Automatic financial information processing method based on AI

Abstract

The invention discloses an AI-based automatic financial information processing method, and relates to the field of financial automation, and the method comprises the steps: achieving the automatic collection and storage of structured and unstructured data through the access of enterprise multi-source financial data; systematic preprocessing is carried out on the collected multi-source heterogeneous financial data, and a unified and high-quality financial data set is constructed; based on natural language processing and a knowledge graph technology, performing text semantic understanding, transaction automatic classification, field standardization and label generation on the cleaned and integrated financial data; comprehensively quantifying enterprise operation and financial performance based on the structured transaction data and the semantic annotation result; based on historical financial indexes, establishing a multi-model architecture to predict key financial variables; and based on the structured data, the prediction result and the historical rule, identifying potential financial abnormity and risk behaviors, and realizing intelligent early warning. According to the method, the intelligence, the real-time performance and the accuracy of financial information processing can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of financial automation, and in particular to an AI-based automated financial information processing method. Background Art

[0002] As internet platform companies diversify their businesses, their financial data is characterized by diverse sources, complex structures, and frequent updates. Traditional financial information processing relies on manual data entry, manual calculations, and empirical analysis, which is not only inefficient but also difficult to adapt to real-time decision-making, risk management, and compliance requirements.

[0003] Most current systems only support the processing of structured data and have difficulty covering unstructured text information such as invoices and reimbursement forms; semantic classification relies on fixed templates and cannot automatically aggregate complex financial events; the calculation of financial indicators lacks scalability and cannot flexibly adapt to dynamic business changes; at the same time, there is a lack of intelligent prediction mechanisms and anomaly monitoring capabilities, which makes it difficult to identify business risks in a timely manner, affecting enterprise management efficiency.

[0004] To address the above problems, the present invention proposes an AI-based automated financial information processing method that integrates OCR, NLP, machine learning, and time series prediction technologies to achieve intelligent processing of the entire process from financial data collection, cleaning, semantic classification, indicator modeling to risk warning and report output, significantly improving data processing capabilities and decision support levels. Summary of the Invention

[0005] In response to the above problems, the present invention provides an AI-based automated financial information processing method to solve the problems of inaccurate semantic classification, non-automatic indicator calculation and untimely risk identification in the prior art.

[0006] To solve the above technical problems, the present invention provides the following technical solution: an AI-based automated financial information processing method, comprising the following steps: Step S1: By accessing the enterprise's multi-source financial data, the system automatically collects and stores structured and unstructured data. The automatic collection covers four aspects: connection method, identification method, legality verification, and storage strategy. In step S1, the following sub-steps are also included: S1-1: Identify and connect to the enterprise's multi-source financial data with a formatted structure. Formatted structures refer to data types with fixed field formats that can be directly parsed. These include: e-commerce platform transaction flows, business vouchers generated by ERP systems, third-party payment platform logs, account tables generated by financial accounting systems, reports exported from tax filing systems, and bank statement reconciliation files. The above data is acquired in real time and in batches through the RESTful API interface. The data acquisition process retains the original format of the data and adds metadata tags, including the source system, acquisition time and business module. S1-2: For unstructured financial documents in paper and image formats, including invoices, receipts, contract scans, and expense reports, data structuring is accomplished through the following process: Call the high-precision OCR engine to recognize image data and use the predefined template extraction algorithm to map the recognized information in the image into standard fields; Among them, standard fields include: invoice code, invoice date, amount, tax rate, supplier name and taxpayer identification number; S1-3, during the collection process, perform the following verification operations on all structured and structured converted data: Permission and encryption protection: Determine whether access to a certain type of data source is allowed based on user identity and permission control logic; HTTPS / SSL encryption is used during data transmission; Authenticity verification means, performing data signature verification and certificate validity check on the received transaction data; Structural integrity assessment: using field integrity statistics, primary key duplication rates, and field value validity verification to determine whether data is missing or abnormal. S1-4: Implement hierarchical management of collected data. All unprocessed raw data, including OCR original images and structured source tables, are stored in the raw data warehouse to retain audit traceability evidence. For structured data that has undergone OCR conversion and field extraction, as well as multi-source financial data with formatted structure collected by S1-1, they are written into the intermediate storage area according to the three-dimensional modeling of "business entity-time-data type".

