementary information for each period via API or RPA technology.

[0049] Step S202 involves extracting multi-modal heterogeneous features from the multi-source heterogeneous financial data using an intelligent processing layer to generate feature vectors for each document. Based on these feature vectors, a graph matching model is used to perform three-document matching and business-finance association to construct a business-finance fusion graph. Account reconciliation is then performed, and an anomaly detection model is used to intelligently diagnose discrepancies between accounts and statements, obtaining diagnostic results. Based on these diagnostic results, an executable adjustment and rectification plan and risk warning information are automatically generated. Specifically, this includes: Through the multi-modal heterogeneous feature fusion and vectorization module, a high-dimensional dense feature vector that integrates semantic features, numerical features, temporal features and structural features is generated for each document collected by the data acquisition layer. Through the intelligent matching engine for business and financial invoices, based on the feature vectors and using a graph matching model, candidate matching pairs are determined according to the rigid features of the documents, and an initial document relationship graph is constructed. All documents are treated as nodes, and candidate relationships are treated as edges, constructing a dynamic graph attention network. Based on the multi-modal feature vectors carried by the nodes, the graph attention network aggregates neighbor information and performs inference, dynamically calculates the association probability between nodes, and performs complex one-to-many and many-to-one matching deductions. The matched invoices are then matched with accounting vouchers through constrained secondary graph matching to form association edges between business documents and accounting vouchers. Based on a pre-established context-adaptive dynamic rule base, conflicts in the matching process are resolved, and finally, a business-finance fusion graph is generated. The dynamic rule base includes: rigid rules and flexible rules. The rigid rules are used for hard blocking; the flexible rules are used to influence the matching score in the form of weights, and the tolerance weights of the flexible rules are automatically adjusted according to historical transaction patterns. Through intelligent analysis and attribution engine for discrepancies between accounts and statements, the data scope of accounting entries and columns in the declaration form are aligned and aggregated over time to generate comparable summary figures for accounts and declarations, and the accounts and statements are reconciled. An anomaly detection model combining rule-based reasoning and case-based reasoning is used to perform intelligent diagnosis of discrepancies through penetrating root cause analysis, and the diagnostic results are obtained. The diagnostic results specifically include: technical discrepancies, operational discrepancies, accounting processing discrepancies, and declaration discrepancies. Based on the diagnostic results, the intelligent rectification suggestion generation engine automatically generates executable adjustment and rectification plans and risk warning information covering three levels: business, accounting, and taxation. The executable adjustment and rectification plans specifically include at least one of the following: accounting adjustment suggestions, tax declaration rectification suggestions, and business optimization measures.

[0050] Furthermore, through a self-evolving anomaly detection and closed-loop management engine, based on the aforementioned business-finance fusion graph, accounting discrepancies, and historical data, and utilizing a dual mechanism of graph anomaly detection and time-series pattern mining, risk patterns such as empty fund transfer graphs, long-term accounting-account deviation patterns, and accounting subject manipulation are identified. A manual annotation interface is provided; when a warning is confirmed or a false alarm is detected, annotation signals are obtained through this interface. These annotation signals are used as reward signals for reinforcement learning or as fine-tuning datasets, and are automatically fed back to the graph matching model and anomaly detection model for online learning and continuous performance evolution. Adopted executable adjustment and rectification plans are continuously tracked, and the corresponding accounting discrepancies are automatically verified in the next verification cycle to ensure they have been eliminated.

[0051] Step S203: Output the business-finance integration diagram, the diagnostic results, the executable adjustment and rectification plan, and the risk warning information through the result output layer using a visual interface or API interface.

[0052] The technical solutions described above in the embodiments of the present invention will be illustrated below.

[0053] S1: Collect and preprocess five types of heterogeneous data: invoices, delivery notes, payment vouchers, accounting records, and tax returns.

[0054] S2: Extract multi-modal heterogeneous features from the data to generate feature vectors for each document.

