Intelligent auditing system for financial statements based on stable learning

Abstract

The application discloses a financial statement intelligent auditing system based on stable learning, which comprises a three-layer cooperative architecture: a criterion-data double alignment preprocessing layer, which is used for domain generalization and incremental learning mechanism in stable learning; a cross-criterion associated feature stable extraction layer, which is connected with the criterion-data double alignment preprocessing layer and is used for knowledge reservation and feature selection mechanism based on stable learning to extract implicit association rules and features which are resistant to data distribution interference, cross-subject and cross-criterion from a data base; and a double dynamic adaptation exception verification layer, which is connected with the cross-criterion associated feature stable extraction layer and is used for a multi-dimensional verification model which fuses stable learning regularization constraints to verify the consistency of cross-subject financial data and the compliance to specific accounting criterion clauses synchronously. Through a criterion updating interface and an incremental learning mechanism, the application can automatically analyze new accounting criteria and convert them into executable auditing rules, realizing seamless and rapid updating of the auditing model.

Description

Technical Field

[0001] This invention belongs to the field of intelligent financial auditing technology, and in particular relates to an intelligent financial statement auditing system based on stable learning. Background Technology

[0002] Currently, intelligent auditing of financial statements primarily relies on two technological approaches. One is traditional audit software based on predefined rule engines. This type of software automates compliance checks on financial data by solidifying accounting standards and audit procedures into a set of executable verification rules (such as account reconciliation relationships and ratio fluctuation thresholds) and scripts. Such systems offer high processing speed and standardization, and have become a fundamental tool for large accounting firms. The other approach incorporates data mining and machine learning technologies. By analyzing historical financial data and audit adjustment records, models are built to identify abnormal transactions or potential misstatement patterns. For example, clustering algorithms can be used to discover outlier expense records, or classification models can be used to predict high-risk accounts. Furthermore, with the development of corporate groups, financial consolidation systems supporting multiple ledgers and multiple accounting standards conversions are widely used in the preparation of consolidated financial statements, providing a structured data foundation for auditing. These technologies collectively constitute the main practices of digital and automated transformation in the current financial auditing field.

[0003] While existing technologies have achieved some success, they still have significant limitations in addressing the core complexities of financial auditing. First, regarding adaptability and maintenance costs, existing systems struggle to adapt to frequent iterations of accounting standards. Rule engines require manual interpretation of new standards and rewriting of numerous scripts, a lengthy and error-prone process; traditional machine learning models face the "catastrophic forgetting" problem, where learning new standards impairs the performance of existing models. Second, in handling data heterogeneity and in-depth analysis, existing tools lack the ability to integrate heterogeneous financial data across legal entities, regions, and accounting periods within a group, and lack intelligent data alignment and noise filtering mechanisms, resulting in a weak analytical foundation. More importantly, their audit logic is often fragmented and superficial; rule checks and data analysis are often conducted independently, failing to simultaneously penetrate and verify the compliance of business operations with standards and the consistency of cross-entity related logic, thus making it difficult to accurately pinpoint the root causes of complex misstatements such as consolidation offsetting and related-party transaction pricing. Finally, in terms of intelligence, existing methods rely on expert-preset rules or historical labels, failing to proactively uncover hidden risk characteristics arising from standard iterations or new business models. These shortcomings collectively lead to the high cost and limited efficiency of the existing audit technology system. Furthermore, when dealing with complex group audits and rapid changes in standards, its audit depth, breadth, and timeliness face severe challenges. There is an urgent need for a new generation of audit system that can achieve self-evolution, intelligent integration, and deep insight. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a financial statement intelligent auditing system based on stable learning, which solves the problems in the prior art where financial auditing systems are difficult to adapt to standard iterations and cannot deeply process heterogeneous data and risks across entities.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A financial statement intelligent auditing system based on stability learning, comprising a three-tier collaborative architecture: The standard-data double alignment preprocessing layer is used to align the inconsistent financial data standards of different accounting entities due to policy differences and accounting standard iterations based on the domain generalization and incremental learning mechanism in stable learning, filter data noise and human adjustment items, and form a unified and standard-adapted cross-entity data base. A stable extraction layer for cross-criterion association features and a preprocessing layer for criterion-data double alignment are used for knowledge preservation and feature selection mechanisms based on stable learning to extract implicit association rules and features that are resistant to data distribution interference and cross-subject and cross-criterion from the data base. The dual dynamic adaptation anomaly verification layer connects to the cross-standard correlation feature stable extraction layer. It is used to adopt a multi-dimensional verification model with integrated stable learning regularization constraints to simultaneously verify the consistency of cross-entity financial data and compliance with specific accounting standard provisions, locate the data root cause of misstatement or fraud issues or the root cause of standard implementation deviations, and output structured risk warnings. The system achieves incremental learning and adaptive model updates to the new accounting standards through the stable learning mechanism.