[0007] Step S2: Systematically preprocess the collected multi-source heterogeneous financial data, including removing duplicate records, standardizing fields, handling missing and outlier values, and merging multiple sources, to build a unified, high-quality financial data set. In step S2, the following sub-steps are also included: S2-1, based on a set of fields with unique identification capabilities, duplicate records are identified using the following method: Field matching and deduplication: extracting single fields and field combinations as unique identifiers, determining conflicts between records with the same identifier value, and retaining only the latest version of the data. Hash fingerprint comparison: Generate hash values for key fields to detect duplicate records with consistent semantics but different data sources; Logically invalid record filtering removes records with a zero amount, "invalid" status, or missing customer number core fields; S2-2, data format standardization, convert the format of the cleaned data and perform standardization processing, specifically: Standardize time fields and convert all date and time fields into ISO format; Amount units are standardized, and all currency fields are converted to the base currency set by the system, with precision unified to two decimal places; Unified the precision of numeric fields, and unified the format of sales, tax, and exchange rate numeric fields to standard precision; Semantic field normalization: mapping enumeration values for supplier name, payment method, and invoice type fields to eliminate semantic inconsistencies caused by spelling and aliases. S2-3, missing value processing, different methods are used to fill missing values in key fields. For key fields, the historical mean backfill algorithm is used, and for category fields, the "highest historical frequency value" is used; Identify and eliminate outliers. Use the standard deviation detection algorithm for the amount and date fields to calculate the Z-score, as shown in the formula: in, Normal score for standard difference, is the current financial data value to be tested, is the historical average, is the standard deviation; S2-4, multi-source data integration, based on the unified field standards, associates and merges the same business entities from different systems, as follows: Business primary key binding, based on the customer ID, contract number, invoice number, transaction serial number primary key fields, perform association alignment; Cross-system field mapping: building a field mapping table to unify fields with the same meaning but different names in different systems; Entity-level integration output generates unified financial data records for each merged entity, forming an analysis-oriented "wide table" structure of financial data.

[0008] Step S3: Based on natural language processing and knowledge graph technology, the cleaned and integrated financial data is subjected to text semantic understanding, automatic transaction classification, field standardization, and label generation to form a "semantic financial data set" for intelligent analysis. In step S3, the following sub-steps are also included: S3-1, text feature extraction, for various text fields, text structure processing, extracting semantic vector features: text preprocessing, performing word segmentation, stop word removal, and part-of-speech tagging on financial texts; word vector generation, using TF-IDF vector construction for short texts, as shown in the formula: in, is the TF-IDF vector, is the frequency of word t in document d, is the number of documents in which word t appears, and N is the total number of documents. For long texts with strong contextual relevance, the pre-trained BERT model is used to output semantic sentence vectors. S3-2, transaction classification and account grouping, inputs text vectors into the classification model to achieve automatic transaction classification and accounting account labeling, uses a multi-layer perceptron to classify transaction texts; and generates semantic transaction labels. Accounting account mapping: automatically binds semantic tags to specific account codes based on the standard accounting chart configured by the enterprise; The classification results are written back into the data table, forming a semantic structure of "transaction data + transaction type + accounting subject"; S3-3, Entity Recognition and Standardization, identifies and standardizes proprietary entities such as company names, suppliers, and category names in the text: A conditional random field model combined with a dictionary is used to identify company names, banks, platforms, and product names that appear in the text; standard code binding is used to uniformly bind different descriptions of the same manufacturer to a standard supplier code; Rule base updates, with automatic incremental learning of custom dictionaries and mapping rules, to adapt to new businesses and new expression variants; S3-4, semantic tag management, unifies the processing results of S3-1 to S3-3 and forms structured tags: generates several semantic tags for each transaction record, and the semantic tags are bound to the original records to form a multi-dimensional tag structure.

[0009] Step S4: Based on the structured transaction data and semantic annotation results, automatically generate accounting indicators, business operation ratios, trend statistics, and custom business indicators to comprehensively quantify the enterprise's operations and financial performance; In step S4, the following sub-steps are also included: S4-1, basic indicator calculation, based on the collected accounting items, summarize each type of transaction within the item and calculate basic indicators, including total revenue, total cost, net profit, and non-operating income. All indicators are dynamically generated on a daily / monthly / quarterly basis; S4-2, Ratios and Comprehensive Indicators: Based on accounting and operational indicators, construct financial ratios that reflect the health of the company. Financial ratios include classic financial ratios and comprehensive business indicators. Classic financial ratios include debt-to-asset ratio, current ratio, and return on equity; comprehensive business indicators include customer conversion rate, platform commission rate, and return rate. S4-3, time series indicators, generates trend indicators based on historical data, including year-on-year growth rate, month-on-month growth rate, and rolling average. The year-on-year growth rate can be expressed as ; S4-4, custom analysis indicators, calculates user dimension indicators and delivery indicators based on the business characteristics of the Internet platform. User dimension indicators include: average revenue per user and customer acquisition cost; delivery indicators include: advertising ROI and inventory turnover rate; all generated financial indicators are stored in the database.

