Tax, accounting and auditing method and system driven by AI artificial intelligence

Abstract

The invention discloses an AI artificial intelligence-driven tax, accounting and auditing method and system, and relates to the technical field of tax and financial management, and the method comprises the following steps: S1, data collection and preprocessing; s2, performing iterative optimization on the xLSTM model through an improved Adame optimizer; s3, extracting a financial data articulation knowledge graph; s4, tracking an abnormal source by using the articulation knowledge graph; s5, generating a comprehensive regulation suggestion; and S6, generating a standardized audit report. According to the method, the problems of long manual calculation time, low tax risk prediction precision, auditing response lag and the like in a traditional financial management method are solved, and an efficient, accurate and real-time financial management technical solution is provided through intelligent automatic analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tax and financial management, and particularly relates to an AI artificial intelligence driven tax, accounting and auditing method and system. BACKGROUND

[0002] With the continuous expansion of enterprise operation scale and the increasing complexity of financial data, traditional tax, accounting and auditing methods have been difficult to meet the needs of modern enterprises for efficient and accurate financial management. Especially in the face of large-scale, high-frequency, heterogeneous financial data, traditional methods often show insufficient processing capacity and cannot timely identify potential tax risks, financial abnormalities and compliance problems. Therefore, how to cope with the complexity and dynamics in financial management has become an important challenge faced by modern enterprises.

[0003] In traditional tax and financial management, financial personnel usually rely on manual operation or manual calculation based on fixed rules to complete tax calculation, account checking and auditing detection. Although these methods have certain effects in simple scenarios, they often expose problems such as reaction lag, tedious process and insufficient accuracy when dealing with a large amount of complex financial data. Especially in tax accounting, traditional methods rely on manual understanding and regular review of tax regulations by financial personnel, lack of automated analysis and processing capacity for real-time financial data, and are difficult to quickly identify tax risks and financial data anomalies, resulting in greater compliance risks and financial risks for enterprises.

[0004] In the aspect of auditing, traditional auditing methods often rely on linear statistical analysis or manual experience judgment, ignoring the deep relevance and potential rules in financial data, resulting in insufficient accuracy and efficiency of auditing. Especially for the complex correlation between enterprise income, expenditure, asset status, liability level and other multi-dimensional data, traditional auditing methods are difficult to fully capture and utilize, resulting in serious deficiencies in financial risk prediction and auditing detection accuracy, and easy to appear misjudgment or auditing loopholes.

[0005] In addition, with the deepening of modern enterprises' dependence on financial data, traditional methods have not been able to fully utilize the advantages of artificial intelligence and deep learning technology, especially in the field of multi-modal data feature mining and intelligent decision-making. Modern financial data often has characteristics of high dimensionality, nonlinearity, heterogeneity, etc., and traditional methods cannot efficiently process these complex data patterns, greatly limiting the accuracy and timeliness of financial decision-making.

[0006] Therefore, how to provide an AI artificial intelligence driven tax, accounting and auditing method and system has become an important problem to be solved in the current enterprise financial management field. SUMMARY

[0007] The application provides an AI artificial intelligence driven tax, accounting and auditing method and system, which can more accurately realize real-time monitoring of enterprise financial data, tax calculation, financial risk assessment and auditing compliance detection. This method can not only automatically process and analyze the financial data, transaction records and tax regulation information of enterprises, but also can take advantage of the xLSTM model in learning the time dependence of financial data to deeply mine the potential patterns and trends in financial data. The application adopts an improved AdamE optimizer, introduces a learning rate preheating and cosine annealing mechanism on the basis of the traditional AdamE optimizer, improves the optimization accuracy, makes the training reach a better global optimal solution, and can more efficiently optimize the xLSTM model. Through this optimization, the model can quickly adapt to the high dimensionality, nonlinearity and diversity of enterprise financial data, update the model parameters in real time, and significantly improve the accuracy and stability of tax calculation, financial risk prediction and auditing compliance detection. The application overcomes the problems of long manual calculation time, low tax risk prediction accuracy and auditing response lag in traditional financial management methods, and provides an efficient, accurate and real-time financial management technical solution through intelligent and automated analysis. The application can significantly improve the response speed of enterprises in tax calculation, financial risk management and auditing supervision, improve decision-making efficiency, reduce financial risk, and promote the intelligent transformation of enterprise financial management.

[0008] An AI artificial intelligence driven tax, accounting and auditing method according to an embodiment of the application comprises the following steps: S1, collecting and importing multi-source heterogeneous financial data, performing data preprocessing on the financial data, and storing the financial data into a historical financial data set; S2, inputting the historical financial data set into an xLSTM model, introducing a learning rate preheating and cosine annealing mechanism, and iteratively optimizing the xLSTM model through an improved AdamE optimizer; S3, defining and configuring an auditing point rule library, and extracting a financial data cross-checking relationship knowledge graph through the xLSTM model; S4, tracking financial data fluctuations in real time, predicting financial data fluctuation results using the xLSTM model, marking abnormal data through prediction errors, tracing abnormal sources using the cross-checking relationship knowledge graph, and generating a financial data abnormal fluctuation report; S5, according to the financial data abnormal fluctuation report and user prompts, matching relevant regulation clauses and historical cases, and generating a comprehensive regulation suggestion; S6, generating a standardized audit report, checking data accuracy, and supporting archiving, sharing and exporting functions after being reviewed and approved by an auditor.

[0009] Optionally, the financial data includes accounting data, Excel spreadsheets, financial records in bank accounts and transaction systems, financial statements, and data from external data sources; the preprocessing includes removing duplicate data, repairing missing data, detecting abnormal data, unifying formats, converting data types, removing noise, merging data from different sources, eliminating redundant information, and unifying time series data.