[0006] Preferably, the standard-data double alignment preprocessing layer quantifies the differences between the old and new accounting standards into audit constraints, including differences in accounting scope, differences in recognition timing standards, and measurement method adjustment parameters, and embeds them into a dynamic rule base for incremental updates.

[0007] Preferably, the differences in data caliber processed by the standard-data double alignment preprocessing layer specifically include deviations in the implementation of accounting policies among different subsidiaries within the group, differences in parallel accounting of the same financial item during the transition period between the old and new standards, and differences in format and granularity caused by inconsistent data entry specifications.

[0008] Preferably, the implicit association rules extracted by the cross-standard association feature stable extraction layer specifically include the mapping logic between accounting standard clauses and audit risk assessment procedures, the deviation characteristics between related party transaction pricing and market fair value, and the automatic elimination entries rules for internal transactions and intercompany transactions in consolidated financial statements.

[0009] Preferably, the multi-dimensional verification model of the dual dynamic adaptation anomaly verification layer is executed according to the hierarchical logic of first verifying compliance with single-entity criteria and then verifying cross-entity data consistency, and the audit conclusion is generated by combining the verification results of the two layers.

[0010] Preferably, the stable learning mechanism specifically includes: a domain generalization algorithm for aligning data distributions of different subjects, an incremental learning algorithm for incorporating new criterion knowledge, and a regularization algorithm for preventing the forgetting of effective patterns under old criteria when learning new knowledge.

[0011] Preferably, it also includes a standard update interface, which is used to receive external accounting standard change information and automatically parse the key points of the standard change through natural language processing technology, and convert them into quantitative audit constraints that can be executed by the preprocessing layer.

[0012] Preferably, after generating new audit constraints, the criterion update interface synchronously triggers the rule base update of the preprocessing layer, the incremental learning and training of the association rules of the feature extraction layer, and the regularization fine-tuning of the model parameters of the anomaly verification layer, thereby realizing the automatic transmission and collaborative adaptation of criterion updates across all layers of the system.

[0013] Preferably, the structured risk alerts output by the dual dynamic adaptation anomaly verification layer include at least the risk level, the subject involved, the specific rules and regulations violated, the related transaction path, and the root cause location explanation.

[0014] Preferably, it also includes a report generation module, which is used to automatically generate a structured audit working paper or risk warning report containing risk items, standard basis, scope of impact and audit recommendations based on the structured risk warning and verification conclusion.

[0015] The technical effects and advantages of the intelligent financial statement auditing system based on stable learning in this invention are as follows: 1. This invention, through a standard update interface and incremental learning mechanism, can automatically parse new accounting standards and transform them into executable audit rules, achieving seamless and rapid updates to the audit model. This fundamentally avoids the high costs and time delays associated with traditional methods that require manual study of standards and rewriting of numerous verification scripts, ensuring that audit capabilities are always in sync with regulatory requirements.

[0016] 2. This invention utilizes techniques such as domain generalization in stable learning to intelligently align the distribution deviations of financial data across different subsidiaries and periods caused by differences in accounting policies, account settings, and data entry standards. This automates the tedious data cleaning and standardization work, constructing a high-quality, unified audit data base and laying a reliable foundation for effective multi-dimensional analysis at the group level.

[0017] 3. This invention not only uncovers deeper cross-entity and cross-standard characteristics beyond preset fixed rules (such as consolidation elimination logic and pricing fairness patterns), but also prevents the "forgetting" of existing effective audit experience when learning new standards through a knowledge retention mechanism. This enhances the system's ability to detect complex and novel financial misstatements and fraud risks, and ensures the continuity and stability of the evolution of audit logic.