[0010] Step S5: Based on historical financial indicators and integrating external influencing factors, a multi-model architecture is established to predict key financial variables, providing forward-looking operational decision support for the enterprise; In step S5, the following sub-steps are also included: S5-1, feature construction and selection, the financial indicators and other relevant features generated above are combined as model input features, and the input features are composed of core features and external features; Core characteristics are historical indicators; external characteristics include macroeconomic data, promotion intensity and industry index; Feature engineering is used to extract time features and derived features to improve model performance. Time features include extracting monthly, quarterly, and holiday periodic variables. Derivative features include constructing sliding averages, year-on-year, and month-on-month features. S5-2, model training, select the appropriate algorithm according to the prediction requirements of different indicators. Linear modeling example, use the multivariate linear regression model to predict a financial indicator Y, as shown in the formula: in, is the target financial indicator value predicted by the model, is the intercept term of the regression model, is the regression coefficient of the i-th feature, is the i-th input eigenvalue, is the number of feature dimensions of the input; For indicators with time correlation, the neural network-based time series model LSTM is used. The LSTM model is used to learn long-term dependencies, and its input is a continuous financial indicator vector sequence within a time window. S5-3, prediction output. After the trained model is deployed, the new period features are input into the prediction engine, and the indicator forecast values for the next 1 to 3 periods are output. At the same time, the prediction confidence interval is calculated, and the upper and lower bounds are calculated using the 95% confidence interval to reflect the uncertainty of the prediction. S5-4, model iterative update. As new data continues to arrive, the model is retrained and updated regularly. If the error is greater than the set threshold, the system triggers automatic retraining of the model.

[0011] Step S6: Based on structured data, prediction results, and historical patterns, identify potential financial anomalies and risky behaviors, and implement intelligent early warning through scoring modeling and rule engines; In step S6, the following sub-steps are also included: S6-1, Outlier Detection, performs periodic anomaly monitoring on key indicators using a sliding statistics method: constructing a sliding window mean and standard deviation for each indicator to determine whether its mean exceeds 3σ. In addition, using isolation forests and rarity scores based on sample segmentation, unsupervised anomaly detection is performed. S6-2, risk score calculation, in order to quantify the risk level, a scoring model is used to assign scores to key indicators and abnormal events. Specifically, abnormal events and characteristic variables are input into the logistic regression risk scorer, as shown in the formula: in, Indicates the probability value of the transaction being "high risk", is the model bias term, is the weight coefficient of the j-th risk factor, is the j-th risk characteristic value, is the number of risk factors, It is a natural exponential function; the risk level is determined based on the risk probability and the comparison between the score and the threshold; S6-3, early warning rule execution, comparing detected anomalies and high-risk items with pre-set rules. If losses continuously exceed the threshold, an alarm is triggered and the system automatically generates an early warning event. Each early warning event includes: indicator name, abnormal magnitude, timestamp, associated predicted value, and recommended treatment measures, which are sent to the operator; S6-4, decision recommendation output, based on the analysis results and risk level, the system intelligently generates decision recommendations.

[0012] Step S7: Organize the analysis results and generate an interactive report with both pictures and text to support management's daily business decisions and audit compliance needs.

[0013] In step S7, the following sub-steps are also included: S7-1, Report content arrangement, organizes key financial indicators, historical trends, forecast results and risk warning content into a report structure according to predefined templates, including summary, financial overview, forecast analysis, and risk warning; S7-2, Visual Presentation, uses graphic elements generated by chart components to visualize important data. Line charts show revenue growth trends, pie charts show cost structures, heat maps indicate risk levels, and donut charts and funnel charts show user conversion and return on investment analysis. S7-3, natural language summary, uses natural language generation technology to automatically write text descriptions based on the analysis results to summarize the key findings in the report; S7-4, output and distribution, exports the generated reports into PDF, PPT, and web page formats, and distributes them to relevant decision makers or system archives according to their permissions. The report generation process can be automatically executed on a regular basis.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention supports automatic access to multiple heterogeneous data sources such as ERP systems, payment platforms, tax interfaces, bank receipts, e-commerce transaction platforms, etc., and combines structured data interfaces and OCR analysis of unstructured bill images to achieve unified collection of all types of financial data. Through field standardization, semantic mapping, currency conversion, time format unification and other operations, a consistent data model is constructed, which solves the problems of fragmented system data sources, inconsistent formats and difficult integration, and significantly improves the data base quality of the financial information system.