[0010] Optionally, the improved AdamE optimizer specifically includes: Initialize the model parameters, including the learning rate, decay factor, first moment estimate, second moment estimate, and bias correction term; By backpropagation, the gradient of the loss function with respect to each parameter is calculated; Based on the gradient, calculate the first-order moment estimate and the second-order moment estimate; Perform bias correction based on the bias correction term, and update the first-order moment estimate and the second-order moment estimate; Using the decay factor multiplied by the negative number of training steps as the exponent and the initial learning rate as the base, perform exponential calculation to obtain an updated learning rate that gradually decreases as training progresses. Update the model parameters using gradient, first-order moment estimate, second-order moment estimate, and updated learning rate; The learning rate is dynamically adjusted based on the current iteration number. During the warm-up phase, the learning rate gradually increases linearly from the initial learning rate to the target learning rate for the warm-up phase. After the preheating phase, the proportion of the current iteration number to the total number of iterations in the cosine annealing phase is calculated to obtain the progress ratio. The progress ratio is then mapped to the cosine function to calculate the corresponding cosine value. The cosine value is then linearly transformed to obtain the cosine annealing adjustment factor. Based on the cosine annealing adjustment factor, the difference between the target learning rate and the minimum learning rate in the warm-up stage is calculated, and then multiplied by the cosine annealing adjustment factor to obtain the learning rate adjustment amount as the number of iterations changes. The learning rate value of the current iteration step is obtained by subtracting the learning rate adjustment amount from the target learning rate in the warm-up stage. As the number of iterations increases, the learning rate gradually decreases in a cosine form until the minimum learning rate is reached. Continue iterative optimization until the loss function converges.

[0011] Optionally, S3 specifically includes: S31. Based on financial industry standards and historical business experience, define a rule base for audit key points, including the completeness of vouchers, compliance of accounts, and accuracy of amounts; S32. Transform the audit key point rule base into a fixed-dimensional vector representation, use the embedding layer to transform the audit key point rule base into an audit key point rule feature vector through dimensionality reduction operation, and transform the historical financial dataset into a feature vector of each financial item. S33. Input the feature vectors of each financial item and the feature vectors of the audit key point rules into the LSTM layer, and gradually learn the long-term dependencies in the financial data through time steps. At each time step, the LSTM layer calculates and updates the hidden state vector based on the latest input financial feature vector and the hidden state of the previous time step. The financial items include accounts, amounts, and vouchers. S34. Based on the hidden state vector, use the attention mechanism to calculate the weight of each financial item according to the hidden state vector and the feature vector of each financial item; according to the weight of each financial item, the feature vectors of each financial item are weighted and summed to generate a mapping relationship aggregation vector, which serves as the set of financial item mapping relationships. S35. Based on the set of financial item mapping relationships, with each financial item as a node and the mapping relationships between financial items as edges, construct a knowledge graph of financial data reconciliation relationships.

[0012] Optionally, S4 specifically includes: S41. Real-time monitoring and tracking of financial data fluctuations, inputting historical financial data into the xLSTM model, and using the embedding layer to transform the historical financial dataset into feature vectors of each financial item through dimensionality reduction operations; S42. Input the feature vectors of each financial item and the feature vectors of the audit key point rules into the LSTM layer. Gradually learn the long-term dependencies in the financial data through time steps. At each time step, the LSTM layer calculates and updates the hidden state vector based on the latest input financial feature vector and the hidden state of the previous time step. S43. Input the hidden state vector into the fully connected layer, process it through linear transformation and linear activation function, predict the feature vector of each financial item in the future, and obtain the predicted value of financial data. S44. Compare the predicted financial data with the actual financial data, and calculate the prediction error; S45. Set a prediction error threshold. If the prediction error exceeds the prediction error threshold, mark the actual value of the financial data as an abnormal data item and record the abnormality type. S46. Based on the abnormal data items, extract the reconciliation relationships of the abnormal data items through the knowledge graph of financial data reconciliation relationships to obtain all related data items; S47. Examine all related data items layer by layer, trace the correlation between abnormal data, and gradually narrow down the scope. First, examine the general ledger in the financial statements, and then examine the specific accounts, vouchers and subsidiary ledgers until the specific source of the abnormality is located. S48. Generate a report on abnormal fluctuations in financial data based on abnormal data items, abnormal types, related data items, reconciliation relationships, and abnormal sources.

[0013] Optionally, S5 specifically includes: S51. Collect the latest financial and audit regulations, build and continuously update the financial regulations database; collect historical audit case data, including common financial misstatements, non-compliance clauses and handling suggestions during the audit process, build and continuously update the financial case database; S52. Based on the report of abnormal fluctuations in financial data and the prompts entered by the user, use natural language processing technology to match relevant legal provisions, accounting standards, and tax regulations in the financial regulations database, and provide prompts for violations. S53. Based on reports of abnormal fluctuations in financial data, warnings of violations, and prompts entered by the user, use natural language processing technology to match relevant historical audit cases in the financial case database and recommend historical cases and corresponding solutions. S54. Generate comprehensive regulatory recommendations based on violation clause prompts, historical cases, and corresponding solutions.

[0014] Optionally, S6 specifically includes: S61. Based on financial data, abnormal data items, reconciliation relationships, and comprehensive regulatory recommendations, financial audit data is obtained by combining them. S62. Design a standardized audit report template, map financial audit data to the corresponding fields in the audit report template, and perform formatting and calculation to generate an audit report; S63. Check whether the data entered in the audit report meets the requirements, including the correctness of the amount, the matching of accounts, and the consistency of logic. Check whether there are any errors or omissions according to the audit key point rule base. If any inconsistencies or errors are found, the system will give an error message to remind the auditor to modify the audit report. S64. Auditors review and revise the generated audit report, and finally confirm and approve it through electronic signature; S65. Archive the final confirmed audit report and provide sharing and access control functions to ensure the secure storage and sharing of the report; S66 provides the function of exporting audit reports, supporting multiple export formats.