[0018] 4. This invention employs a dual dynamic adaptation verification mechanism, enabling simultaneous penetrating checks on the compliance with standards and cross-entity data consistency for the same business matter. This changes the fragmented nature of various checks in the traditional audit process, allowing for precise identification of the underlying root causes of complex misstatements (such as inflated profits through related-party transactions) (whether it is data error or deviation in standard implementation), greatly enhancing the value and depth of audit findings.

[0019] 5. By integrating the aforementioned capabilities, this invention frees auditors from a large amount of repetitive and mechanical verification and inspection work, allowing them to focus on higher-value professional judgment and risk assessment. This not only significantly shortens the audit cycle and reduces operating costs, but also promotes the transformation of the entire audit work model from ex-post sampling inspection to continuous, comprehensive, and intelligent risk monitoring and data analysis decision support. Attached Figure Description

[0020] Figure 1 This is a system block diagram of the intelligent financial statement auditing system based on stable learning proposed in this invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0023] refer to Figure 1 This invention provides an intelligent financial statement auditing system based on stable learning, characterized by a three-layer collaborative architecture: a standard-data dual-alignment preprocessing layer, used to align cross-entity data heterogeneity and adapt to accounting standard iterations based on domain generalization and incremental learning mechanisms, forming a unified data base; a cross-standard correlation feature stable extraction layer, used to extract interference-resistant cross-entity correlation features from the base based on knowledge retention and feature selection mechanisms; and a dual dynamic adaptation anomaly verification layer, used to verify the verification model using fusion regularization technology, simultaneously verifying data consistency and standard compliance, and locating the root cause of misstatements to output risk warnings. This system achieves incremental learning and adaptive updates to new accounting standards through a stable learning core algorithm, enabling standard differences to be automatically propagated to each layer. It maintains stable cross-entity audit performance without model reconstruction, effectively solving the problems of traditional audit systems' inability to adapt to rapid standard changes and their inability to deeply process complex group-related data.

[0024] Example 1

[0025] This embodiment provides a financial statement intelligent auditing system based on stable learning for iterative adaptation auditing of new revenue accounting standards: Purpose of implementation: This embodiment aims to demonstrate how the system can automatically adapt to major iterations of accounting standards (such as switching from the old revenue standard to the new revenue standard), solve the problem of data incomparability caused by different subsidiaries' different implementation progress during the transition period, and automatically identify the risk of misstatement caused by improper or inconsistent implementation of the standards.

[0026] Implementation System: The system's criteria-data double alignment preprocessing layer, cross-criterion correlation feature stable extraction layer, and dual dynamic adaptation anomaly verification layer are enabled. Specifically, the system's domain generalization module, incremental learning module, and criterion update interface functions are invoked.

[0027] Implementation steps: Step 1: Standard Update and Interpretation: The standard update interface receives the official text and interpretation of US Accounting Standard ASC 606 (or International Standard IFRS 15). The system automatically parses out key changes such as "identifying performance obligations in a contract," "allocating the transaction price to each performance obligation," and "recognizing revenue when the performance obligation is fulfilled."

[0028] Step 2: Rule Transformation and Data Alignment: The system quantifies the parsed change points into specific audit constraint rules, such as: "For software sales and service contracts containing multiple performance obligations, revenue must be allocated proportionally to the individual selling price of each obligation." An incremental learning mechanism seamlessly adds this new rule to the existing rule base. Simultaneously, for subsidiary B, which has already implemented the new standard, and subsidiary C, which still implements the old standard, the system uses a domain generalization algorithm to analyze their revenue data from the past three years, reconstructing B's new standard data into a comparable caliber under the old standard, forming a unified cross-entity data base.

[0029] Step 3: Feature Extraction: At the feature extraction layer, the system retains the strong correlation between "contract signing time" and "revenue recognition" under the old standard, and incrementally learns the new correlation between "performance obligation completion progress" and "revenue recognition" under the new standard. In addition, it also extracts "logical consistency of revenue recognition time for the same cross-subsidiary contract within the group" as a key cross-entity monitoring feature.