[0055] S3: Based on feature vectors and graph attention networks, perform "business three-document matching" and "business-finance voucher association" to construct a business-finance fusion graph.

[0056] S4: Perform “ledger-table” data alignment and use a hybrid diagnostic model to perform multidimensional root cause automatic diagnosis and classification of discrepancies.

[0057] S5: Based on the diagnostic results, automatically generate an executable rectification plan that includes accounting adjustment entries, reporting correction guidelines, and business optimization measures.

[0058] S6: Input the matching relationship graph and the difference data of the account table into the self-evolutionary anomaly detection engine to identify anomalies in graph structure and time series dimension and generate risk warnings.

[0059] S7: Receives human feedback signals, updates the parameters of the matching and detection model online, and completes self-evolution.

[0060] S8: Track the implementation status of the rectification plan, automatically verify whether the differences are closed in the next cycle, and output a visual report that includes full-link traceability, difference analysis and risk warning.

[0061] In this embodiment, a self-evolving anomaly detection and closed-loop management engine is used. Based on the business-finance fusion graph, account-statement discrepancies, and historical data, it comprehensively utilizes graph anomaly detection to identify structural anomaly patterns such as the cash flow loop diagram. Combined with time-series pattern mining, it discovers hidden risks such as long-term account-statement deviations and accounting subject manipulation. This overcomes the limitations of traditional methods that rely solely on immediate discrepancies, achieving accurate identification of deep-seated and persistent financial risks. Simultaneously, through a manual annotation interface, it receives user confirmation or false alarm feedback on warning results, converting the annotation signal into a reward signal for reinforcement learning or high-quality training data for model fine-tuning. According to reports, the judgment capabilities of the dynamic optimization graph matching model and the anomaly detection model enable the system to have a closed-loop capability of online learning and self-evolution. In addition, the system continuously tracks the implementation status of adopted rectification plans and automatically compares relevant accounting data in the next verification cycle to verify whether the discrepancies have been eliminated, forming a complete closed loop of "diagnosis-rectification-verification-feedback". This effectively solves the problems of no tracking of rectification and no system evolution, and ultimately achieves a leapfrog improvement in financial risk identification from passive response to proactive prediction and from single processing to continuous optimization, significantly enhancing the accuracy, robustness and long-term applicability of intelligent financial verification.

[0062] The embodiments of the present invention are method embodiments corresponding to the system embodiments described above. The specific operations of each step can be understood by referring to the description of the system embodiments, and will not be repeated here.

[0063] Device Example 1 This invention provides an electronic device, such as... Figure 3 As shown, it includes: a memory 30, a processor 32, and a computer program stored in the memory 30 and executable on the processor 32, wherein the computer program, when executed by the processor 32, performs the steps as described in the method embodiment.

[0064] Device Example 2 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor 32, performs the steps described in the method embodiment.

[0065] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A financial statement intelligent reconciliation system, characterized in that, include: The data acquisition layer is used to collect and preprocess multi-source heterogeneous financial data, wherein the multi-source heterogeneous financial data specifically includes at least one of the following: business vouchers, accounting data, and tax declaration data; The intelligent processing layer is used to extract multi-modal heterogeneous features from the multi-source heterogeneous financial data, generate feature vectors for each document, and perform three-document matching and business-finance association based on the feature vectors through a graph matching model to construct a business-finance fusion graph; it also performs account and statement verification, and intelligently diagnoses the differences between account and statement through an anomaly detection model to obtain diagnostic results, and automatically generates executable adjustment and rectification plans and risk warning information based on the diagnostic results. The results output layer is used to output the business-finance integration diagram, the diagnostic results, the executable adjustment and rectification plan, and the risk warning information through a visual interface or API interface.

2. The system according to claim 1, characterized in that, The data acquisition layer is specifically used for: The system uses OCR to recognize VAT invoice images, extracts structured fields and unstructured clues, and obtains invoices and payment vouchers from the ERP system via API. It also obtains detailed ledgers of VAT-related accounting subjects from the general ledger system via API. Furthermore, it uses API or RPA technology to obtain column-level structured data of VAT and surcharge tax returns and their supplementary information for each period.