[0015] According to an embodiment of the present invention, an AI-driven tax, accounting, and auditing system includes the following modules: The data acquisition and preprocessing module is used to collect and import multi-source heterogeneous financial data, preprocess the financial data, and store it into a historical financial dataset. The optimization algorithm module is used to input historical financial datasets into the xLSTM model, introduce learning rate warm-up and cosine annealing mechanisms, and iteratively optimize the xLSTM model through an improved AdamE optimizer. The knowledge graph module for reconciliation relationships is used to define and configure the rule base for audit key points. It extracts the knowledge graph of reconciliation relationships of financial data through the xLSTM model. The financial fluctuation tracking module is based on real-time tracking of financial data fluctuations. It uses an xLSTM model to predict the financial data fluctuation results, marks abnormal data by prediction errors, and uses a knowledge graph of correlation relationships to track the source of anomalies and generate a report on abnormal financial data fluctuations. The comprehensive regulatory advice module is used to match relevant regulatory provisions and historical cases with reports of abnormal fluctuations in financial data and user prompts to generate comprehensive regulatory advice. The audit report module is used to generate standardized audit reports, verify data accuracy, and support archiving, sharing, and export functions after auditors review and approve them.

[0016] The beneficial effects of this invention are: (1) This invention, by introducing the xLSTM model, successfully overcomes the shortcomings of traditional tax, accounting, and auditing methods in handling large-scale, high-dimensional, nonlinear, and dynamically changing financial data. Traditional methods typically rely on manual calculation and rule-driven analysis, which often exhibit low efficiency and poor accuracy when dealing with complex financial data (such as multi-dimensional data on corporate revenue, expenditure, assets, and liabilities). The xLSTM model, through its unique long short-term memory structure, can efficiently capture long-term dependencies in financial data and adapt to dynamic changes in data, significantly improving the accuracy and real-time performance of financial data analysis. This model can handle heterogeneous financial data and deeply explore potential patterns and trends in financial data, improving the accuracy and real-time performance of tax accounting, financial risk prediction, and audit compliance detection. With the help of the xLSTM model, enterprises can quickly identify and handle tax risks, optimize financial decision-making processes, and promote the intelligent transformation of financial management.

[0017] (2) This invention introduces an improved AdamE optimizer, which, based on the traditional AdamE optimizer, combines learning rate warm-up and cosine annealing optimization strategies to significantly improve the efficiency, stability, and accuracy of the model training process. The improved AdamE optimizer avoids oscillations caused by excessive gradients during training by adjusting the dynamic decay strategy of the learning rate, and effectively reduces the risk of gradient vanishing or exploding. By introducing bias correction, gradient pruning, and dynamic decay mechanisms, the optimizer can quickly adapt to the high dimensionality and diversity of complex financial data, ensuring that the model can converge stably when facing high-dimensional and nonlinear financial data, thereby effectively improving the robustness and accuracy of financial data analysis. This optimization method ensures efficient training of the xLSTM model in tax, accounting, and auditing tasks, reduces training time, and provides enterprises with more accurate and real-time financial decision support. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a schematic diagram of the structure of an AI-driven tax, accounting, and auditing system proposed in this invention. Figure 2 This is an overall flowchart of an AI-driven tax, accounting, and auditing method proposed in this invention; Figure 3 This is a flowchart of the AdamE optimizer, which is an AI-driven improvement method for taxation, accounting, and auditing proposed in this invention. Detailed Implementation

[0020] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 2 and Figure 3 An AI-driven method for taxation, accounting, and auditing, comprising the following steps: S1. Collect and import multi-source heterogeneous financial data, perform data preprocessing on the financial data, and store it in the historical financial dataset; S2. Input the historical financial dataset into the xLSTM model, introduce learning rate warm-up and cosine annealing mechanisms, and iteratively optimize the xLSTM model through the improved AdamE optimizer. S3. Define and configure the audit key point rule base, and extract the knowledge graph of financial data correlation through the xLSTM model; S4. Track financial data fluctuations in real time, use the xLSTM model to predict the results of financial data fluctuations, mark abnormal data through prediction errors, use the knowledge graph of correlation relationships to track the source of anomalies, and generate a report on abnormal fluctuations in financial data. S5. Based on reports of abnormal fluctuations in financial data and user prompts, match relevant legal provisions and historical cases to generate comprehensive legal recommendations; S6. Generate standardized audit reports, verify data accuracy, and support archiving, sharing, and export functions after auditors review and approve them.

[0022] In this embodiment, the financial data includes accounting data, Excel spreadsheets, bank accounts and financial records in the transaction system, financial statements, and external data source data; the preprocessing includes removing duplicate data, repairing missing data, detecting abnormal data, unifying formats, converting data types, removing noise, merging data from different sources, eliminating redundant information, and unifying time series data.

[0023] This invention comprehensively cleans and integrates financial data from various sources, ensuring data accuracy, consistency, and efficiency, thereby providing high-quality foundational data for subsequent data analysis and modeling. The financial data includes accounting records, Excel spreadsheets, financial records from bank accounts and transaction systems, financial statements, and data from external data sources. By removing duplicate data, redundant information caused by data duplication is eliminated, reducing computational and storage burdens and ensuring data uniqueness. Repairing missing data effectively fills gaps in the financial data, avoiding analytical errors caused by missing data and improving data integrity. Anomaly detection can promptly identify and remove extreme values ​​or erroneous data, ensuring data rationality and accuracy. Standardized formats and data type conversions ensure that data from different data sources can be standardized, facilitating subsequent analysis and avoiding errors caused by inconsistent formats. Noise removal improves data clarity and reduces interference with analysis results. Merging data from different sources integrates financial information scattered across different systems into a unified dataset, enhancing data comprehensiveness and consistency, and improving analytical accuracy. Eliminating redundant information further optimizes the data structure and reduces the waste of computational resources. Time series standardization can standardize data from different time formats, ensuring the consistency of time series data and providing accurate data support for time series analysis and forecasting. Through these preprocessing steps, this invention can significantly improve the quality and processing efficiency of financial data, providing a solid data foundation for subsequent intelligent analysis, tax calculation, financial risk assessment, and audit compliance testing.