[0030] Step 4: Anomaly Verification and Output: During the audit verification phase, the system detected that subsidiary C recognized full revenue for a cross-year service contract on the contract signing date, while the system's performance progress report only showed 30%. The dual dynamic adaptation anomaly verification layer simultaneously triggers two types of alerts: First, a compliance alert, indicating that this operation violates ASC 606's requirement for revenue recognition based on performance progress; second, a cross-entity consistency alert, showing that subsidiary B, responsible for other parts of the same contract, has recognized revenue based on performance progress. The system ultimately outputs a structured risk warning, clearly indicating the root cause of the risk, the relevant accounts, the specific regulatory clauses violated, and the related-party transaction path.

[0031] Implementation results: The system successfully achieved seamless adaptation to the new accounting standards, eliminating the need for manual rewriting of audit rules. During the transition period of mixed standards implementation, the system automatically resolved data discrepancies and accurately identified significant revenue misstatement risks caused by delays in standard implementation, improving audit efficiency by approximately 60% and avoiding risks caused by omissions in manual comparisons.

[0032] Example 2

[0033] This embodiment provides a financial statement intelligent auditing system based on stable learning for auditing the fairness of transfer pricing in cross-border related-party transactions: Purpose of implementation: This embodiment demonstrates how the system addresses accounting differences and data heterogeneity among subsidiaries in different countries, and automatically audits the fairness of related-party transaction pricing based on multi-source data to identify potential profit shifting and tax risks.

[0034] Implementation System: The system primarily employs a criterion-data dual-alignment preprocessing layer for data standardization, utilizes a cross-criterion correlation feature stable extraction layer to mine pricing fairness features, and employs a dual dynamic adaptation anomaly verification layer to simultaneously verify compliance and rationality.

[0035] 3. Implementation Steps

[0036] Step 1: Multi-Source Data Access and Standardization: The system accesses financial data from the Chinese parent company D and the overseas subsidiary E, as well as an independent industry market database. The preprocessing layer first uses a domain generalization algorithm to map the differentiated accounting item codes of both parties, unifying the financial data format and currency unit. Simultaneously, it retrieves batches of contemporaneous third-party transaction prices for similar raw materials from the market database as a fair price benchmark.

[0037] Step 2: Fairness Feature Mining: The feature extraction layer does not simply compare prices, but deeply integrates business rules. It systematically learns the logic of clauses such as the "comparable uncontrolled price method" in China's Corporate Income Tax Law and the OECD Transfer Pricing Guidelines, and extracts the differences between internal and external transactions in terms of credit periods, transportation terms, etc., to construct a comprehensive pricing fairness judgment model. This model resists normal market fluctuation noise through stable learning.

[0038] Step 3: Multidimensional Risk Verification: The verification layer model calculation revealed that the price at which Company D sold raw materials to Company E was consistently 20% lower than the median price of comparable third-party prices, with extremely high statistical significance. Based on this, the system generated risk alerts: firstly, tax compliance risk, indicating a potential violation of the arm's length principle; and secondly, commercial rationality risk, indicating a significant deviation between the internal pricing model and the market model. The report detailed the deviation percentage, statistical confidence level, specific transaction details, and related parties involved.

[0039] Step 4: Retrospective Support: The risk documentation generated by the system can be directly linked to relevant vouchers, contract terms, and external market price sources, providing auditors with a complete chain of evidence to begin performing further transfer pricing due diligence.

[0040] Implementation results: The system overcomes the limitations of manual auditing, such as the difficulty in obtaining comparable external data and the limited analytical dimensions, enabling automated and continuous multi-dimensional monitoring of related-party transaction pricing. It can quickly and quantitatively identify pricing anomalies and increases the coverage of tax risk audits from sampling to near-full scanning, significantly enhancing the group's tax compliance assurance.

[0041] Example 3

[0042] This embodiment provides a stable learning-based intelligent auditing system for financial statements, used for automated auditing of internal eliminations in consolidated financial statements of large groups: Purpose of implementation: This example demonstrates how a system can automatically process the reconciliation and offsetting of massive amounts of internal transaction data for a large group with hundreds of member units, and accurately identify the risk of misstatement in consolidated financial statements caused by incorrect account mapping and failed transaction matching.