3. The system according to claim 1, characterized in that, The intelligent processing layer specifically includes: The multi-modal heterogeneous feature fusion and vectorization module is used to generate a high-dimensional dense feature vector that integrates semantic features, numerical features, temporal features and structural features for each document collected by the data acquisition layer. The business and financial invoice intelligent matching engine is used to determine candidate matching pairs based on the feature vectors and the rigid features of the documents through a graph matching model. It constructs an initial document relationship graph, using all documents as nodes and candidate relationships as edges, and builds a dynamic graph attention network. Based on the multi-modal feature vectors carried by the nodes, it aggregates neighbor information and performs inference through the graph attention network, dynamically calculates the association probability between nodes, and performs complex one-to-many and many-to-one matching deductions. It then performs constrained secondary graph matching between matched invoices and accounting vouchers, forming association edges between business documents and accounting vouchers. Based on a pre-established context-adaptive dynamic rule base, it resolves conflicts in the matching process, ultimately generating a business-finance fusion graph. The dynamic rule base includes rigid rules and flexible rules. Rigid rules are used for hard blocking; flexible rules influence the matching score in the form of weights, and automatically adjust the tolerance weights of the flexible rules according to historical transaction patterns. The intelligent analysis and attribution engine for discrepancies between accounts and reports aligns the data scope and aggregates the periods of accounting entries with the columns of the tax return, generating comparable summary figures for accounts and reports, and performing account and report verification. It employs an anomaly detection model combining rule-based reasoning and case-based reasoning to perform intelligent diagnosis of discrepancies through penetrating root cause analysis, obtaining diagnostic results. These diagnostic results specifically include: technical discrepancies, operational discrepancies, accounting processing discrepancies, and reporting discrepancies. The intelligent rectification suggestion generation engine is used to automatically generate executable adjustment and rectification plans and risk warning information covering three levels: business, accounting, and taxation, based on the diagnostic results. The executable adjustment and rectification plans specifically include at least one of the following: accounting adjustment suggestions, tax declaration rectification suggestions, and business optimization measures.

4. The system according to claim 1, characterized in that, The intelligent processing layer further includes: The self-evolving anomaly detection and closed-loop management engine is used to identify risk patterns such as cash flow manipulation, long-term discrepancies between accounts and statements, and manipulation of accounting subjects based on the aforementioned business-finance fusion graph, accounting discrepancies, and historical data, utilizing a dual mechanism of graph anomaly detection and time-series pattern mining. It provides a manual annotation interface; when a warning is confirmed or a false alarm is detected, annotation signals are obtained through this interface. These signals are then used as reward signals for reinforcement learning or as fine-tuning datasets, automatically fed back to the graph matching model and anomaly detection model for online learning and continuous performance evolution. It continuously tracks adopted and executable adjustment and rectification plans, automatically verifying in the next verification cycle whether the corresponding accounting discrepancies have been eliminated.

5. A method for intelligent reconciliation of financial statements, characterized in that, The method for the intelligent reconciliation system of financial statements according to any one of claims 1 to 4 specifically includes: The data acquisition layer collects and preprocesses multi-source heterogeneous financial data, which specifically includes at least one of the following: business vouchers, accounting data, and tax declaration data. The intelligent processing layer extracts multi-modal heterogeneous features from the multi-source heterogeneous financial data to generate feature vectors for each document. Based on the feature vectors, a graph matching model is used to match the three documents and associate them with business and finance to construct a business-finance fusion graph. The accounting and financial statements are checked, and an anomaly detection model is used to intelligently diagnose the discrepancies between the accounting and financial statements to obtain diagnostic results. Based on the diagnostic results, an executable adjustment and rectification plan and risk warning information are automatically generated. The results output layer outputs the business-finance integration diagram, the diagnostic results, the executable adjustment and rectification plan, and the risk warning information through a visual interface or API interface.