[0024] In this embodiment, the improved AdamE optimizer specifically includes: Initialize the model parameters, including the learning rate, decay factor, first moment estimate, second moment estimate, and bias correction term; By backpropagation, the gradient of the loss function with respect to each parameter is calculated; Based on the gradient, calculate the first-order moment estimate and the second-order moment estimate; Perform bias correction based on the bias correction term, and update the first-order moment estimate and the second-order moment estimate; Using the decay factor multiplied by the negative number of training steps as the exponent and the initial learning rate as the base, perform exponential calculation to obtain an updated learning rate that gradually decreases as training progresses. Update the model parameters using gradient, first-order moment estimate, second-order moment estimate, and updated learning rate; The learning rate is dynamically adjusted based on the current iteration number. During the warm-up phase, the learning rate gradually increases linearly from the initial learning rate to the target learning rate for the warm-up phase. After the preheating phase, the proportion of the current iteration number to the total number of iterations in the cosine annealing phase is calculated to obtain the progress ratio. The progress ratio is then mapped to the cosine function to calculate the corresponding cosine value. The cosine value is then linearly transformed to obtain the cosine annealing adjustment factor. Based on the cosine annealing adjustment factor, the difference between the target learning rate and the minimum learning rate in the warm-up stage is calculated, and then multiplied by the cosine annealing adjustment factor to obtain the learning rate adjustment amount as the number of iterations changes. The learning rate value of the current iteration step is obtained by subtracting the learning rate adjustment amount from the target learning rate in the warm-up stage. As the number of iterations increases, the learning rate gradually decreases in a cosine form until the minimum learning rate is reached. Continue iterative optimization until the loss function converges.

[0025] This invention effectively improves the efficiency, accuracy, and stability of model training through a series of refined steps. Initializing model parameters, learning rate, decay factor, first-moment estimate, second-moment estimate, and bias correction term provides a solid foundation for subsequent optimization, ensuring the model has reasonable initial parameters at the start of training and avoiding training problems caused by improper initialization. Backpropagation calculates the gradient of the loss function with respect to each parameter, effectively determining the model's optimization direction and providing accurate information for parameter updates. Based on the gradient, calculating the first-moment and second-moment estimates smooths the model's gradient history, reducing fluctuations during training and improving stability. Performing bias correction based on the bias correction term updates the first-moment and second-moment estimates, eliminating bias in gradient estimation, ensuring more accurate gradient calculation, and thus accelerating model convergence. Using the decay factor multiplied by the negative of the training steps as the exponent and the initial learning rate as the base, an exponential calculation is performed to obtain an updated learning rate that gradually decreases as training progresses. This allows the model to update quickly in the early stages of training and adjust more finely in later stages, avoiding premature convergence. The model parameters are updated using gradients, first-order moment estimates, second-order moment estimates, and the updated learning rate to ensure efficient updates in each iteration and optimize model performance. The learning rate is dynamically adjusted based on the current iteration count. During the warm-up phase, the learning rate gradually increases linearly from the initial learning rate to the target learning rate, ensuring stable convergence in the early stages of training and preventing slow convergence due to an excessively small learning rate. After the warm-up phase, the proportion of the current iteration count to the total number of iterations in the cosine annealing phase is calculated to obtain the progress ratio. This progress ratio is mapped to a cosine function to calculate the corresponding cosine value. A linear transformation of the cosine value yields the cosine annealing adjustment factor, which helps the model smoothly reduce the learning rate in the later stages of training, ensuring fine-tuning. Based on the cosine annealing adjustment factor, the difference between the target learning rate and the minimum learning rate during the warm-up phase is calculated and multiplied by the cosine annealing adjustment factor to obtain the learning rate adjustment amount as the number of iterations changes. This further refines the change in the learning rate, allowing for more precise model optimization and avoiding unnecessary training fluctuations. The learning rate for the current iteration is obtained by subtracting the learning rate adjustment from the target learning rate during the warm-up phase. As the number of iterations increases, the learning rate gradually decreases in a cosine manner until it reaches the minimum learning rate, ensuring that the model can be finely tuned in the final stage. Continuous iterative optimization continues until the loss function converges. This process guarantees the stability of the optimization and efficient convergence, enabling the model to reach the global optimum or near-optimal solution in a relatively short time.

[0026] In this embodiment, S3 specifically includes: S31. Based on financial industry standards and historical business experience, define a rule base for audit key points, including the completeness of vouchers, compliance of accounts, and accuracy of amounts; S32. Transform the audit key point rule base into a fixed-dimensional vector representation, use the embedding layer to transform the audit key point rule base into an audit key point rule feature vector through dimensionality reduction operation, and transform the historical financial dataset into a feature vector of each financial item. S33. Input the feature vectors of each financial item and the feature vectors of the audit key point rules into the LSTM layer, and gradually learn the long-term dependencies in the financial data through time steps. At each time step, the LSTM layer calculates and updates the hidden state vector based on the latest input financial feature vector and the hidden state of the previous time step. The financial items include accounts, amounts, and vouchers. S34. Based on the hidden state vector, use the attention mechanism to calculate the weight of each financial item according to the hidden state vector and the feature vector of each financial item; according to the weight of each financial item, the feature vectors of each financial item are weighted and summed to generate a mapping relationship aggregation vector, which serves as the set of financial item mapping relationships. S35. Based on the set of financial item mapping relationships, with each financial item as a node and the mapping relationships between financial items as edges, construct a knowledge graph of financial data reconciliation relationships.