[0043] Implementation System: The system fully utilizes a three-tier architecture. The preprocessing layer handles data cleaning and unifies related party identifiers; the feature extraction layer focuses on learning complex logic rules for merging and offsetting; and the anomaly verification layer performs full offsetting verification and false alarm localization.

[0044] Implementation steps: Step 1: Data Base Construction: The system imports all individual reports, internal transaction details, and organizational structure trees. The preprocessing layer uses algorithms to automatically identify and standardize related party names, clean up duplicate and invalid transaction records, and form clean and standardized consolidation and offsetting working paper data.

[0045] Step 2: Learning and Consolidating Offset Logic: The feature extraction layer does not use fixed formulas. Instead, it uses machine learning to autonomously extract complex logic from historically successful merger processes, such as matching and offsetting rules for "parent company accounts receivable - subsidiary" and "subsidiary accounts payable - parent company," as well as calculation and offsetting rules for "internal sales unrealized profits." Through regularization techniques, the system can firmly retain the core offsetting logic it has learned when learning new merger and acquisition subsidiary transaction patterns, preventing "forgetting."

[0046] Step 3: Full Elimination Verification and Misstatement Location: The anomaly verification layer performs fully automated elimination tests on all intra-group transactions. For example, it discovers that a lease deposit for a newly acquired subsidiary G has inconsistent accounting entries (one party records it as "Other Receivables - Lease Deposit," while the other records it as "Other Payables - Equipment Transactions"), causing the system's automatic elimination to fail. The system immediately locates the problem, indicating that it may cause an overstatement of assets and liabilities in the consolidated balance sheet, and checks whether the accounting treatment of the transaction complies with lease standards.

[0047] Step 4: Generate audit evidence: The system automatically generates a consolidation elimination audit procedure table, listing all transactions that have been successfully eliminated, partially eliminated but failed to be eliminated, and completely eliminated but failed to be eliminated, along with an analysis of the reasons and adjustment suggestions.

[0048] Implementation results: This completely frees auditors from the massive and tedious work of reconciling internal transactions, enabling 100% full auditing of the consolidation elimination process. The elimination accuracy is close to 100%, and the root cause of errors can be precisely located. The consolidation elimination audit work, which previously took weeks to complete, is now reduced to hours, significantly improving the quality of consolidated financial statement preparation and audit efficiency.

[0049] Example 4

[0050] This embodiment provides a financial statement intelligent auditing system based on stable learning for auditing financial instruments using the expected credit loss (ECL) model.

[0051] Purpose of implementation: This embodiment demonstrates the system's ability to handle highly complex audits of the Financial Instruments Accounting Standard (IFRS 9), particularly how it verifies the rationality of the division logic and provision amount of the bank's three-stage "Expected Credit Loss (ECL)" model.

[0052] Implementation System: The system's incremental learning capability is utilized to grasp the differences between the old and new standards (IAS39 vs IFRS9), credit risk early warning features are constructed using the feature extraction layer, and data-based model verification is implemented through the validation layer.

[0053] Implementation steps: Step 1: Criterion Parsing and Model Parameterization: The criterion update interface parses IFRS 9, transforming abstract concepts such as "significant increase in credit risk (SICR)," "stage division," and "adjustment of forward-looking information" into quantifiable and auditable rules and parameter thresholds, which are then embedded into the system rule base.

[0054] Step 2: Risk Feature Extraction Across Criteria: Through incremental learning, the system retains key features such as "number of overdue days" under the old loss model while incrementally learning multiple early warning features under the new ECL model, such as subtle changes in borrowers' repayment behavior patterns, declines in the industry's prosperity index, and macroeconomic forecast variables (such as GDP growth rate). The system can also compare the consistency of different branches in the stage segmentation of similar risk customer groups and extract management deviation features.