6. The method according to claim 5, characterized in that, The collection of multi-source heterogeneous financial data through the data acquisition layer specifically includes: The system uses OCR to recognize VAT invoice images, extracts structured fields and unstructured clues, and obtains invoices and payment vouchers from the ERP system via API. It also obtains detailed ledgers of VAT-related accounting subjects from the general ledger system via API. Furthermore, it uses API or RPA technology to obtain column-level structured data of VAT and surcharge tax returns and their supplementary information for each period.

7. The method according to claim 5, characterized in that, The intelligent processing layer extracts multi-modal heterogeneous features from the multi-source heterogeneous financial data, generating feature vectors for each document. Based on these feature vectors, a graph matching model is used to match the three documents and associate them with business and finance data, constructing a business-finance fusion graph. Account reconciliation is performed, and an anomaly detection model is used to intelligently diagnose discrepancies, obtaining diagnostic results. Based on these results, an executable adjustment and rectification plan and risk warning information are automatically generated, specifically including: Through the multi-modal heterogeneous feature fusion and vectorization module, a high-dimensional dense feature vector that integrates semantic features, numerical features, temporal features and structural features is generated for each document collected by the data acquisition layer. Through the intelligent matching engine for business and financial invoices, based on the feature vectors and using a graph matching model, candidate matching pairs are determined according to the rigid features of the documents, and an initial document relationship graph is constructed. All documents are treated as nodes, and candidate relationships are treated as edges, constructing a dynamic graph attention network. Based on the multi-modal feature vectors carried by the nodes, the graph attention network aggregates neighbor information and performs inference, dynamically calculates the association probability between nodes, and performs complex one-to-many and many-to-one matching deductions. The matched invoices are then matched with accounting vouchers through constrained secondary graph matching to form association edges between business documents and accounting vouchers. Based on a pre-established context-adaptive dynamic rule base, conflicts in the matching process are resolved, and finally, a business-finance fusion graph is generated. The dynamic rule base includes: rigid rules and flexible rules. The rigid rules are used for hard blocking; the flexible rules are used to influence the matching score in the form of weights, and the tolerance weights of the flexible rules are automatically adjusted according to historical transaction patterns. Through intelligent analysis and attribution engine for discrepancies between accounts and statements, the data scope of accounting entries and the columns of the declaration form are aligned and aggregated over time to generate comparable summary figures for accounts and declarations, and the accounts and statements are reconciled. An anomaly detection model combining rule-based reasoning and case-based reasoning is used to perform intelligent diagnosis of discrepancies through penetrating root cause analysis, and the diagnostic results are obtained. The diagnostic results specifically include: technical discrepancies, operational discrepancies, accounting processing discrepancies, and declaration discrepancies. Based on the diagnostic results, the intelligent rectification suggestion generation engine automatically generates executable adjustment and rectification plans and risk warning information covering three levels: business, accounting, and taxation. The executable adjustment and rectification plans specifically include at least one of the following: accounting adjustment suggestions, tax declaration rectification suggestions, and business optimization measures.

8. The method according to claim 5, characterized in that, The method further includes: Through a self-evolving anomaly detection and closed-loop management engine, based on the aforementioned business-finance fusion graph, accounting discrepancies, and historical data, and utilizing a dual mechanism of graph anomaly detection and time-series pattern mining, it identifies risk patterns such as empty fund transfer graphs, long-term accounting-account deviation patterns, and accounting subject manipulation. It provides a manual annotation interface; when a warning is confirmed or a false alarm is detected, annotation signals are obtained through this interface. These signals are then used as reward signals for reinforcement learning or as fine-tuning datasets, automatically fed back to the graph matching model and anomaly detection model for online learning and continuous performance evolution. It continuously tracks adopted and executable adjustment and rectification plans, automatically verifying in the next verification cycle whether the corresponding accounting discrepancies have been eliminated.

9. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the intelligent reconciliation method for financial statements as described in any one of claims 5 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the intelligent financial statement reconciliation method as described in any one of claims 5 to 8.

Images