[0027] This invention defines an audit point rule base based on financial industry standards and historical business experience. This step ensures the compliance and reliability of financial data analysis, laying a solid foundation for subsequent analysis. The audit point rule base is transformed into a fixed-dimensional vector representation. An embedding layer is used to reduce the dimensionality of the audit point rule base into audit point rule feature vectors, and historical financial datasets are transformed into feature vectors for each financial item. This standardization of rules and financial data into vector form facilitates efficient model processing and ensures data consistency and accuracy in subsequent analysis. The feature vectors of each financial item and the audit point rule feature vectors are input into an LSTM layer. Through time steps, the LSTM layer gradually learns the long-term dependencies in the financial data. At each time step, the LSTM layer calculates and updates the hidden state vector based on the latest input financial feature vector and the hidden state of the previous time step. This allows the model to capture the deep-seated relationships in financial data over time, ensuring effective learning of long-term dependencies. Based on the hidden state vectors, an attention mechanism is used to calculate the weight of each financial item based on the hidden state vector and the feature vector of each financial item. Through this process, the model automatically focuses on important financial items relevant to the task, ensuring efficient processing of key financial data. Based on the weight of each financial item, the feature vectors of each financial item are weighted and summed to generate a mapping relationship aggregation vector, which serves as the financial item mapping relationship set. This step effectively integrates data from different financial items, forming a holistic view of the financial data, facilitating subsequent analysis and decision-making. Based on this financial item mapping relationship set, with each financial item as a node and the mapping relationships between financial items as edges, a knowledge graph of financial data correlation is constructed. This graph intuitively displays the inherent connections between financial items, providing in-depth financial data insights and aiding in real-time auditing and financial risk detection.

[0028] In this embodiment, S4 specifically includes: S41. Real-time monitoring and tracking of financial data fluctuations, inputting historical financial data into the xLSTM model, and using the embedding layer to transform the historical financial dataset into feature vectors of each financial item through dimensionality reduction operations; S42. Input the feature vectors of each financial item and the feature vectors of the audit key point rules into the LSTM layer. Gradually learn the long-term dependencies in the financial data through time steps. At each time step, the LSTM layer calculates and updates the hidden state vector based on the latest input financial feature vector and the hidden state of the previous time step. S43. Input the hidden state vector into the fully connected layer, process it through linear transformation and linear activation function, predict the feature vector of each financial item in the future, and obtain the predicted value of financial data. S44. Compare the predicted financial data with the actual financial data, and calculate the prediction error; S45. Set a prediction error threshold. If the prediction error exceeds the prediction error threshold, mark the actual value of the financial data as an abnormal data item and record the abnormality type. S46. Based on the abnormal data items, extract the reconciliation relationships of the abnormal data items through the knowledge graph of financial data reconciliation relationships to obtain all related data items; S47. Examine all related data items layer by layer, trace the correlation between abnormal data, and gradually narrow down the scope. First, examine the general ledger in the financial statements, and then examine the specific accounts, vouchers and subsidiary ledgers until the specific source of the abnormality is located. S48. Generate a report on abnormal fluctuations in financial data based on abnormal data items, abnormal types, related data items, reconciliation relationships, and abnormal sources.

[0029] This invention inputs the feature vectors of each financial item and the feature vectors of audit key rules into an LSTM layer. Through time steps, it gradually learns the long-term dependencies in the financial data. At each time step, the LSTM layer calculates and updates the hidden state vector based on the latest input financial feature vector and the hidden state from the previous time step. This ensures the model can capture the dynamic patterns of financial data changing over time, improving its ability to model complex financial data. The hidden state vectors are then input into a fully connected layer and processed through linear transformations and linear activation functions to predict the feature vectors of future financial items, obtaining predicted financial data values. This allows for accurate estimation of future financial data through a precise prediction model, improving the accuracy and real-time performance of financial forecasts. Comparing the predicted financial data values ​​with the actual financial data values ​​to calculate the prediction error helps quantify the model's performance and perform error analysis, ensuring continuous optimization of the model's predictive ability. A prediction error threshold is set; if the prediction error exceeds this threshold, the actual financial data value is marked as an anomalous data item, and the anomalous type is recorded. This automatic identification and marking of anomalous data facilitates subsequent auditing and risk control. Based on the abnormal data items, a knowledge graph of financial data correlation relationships is used to extract the correlation relationships between the abnormal data items, obtaining all related data items. This ensures that financial data anomalies can be traced from a global perspective, identifying potential associated risks. All related data items are examined layer by layer, tracing the correlation relationships of abnormal data and gradually narrowing the scope. First, the general ledger in the financial statements is examined, then specific accounts, vouchers, and subsidiary ledgers are examined step by step until the specific source of the anomaly is located. This process can deeply explore the causes of anomalies in financial data, ultimately accurately locating the root cause of the anomaly, avoiding the limitations of relying solely on manual judgment in traditional methods. Based on the abnormal data items, anomaly types, related data items, correlation relationships, and anomaly sources, a report on abnormal fluctuations in financial data is generated. This helps auditors and decision-makers clearly understand the specific circumstances and background of the anomaly, thereby taking more effective countermeasures and improving the transparency and compliance of financial management.

[0030] In this embodiment, S5 specifically includes: S51. Collect the latest financial and audit regulations, build and continuously update the financial regulations database; collect historical audit case data, including common financial misstatements, non-compliance clauses and handling suggestions during the audit process, build and continuously update the financial case database; S52. Based on the report of abnormal fluctuations in financial data and the prompts entered by the user, use natural language processing technology to match relevant legal provisions, accounting standards, and tax regulations in the financial regulations database, and provide prompts for violations. S53. Based on reports of abnormal fluctuations in financial data, warnings of violations, and prompts entered by the user, use natural language processing technology to match relevant historical audit cases in the financial case database and recommend historical cases and corresponding solutions. S54. Generate comprehensive regulatory recommendations based on violation clause prompts, historical cases, and corresponding solutions.