[0055] Step 3: Model Output Validation: The system accesses the bank's ECL model output and underlying data. The validation layer performs the following checks: ① Checks whether the stage classification is consistent with the SICR quantitative and qualitative standards; ② Compares the stage classifications of similar loan portfolios across different branches to identify any significant differences without reasonable explanation; ③ Utilizes its learned features to independently assess the credit risk of a subset of sample loans and cross-validates them with the bank's model results. For example, if the system finds that a certain manufacturing loan portfolio is generally classified as Stage 1, but its associated industry forecast data has deteriorated significantly, it will issue an inquiry alert.

[0056] Step 4: Generate specialized working papers: The system outputs ECL audit-specific working papers, which detail the phase-based inspection samples, model parameter rationality analysis, cross-departmental comparison results, and anomalies that require additional explanation from management.

[0057] Implementation results: It provides automated, repeatable, and in-depth tools for auditing highly complex financial models. The system not only checks the correctness of model calculations but also verifies the prudence and rationality of the models from the perspectives of business logic and cross-institutional consistency, significantly improving the professionalism and reliability of audit work in such high-risk areas.

[0058] Example 5

[0059] This embodiment provides a stable learning-based intelligent financial statement auditing system for automated auditing of annual consolidated financial statements of a retail group: Purpose of implementation: This example simulates a complete annual financial statement audit project, demonstrating how the system performs a comprehensive audit task involving multiple accounts and multiple entities from end to end, from data preparation to risk focus, and finally generates structured working papers.

[0060] Implementation System: Enable all system modules and demonstrate the end-to-end process.

[0061] Implementation steps: Step 1: Comprehensive Data Integration and Preprocessing: The system imports trial balances, general ledgers, related party lists, contract ledgers, and historical data of newly acquired companies from all legal entities within the group with a single click. It automates currency conversion, adjustments for differences between old and new accounting standards (such as leases), and reclassification of core accounts, building a group-level audit data platform.

[0062] Step 2: Preliminary Intelligent Risk Screening: The system performs parallel analysis of various key items. For revenue, it analyzes the correlation between sales volume and customer traffic / sales per square meter for each store; for inventory, it calculates and compares the turnover rate of subsidiaries in each region; for fixed assets, it analyzes the rationality of new additions and decommissionings. Based on preset rules and machine learning anomaly detection, the system automatically identifies Subsidiary X, with abnormally slow inventory turnover, and Subsidiary Y, with a significant mismatch between fixed asset growth and revenue growth, as high-risk audit areas.

[0063] Step 3: In-depth verification of high-risk areas: The system initiates in-depth audit procedures for the marked high-risk areas. For subsidiary X, perform an inventory valuation test: verify whether its inventory impairment provision is sufficient (standard compliance), and compare it with the inventory aging structure and impairment provision ratio of subsidiaries in similar regions (cross-entity consistency). For subsidiary Y, perform a capitalization expenditure test: check whether the capitalization of large expenditures complies with the asset definition (standard), and verify whether the capitalized amount is consistent with the asset valuation in the acquisition agreement (data consistency).

[0064] Step 4: Automated Working Paper and Report Generation: The report generation module summarizes the results of all audit steps and automatically generates structured audit working papers. The working papers are organized by account and risk level, and each finding includes: a description of the issue, the amount involved, an index of relevant vouchers, the accounting standard clauses violated, cross-entity comparison data, and system-suggested further audit procedures (such as "perform a full inventory count on subsidiary X").

[0065] Implementation results: This has enabled the process-oriented, automated, and intelligent transformation of annual audit work. Auditors are freed from extensive data collection, processing, and simple verification tasks, allowing them to focus on the highest-risk areas identified by the system for professional judgment and in-depth investigation. The overall audit project cycle is expected to be shortened by 30%-50%, while the depth and breadth of audit coverage are enhanced.

[0066] Comparative Example 1

[0067] This comparison provides traditional, rule-based auditing software, including: Purpose of implementation: By demonstrating the limitations of traditional auditing techniques in addressing the "new revenue standard adaptation" scenario in Implementation Case 1, the technological advancement and practical value of the system of this invention are highlighted.

[0068] Implementation System: A commercial-grade financial auditing software based on predefined rules and script execution.

[0069] Implementation steps and limitations: Step 1: Manual Rule Updates: Faced with new accounting standards, the audit team must invest significant manpower in studying the original text of the standards and manually translating and writing the new requirements into specific scripts (such as SQL queries) or rule packages that can be executed by the software. This process is time-consuming and labor-intensive, and is prone to introducing rule errors due to misunderstandings.

[0070] Step 2: Manual Data Adjustment and Segmentation: Because the software cannot intelligently align data under different standards, the audit team needs to manually export the data of subsidiary B, which has implemented the new standards, use formulas in Excel to reverse-calculate the amounts under the old standards, and then import them into the system. This entire process is highly error-prone and cannot perform effective cross-entity correlation analysis.

[0071] Step 3: Fragmented Audit Procedures: The software can only run different inspection modules sequentially. When running the new revenue standard inspection script, it cannot simultaneously verify the consistency of policies within the group; when running related-party transaction inspections, it cannot incorporate the requirements of the new standard. The inspection dimensions are singular, creating information silos.

[0072] Step 4: Results rely on manual review: The software outputs a large number of hashed anomaly lists, requiring senior auditors to spend a lot of time cross-checking, correlation analysis and root cause judgment, resulting in low automation.

[0073] Comparison results: Compared to traditional methods, this invention's system exhibits disruptive advantages: in terms of adaptability, it achieves automatic learning and adaptation to changes in criteria, rather than manual recoding; in terms of intelligence, it can uncover implicit correlations and complex patterns, rather than simply executing explicit rules; and in terms of integration, it achieves simultaneous verification and correlation localization of multi-dimensional risks, rather than fragmented inspections. Ultimately, it achieves a qualitative leap in audit quality, efficiency, and the ability to respond to complex changes.

[0074] Compared with Examples 1-5 and Comparative Example 1, Examples 1-5 of this invention comprehensively demonstrate the fundamental technological breakthrough achieved by the intelligent audit system based on stable learning compared to the traditional rule engine (Comparative Example 1). Its core advantages are reflected in the three dimensions of the adaptiveness, intelligence, and integration of the technical mechanism. When faced with the iteration of accounting standards, traditional systems rely on auditors to manually read, translate, and rewrite fixed audit rules (such as SQL scripts). This process is not only time-consuming and error-prone, but also cannot handle the differences in cross-entity data caliber caused by the coexistence of old and new standards. Its feature extraction capability is limited to preset explicit rules and cannot uncover implicit patterns under new standards, such as the "correlation between performance progress and revenue recognition". Its verification logic is fragmented and cannot simultaneously link standard compliance and data consistency. In stark contrast, the system of this invention achieves automatic parsing, rule quantification, and seamless embedding of new criteria through a collaborative mechanism of "criterion update interface - stable learning algorithm"; it intelligently aligns heterogeneous data using domain generalization algorithm to construct a unified analysis basis; and with incremental learning and knowledge retention mechanisms, it autonomously mines complex correlation features across criteria and subjects without forgetting effective historical experience; finally, through a dual dynamic adaptation verification model, it performs multi-dimensional and root-cause synchronous location and correlation analysis of risks, realizing a fundamental transformation from a rigid, passive, and fragmented technical paradigm to an autonomous, intelligent, and collaborative paradigm.

[0075] This technological paradigm shift has directly brought about a qualitative leap in audit quality, efficiency, and capabilities. In terms of efficiency and coverage, traditional methods burden auditors with tedious manual rule maintenance, data cleaning, and fragmented alert review. This system liberates auditors from these low-value tasks, achieving 100% full automation of audits in consolidation and offsetting processes, reducing weeks of work to hours, and improving audit efficiency by over 50%. Regarding audit depth and risk detection capabilities, traditional methods are limited by the breadth and dimensionality of rules, making it difficult to detect complex new risks (such as biases in expected credit loss model classification and cross-entity policy inconsistencies). This system can accurately locate deep-seated, interconnected misstatements that traditional methods easily miss, shifting risk auditing from sampling to near-full scanning. Ultimately, this system drives the audit workflow to transform from ex-post, sampling, and manual verification to continuous, comprehensive, and intelligent risk monitoring and insight, providing a reliable technological foundation for addressing increasingly complex business environments and rapidly evolving regulatory requirements.