[0031] This invention collects the latest financial and auditing regulations, constructs and continuously updates a financial regulation database, ensuring the system can acquire and store the latest regulatory information in real time, providing up-to-date regulatory support for financial analysis. It also collects historical audit case data, including common financial misstatements, non-compliant clauses, and handling suggestions during the audit process, constructing and continuously updating a financial case database. This database provides auditors with intelligent recommendations based on historical cases, improving decision-making efficiency. Based on reports of abnormal fluctuations in financial data and user-input prompts, natural language processing technology matches relevant legal provisions, accounting standards, and tax regulations in the financial regulation database, providing prompts for non-compliant clauses to help quickly identify violations in financial operations and provide timely legal basis for subsequent handling. Finally, based on reports of abnormal fluctuations in financial data, non-compliant clause prompts, and user-input prompts, natural language processing technology matches relevant historical audit cases in the financial case database, recommending historical cases and corresponding solutions. This improves the accuracy and efficiency of financial risk response by drawing on past experience and handling solutions. Based on violation clause alerts, historical cases, and corresponding solutions, comprehensive regulatory recommendations are generated to help financial managers develop more compliant financial processing solutions based on the latest regulations and auditing experience, ensuring financial compliance and audit effectiveness, and improving the company's financial management level and decision-making capabilities.

[0032] In this embodiment, S6 specifically includes: S61. Based on financial data, abnormal data items, reconciliation relationships, and comprehensive regulatory recommendations, financial audit data is obtained by combining them. S62. Design a standardized audit report template, map financial audit data to the corresponding fields in the audit report template, and perform formatting and calculation to generate an audit report; S63. Check whether the data entered in the audit report meets the requirements, including the correctness of the amount, the matching of accounts, and the consistency of logic. Check whether there are any errors or omissions according to the audit key point rule base. If any inconsistencies or errors are found, the system will give an error message to remind the auditor to modify the audit report. S64. Auditors review and revise the generated audit report, and finally confirm and approve it through electronic signature; S65. Archive the final confirmed audit report and provide sharing and access control functions to ensure the secure storage and sharing of the report; S66 provides the function of exporting audit reports, supporting multiple export formats.

[0033] This invention combines financial data, abnormal data items, reconciliation relationships, and comprehensive regulatory recommendations to obtain financial audit data, ensuring the accuracy and completeness of financial reports. It automatically integrates information from multiple data sources to generate financial audit data that meets audit standards. A standardized audit report template is designed, mapping financial audit data to corresponding fields in the template, formatting and calculating them to generate the audit report. This ensures the structural standardization and content consistency of the audit report, providing a convenient foundation for subsequent review and approval. The system checks whether the filled data in the audit report meets requirements, including the correctness of amounts, account matching, and logical consistency. It checks for errors or omissions based on the audit key point rule base. If inconsistencies or errors are found, the system provides error messages, reminding auditors to modify the audit report. This step effectively prevents errors during report generation, ensuring the accuracy of the report. Auditors review and modify the generated audit report, ultimately confirming and approving it through electronic signatures. This ensures the standardization and compliance of the report review process, while electronic signatures enhance the legal validity of the report. The system archives the final confirmed audit report and provides sharing and access control functions to ensure the secure storage and sharing of the report. It offers persistent storage and flexible sharing of audit reports for enterprises, enhancing information security and convenience. The system also provides an audit report export function, supporting multiple export formats to meet different business needs and format requirements, enabling flexible sharing of reports across different systems and platforms.

[0034] refer to Figure 1 An AI-driven tax, accounting, and auditing system, comprising the following modules: The data acquisition and preprocessing module is used to collect and import multi-source heterogeneous financial data, preprocess the financial data, and store it into a historical financial dataset. The optimization algorithm module is used to input historical financial datasets into the xLSTM model, introduce learning rate warm-up and cosine annealing mechanisms, and iteratively optimize the xLSTM model through an improved AdamE optimizer. The knowledge graph module for reconciliation relationships is used to define and configure the rule base for audit key points. It extracts the knowledge graph of reconciliation relationships of financial data through the xLSTM model. The financial fluctuation tracking module is based on real-time tracking of financial data fluctuations. It uses an xLSTM model to predict the financial data fluctuation results, marks abnormal data by prediction errors, and uses a knowledge graph of correlation relationships to track the source of anomalies and generate a report on abnormal financial data fluctuations. The comprehensive regulatory advice module is used to match relevant regulatory provisions and historical cases with reports of abnormal fluctuations in financial data and user prompts to generate comprehensive regulatory advice. The audit report module is used to generate standardized audit reports, verify data accuracy, and support archiving, sharing, and export functions after auditors review and approve them.

[0035] Example 1: To verify the feasibility of this invention in intelligent management of taxation, accounting, and auditing, the method of this invention was applied to the tax calculation, financial risk assessment, and audit compliance detection system of a large enterprise's financial intelligent analysis and management platform (hereinafter referred to as "Platform F"). In traditional enterprise financial management systems, manual accounting, experience-based judgment, and rule-based financial risk analysis and auditing methods are commonly used. These methods are not only cumbersome and slow in response, but also struggle to handle complex financial data and tax-related textual information, resulting in insufficient predictive accuracy and low financial management efficiency. To address these issues, Platform F decided to adopt the AI-driven taxation, accounting, and auditing method and system proposed in this invention.