[0076] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0077] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A financial statement intelligent auditing system based on stable learning, characterized in that, Includes a three-tier collaborative architecture: The standard-data double alignment preprocessing layer is used to align the inconsistencies in financial data caliber caused by policy differences and accounting standard iterations among different accounting entities based on the domain generalization and incremental learning mechanism in stable learning, filter data noise and human adjustment items, and form a unified and standard-adapted cross-entity data base. A stable extraction layer for cross-criterion association features and a preprocessing layer for criterion-data double alignment are used for knowledge preservation and feature selection mechanisms based on stable learning to extract implicit association rules and features that are resistant to data distribution interference and cross-subject and cross-criterion from the data base. The dual dynamic adaptation anomaly verification layer connects to the cross-standard correlation feature stable extraction layer. It is used to adopt a multi-dimensional verification model with integrated stable learning regularization constraints to simultaneously verify the consistency of cross-entity financial data and compliance with specific accounting standard provisions, locate the data root cause of misstatement or fraud issues or the root cause of standard implementation deviations, and output structured risk warnings. The system achieves incremental learning and adaptive model updates to the new accounting standards through the stable learning mechanism.

2. The intelligent financial statement auditing system based on stable learning as described in claim 1, characterized in that, The aforementioned standard-data double alignment preprocessing layer quantifies the differences between the old and new accounting standards into audit constraints, including differences in accounting scope, differences in recognition timing standards, and adjustment parameters of measurement methods, and embeds them into a dynamic rule base for incremental updates.

3. The intelligent financial statement auditing system based on stable learning as described in claim 1, characterized in that, The differences in data caliber processed by the standard-data double alignment preprocessing layer include, in particular, deviations in the implementation of accounting policies among different subsidiaries within the group, differences in parallel accounting of the same financial item during the transition period between the old and new standards, and differences in format and granularity caused by inconsistent data entry specifications.

4. The intelligent financial statement auditing system based on stable learning as described in claim 1, characterized in that, The implicit association rules extracted by the cross-standard association feature stable extraction layer specifically include the mapping logic between accounting standard clauses and audit risk assessment procedures, the deviation characteristics of related party transaction pricing and market fair value, and the automatic elimination entries rules for internal transactions and intercompany transactions in consolidated financial statements.

5. The intelligent financial statement auditing system based on stable learning as described in claim 1, characterized in that, The multi-dimensional verification model of the dual dynamic adaptation anomaly verification layer is executed according to the hierarchical logic of first verifying compliance with single-entity criteria and then verifying cross-entity data consistency, and the audit conclusion is generated by combining the verification results of the two layers.

6. The intelligent financial statement auditing system based on stable learning as described in claim 1, characterized in that, The stable learning mechanism specifically includes: a domain generalization algorithm for aligning data distributions of different subjects, an incremental learning algorithm for incorporating new criterion knowledge, and a regularization algorithm for preventing the forgetting of effective patterns under old criteria when learning new knowledge.

7. The intelligent financial statement auditing system based on stable learning as described in claim 1, characterized in that, It also includes a standard update interface, which is used to receive external accounting standard change information and automatically parse the key points of the standard change through natural language processing technology, and convert them into quantitative audit constraints that can be executed by the preprocessing layer.

8. The intelligent financial statement auditing system based on stable learning as described in claim 7, characterized in that, After generating new audit constraints, the rule update interface synchronously triggers the rule base update of the preprocessing layer, the incremental learning and training of the association rules of the feature extraction layer, and the regularization fine-tuning of the model parameters of the anomaly verification layer, thereby realizing the automatic transmission and collaborative adaptation of rule updates across all layers of the system.

9. The intelligent financial statement auditing system based on stable learning as described in claim 1, characterized in that, The structured risk alerts output by the dual dynamic adaptation anomaly verification layer include at least the risk level, the subject involved, the specific rules and regulations violated, the related transaction path, and the root cause analysis.

10. The intelligent financial statement auditing system based on stable learning as described in claim 9, characterized in that, It also includes a report generation module, which automatically generates structured audit working papers or risk warning reports containing risk items, standard basis, scope of impact and audit recommendations based on the structured risk warnings and verification conclusions.

Images