[0036] During implementation, Platform F first collects image and text data from corporate financial statements, tax documents, and audit reports. It then performs image resizing, enhancement, and normalization, while simultaneously cleaning, segmenting, mapping, and normalizing the text data to form a unified multimodal feature vector, which is input into the xLSTM model. Through the xLSTM model, Platform F can deeply learn the long-term dependencies in financial data and, combined with an improved AdamE optimizer, optimize model parameters, thereby efficiently and accurately analyzing and predicting financial data. Through the xLSTM model, Platform F can not only identify potential patterns in financial data but also monitor corporate financial status in real time, performing tax calculations, financial risk assessments, and audit compliance checks.

[0037] Based on the financial intelligent analysis and management system of this invention, Platform F has successfully achieved real-time prediction of corporate tax payables, financial risk scoring analysis, and audit compliance detection, significantly improving system operating efficiency and the accuracy of financial analysis. During implementation, the Platform F technical team found that, compared with traditional methods, the method of this invention significantly improves the accuracy of tax prediction, the precision of financial risk assessment, and the efficiency of audit compliance detection.

[0038] To further verify the effectiveness of the method of the present invention, platform F compared the performance of the method of the present invention with that of traditional methods. The specific comparison data is shown in Table 1: Table 1. Performance Comparison of Financial Intelligent Analysis and Management System on Platform F

[0039] As shown in Table 1, the application of the method of this invention significantly improves the accuracy of financial intelligent analysis and management on platform F in terms of tax calculation accuracy, financial risk assessment precision, and audit compliance detection efficiency. The accuracy rate of tax payable prediction increased from 80.5% using traditional methods to 96.8%, the accuracy rate of financial risk scoring prediction increased from 76.3% to 94.2%, and the accuracy rate of audit compliance detection increased from 78.7% to 95.5%, significantly enhancing the system's ability to predict and detect financial conditions and risks. The average analysis response time was reduced from 180 seconds to 32 seconds, an improvement of 82.2%; data processing capacity increased from 120 records / second to 560 records / second, and the number of clients analyzed simultaneously increased from 3 to 12, greatly improving the system's concurrent processing and analysis efficiency. The average time for financial risk identification and location decreased from 60 minutes to 12 minutes, and the average monthly workload of financial personnel decreased from 160 hours to 48 hours, significantly reducing the need for manual analysis. The overall monthly average financial management cost decreased from 350,000 yuan to 220,000 yuan, representing a cost saving of 37.1%.

[0040] Through the method of this invention, platform F can more accurately and efficiently complete corporate tax calculations, financial risk assessments, and audit compliance checks, monitor changes in financial data in real time, and predict potential risks in advance, significantly improving the intelligence level and decision-making efficiency of corporate financial management. This invention not only improves the system's automation and operational efficiency, significantly reducing the need for manual intervention by financial personnel, but also enhances the stability and robustness of the financial analysis and management system, providing effective technical support for intelligent corporate financial management.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI artificial intelligence driven tax, accounting, auditing method, characterized in that, The method comprises the following steps: S1, collecting and importing multi-source heterogeneous financial data, pre-processing the financial data, and storing the data into a historical financial data set; S2, inputting the historical financial data set into an xLSTM model, introducing a learning rate preheating and cosine annealing mechanism, and iteratively optimizing the xLSTM model through an improved AdamE optimizer; S3, defining and configuring an audit point rule library, extracting a financial data reconciliation relationship knowledge graph through the xLSTM model; S4, real-time tracking of financial data fluctuations, using the xLSTM model to predict financial data fluctuation results, marking abnormal data through prediction errors, tracking abnormal sources using the reconciliation relationship knowledge graph, and generating a financial data abnormal fluctuation report; S5, according to the financial data abnormal fluctuation report and user prompts, matching relevant regulations and historical cases, and generating comprehensive regulation suggestions; S6, generating a standardized audit report, checking data accuracy, and supporting archiving, sharing and exporting functions after being reviewed and approved by an auditor.

2. The AI artificial intelligence driven tax, accounting, auditing method of claim 1, wherein, The financial data includes account data, Excel tables, financial records in bank accounts and transaction systems, financial statements, and external data sources; the preprocessing includes removing duplicate data, repairing missing data, detecting abnormal data, unifying formats, converting data types, removing noise, merging data from different sources, eliminating redundant information, and unifying time series.

3. The AI artificial intelligence driven tax, accounting, auditing method of claim 1, wherein, The improved AdamE optimizer specifically includes: Initializing model parameters, initializing learning rate, decay factor, first moment estimation, second moment estimation, and bias correction term; Through back propagation, the gradient of the loss function with respect to each parameter is calculated; Based on the gradient, the first moment estimation and the second moment estimation are calculated; Based on the bias correction term, the bias correction is performed, and the first moment estimation and the second moment estimation are updated; The negative number of the training step number is multiplied by the decay factor to perform exponential calculation with the initial learning rate as the base, to obtain an updated learning rate that gradually decreases with the training progress; The model parameters are updated using the gradient, the first moment estimation, the second moment estimation, and the updated learning rate; The learning rate value is dynamically adjusted according to the current iteration number, and in the preheating stage, the learning rate gradually increases linearly from the initial learning rate to the target learning rate of the preheating stage; After the preheating stage ends, the ratio of the current iteration number to the total iteration number in the cosine annealing stage is calculated to obtain a progress ratio value, which is mapped into a cosine function to calculate the corresponding cosine value, and the cosine value is linearly transformed to obtain a cosine annealing adjustment factor; Based on the cosine annealing adjustment factor, the difference between the target learning rate of the preheating stage and the minimum learning rate is calculated, and then multiplied by the cosine annealing adjustment factor to obtain a learning rate adjustment amount that varies with the iteration number, and the learning rate value of the current iteration step is obtained by subtracting the learning rate adjustment amount from the target learning rate of the preheating stage; with the increase of the iteration number, the learning rate gradually decreases in a cosine form until it reaches the minimum learning rate; Continuously iteratively optimize until the loss function converges.

4. The AI artificial intelligence driven tax, accounting, auditing method of claim 1, wherein, S3 specifically includes: S31, according to the financial industry standards and historical business experience, defining the audit point rule library, including voucher integrity, subject compliance, and amount accuracy; S32, converting the audit focus rule library into a fixed dimension vector representation, using an embedding layer to convert the audit focus rule library into an audit focus rule feature vector through dimension reduction operation, and converting the historical financial data set into a feature vector of each financial item; S33, inputting the feature vector of each financial item and the audit focus rule feature vector into an LSTM layer, gradually learning the long-term dependency in the financial data through time steps, and at each time step, the LSTM layer calculates and updates the hidden state vector according to the latest input financial feature vector and the hidden state of the previous time step; the financial items include subjects, amounts, and vouchers; S34, based on the hidden state vector, using an attention mechanism to calculate the weight of each financial item according to the hidden state vector and the feature vector of each financial item; and performing weighted summation on the feature vectors of the financial items according to the weight of each financial item to generate a mapping relationship aggregation vector as a financial item mapping relationship set; S35, based on the financial item mapping relationship set, taking each financial item as a node and the mapping relationship between the financial items as an edge, a financial data reconciliation relationship knowledge graph is constructed.

5. The AI artificial intelligence driven tax, accounting, auditing method of claim 1, wherein, The S4 specifically includes: S41, real-time monitoring and tracking of financial data fluctuations, inputting historical financial data into an xLSTM model, and using an embedding layer to convert the historical financial data set into a feature vector of each financial item through dimension reduction operation; S42, inputting the feature vector of each financial item and the audit focus rule feature vector into an LSTM layer, gradually learning the long-term dependency in the financial data through time steps, and at each time step, the LSTM layer calculates and updates the hidden state vector according to the latest input financial feature vector and the hidden state of the previous time step; S43, inputting the hidden state vector into a fully connected layer, processing through linear transformation and linear activation function, predicting the feature vector of each financial item in the future, and obtaining a financial data prediction value; S44, comparing the financial data prediction value with the true value of the financial data to calculate the prediction error; S45, setting a prediction error threshold, if the prediction error exceeds the prediction error threshold, marking the true value of the financial data as an abnormal data item, and recording the abnormal type; S46, according to the abnormal data item, extracting the reconciliation relationship of the abnormal data item through the financial data reconciliation relationship knowledge graph to obtain all associated data items; S47, checking all associated data items layer by layer, tracking the reconciliation relationship of the abnormal data, gradually narrowing the range, first checking the total account in the financial statements, and then gradually checking the specific subjects, vouchers and detailed accounts, until the specific abnormal source is located; S48, generating a financial data abnormal fluctuation report according to the abnormal data item, the abnormal type, the associated data item, the reconciliation relationship and the abnormal source.

6. The AI artificial intelligence driven tax, accounting, auditing method of claim 1, wherein, The S5 specifically includes: S51, collecting the latest financial and auditing regulations, constructing and continuously updating a financial regulation library; collecting historical audit case data, including common financial misstatements, violation clauses and processing suggestions in the audit process, constructing and continuously updating a financial case library; S52, according to the financial data abnormal fluctuation report and the user input prompt statement, through natural language processing technology, matching the relevant legal provisions, accounting standards, tax regulations in the financial regulations library, giving the violation clause prompt; S53, according to the financial data abnormal fluctuation report, violation clause prompt and user input prompt statement, through natural language processing technology, matching the relevant historical audit cases in the financial case library, recommending historical cases and corresponding solutions; S54, based on the violation clause prompt, historical cases and corresponding solutions, generating comprehensive regulations suggestions.

7. The AI artificial intelligence driven tax, accounting, auditing method of claim 1, wherein, The S6 specifically includes: S61, based on the financial data, abnormal data items, reconciliation relationship, comprehensive regulations suggestions, combining to obtain the financial audit data; S62, designing a standardized audit report template, mapping the financial audit data to the corresponding fields in the audit report template, and formatting and calculating to generate the audit report; S63, checking whether the filled data in the audit report meets the requirements, including the correctness of the amount, subject matching, logical consistency, checking whether there are errors or omissions according to the audit point rule library, if inconsistencies or errors are found, the system will give error prompt information to remind the auditor to modify the audit report; S64, the auditor reviews the generated audit report and makes modifications, finally confirms and approves by electronic signature; S65, archiving the final confirmed audit report, and providing sharing and permission management functions to ensure the safe storage and sharing of the report; S66, providing the function of exporting the audit report, supporting multiple export formats.

8. An AI artificial intelligence driven tax, accounting, auditing system, executing an AI artificial intelligence driven tax, accounting, auditing method according to any one of claims 1 to 7, characterized in that, The following modules are included: Data acquisition and preprocessing module, used for collecting and importing multi-source heterogeneous financial data, preprocessing the financial data, and storing it into the historical financial data set; Optimization algorithm module, used for inputting the historical financial data set into the xLSTM model, introducing learning rate preheating and cosine annealing mechanism, and using the improved AdamE optimizer to iteratively optimize the xLSTM model; Reconciliation relationship knowledge graph module, used for defining and configuring the audit point rule library, extracting the financial data reconciliation relationship knowledge graph through the xLSTM model; Financial fluctuation tracking module, based on real-time tracking of financial data fluctuation, using the xLSTM model to predict the financial data fluctuation result, marking abnormal data through prediction error, tracing the abnormal source using the reconciliation relationship knowledge graph, and generating the financial data abnormal fluctuation report; Comprehensive regulations suggestion module, used for matching relevant regulations and historical cases according to the financial data abnormal fluctuation report and user prompts, and generating comprehensive regulations suggestions; Audit report module, used for generating standardized audit reports, checking data accuracy, and supporting archiving, sharing and exporting functions after the auditor reviews and approves.

Images