[0015] The present invention uses NLP technology to perform deep feature extraction on text fields such as transaction descriptions and invoice notes, combines support vector machines or neural network classification models to perform semantic classification of financial transactions, and automatically matches standard accounting subjects, so that transactions with different descriptions can be automatically aggregated into standard subjects such as "sales revenue", "administrative expenses", and "advertising expenses". At the same time, it identifies and unifies entity information such as supplier names and payer identifiers, solving the problem of low accuracy and high maintenance costs of rigid classification methods that rely on templates and rules when facing complex business scenarios.

[0016] The present invention integrates multiple anomaly detection algorithms such as standard deviation recognition and isolation forest, and combines them with financial logic rules to realize real-time monitoring of abnormal transaction amounts, cash flow breakpoints, and abnormal changes in accounts; it assigns a risk level to each record through a risk scoring model, triggers system-level early warning events, generates risk warnings and corresponding management suggestions, and realizes a complete closed loop from anomaly discovery, qualitative assessment to intervention suggestions, thereby improving the company's response speed and handling capabilities to potential financial risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the embodiments. It is understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in combination with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0020] Please refer to Figure 1 , Figure 1 The flowchart of an AI-based automated financial information processing method provided by an embodiment of the present invention includes the following steps: Step S1: By accessing the enterprise's multi-source financial data, the system automatically collects and stores structured and unstructured data. The automatic collection covers four aspects: connection method, identification method, legality verification, and storage strategy. S1-1: Identify and connect to the enterprise's multi-source financial data with a formatted structure. Formatted structures refer to data types with fixed field formats that can be directly parsed. These include: e-commerce platform transaction flows, business vouchers generated by ERP systems, third-party payment platform logs, account tables generated by financial accounting systems, reports exported from tax declaration systems, and bank statement reconciliation files.

[0021] The above data is acquired in real time and in batches through the RESTful API interface. The data is collected in its original format and metadata tags are added, including the source system, collection time, and business module, for subsequent tracing. S1-2: For unstructured financial documents in paper and image formats, including invoices, receipts, contract scans, and expense reports, data structuring is accomplished through the following process: Calling a high-precision OCR (Optical Character Recognition) engine to recognize image data; using a predefined template extraction algorithm to map the recognized information in the image into standard fields; Standard fields include: invoice code, invoice date, amount, tax rate, supplier name, taxpayer identification number, etc. After the above operations, unstructured data is converted into a structured format and enters a unified data preprocessing channel for subsequent verification and cleaning. S1-3, data standardization and compliance verification, during the collection process, all structured and structured converted data are uniformly verified as follows: Permission and encryption protection: Based on user identity and permission control logic, it is determined whether access to a certain type of data source is allowed; HTTPS / SSL encryption is used during data transmission to ensure link security; Authenticity verification means, performing data signature verification and certificate validity check on the received transaction data; Structural integrity assessment: using field integrity statistics, primary key duplication rates, and field value validity verification to determine whether data is missing or abnormal. Through the above operations, the system can promptly detect forged data, records with disordered formats, and illegal collection behaviors, and generate verification logs and error reports; S1-4: Implement hierarchical management of collected data. All unprocessed raw data, including OCR original images and structured source tables, are stored in the original data warehouse or data lake to retain audit traceability evidence. For structured data that has undergone OCR conversion and field extraction, as well as multi-source financial data with formatted structures collected by S1-1, they are written into the intermediate storage area according to the three-dimensional modeling of "business entity-time-data type" for subsequent use.

[0022] It should be noted that all data storage uses the AES-256 encryption mechanism and performs periodic backups; metadata records file fingerprints, update time, operation logs and other information to facilitate permission review and data consistency checks.

[0023] Step S2, data cleaning and preprocessing, systematically preprocesses the collected multi-source heterogeneous financial data, including duplicate record removal, field standardization, missing and outlier processing, and multi-source merging operations, to build a unified, high-quality financial data set to provide consistent input for subsequent intelligent analysis; S2-1, based on a set of fields with unique identification capabilities, such as invoice number, contract number, and transaction serial number, duplicate record identification is performed using the following method: Field matching and deduplication: extract a single field or a combination of fields as a unique identifier, perform conflict determination on records with the same identifier value, and retain only the latest version of the data. Hash fingerprint comparison: Generate hash values for key fields to detect duplicate records with consistent semantics but different data sources; Logically invalid record filtering: remove records with zero amount, "invalid" status, missing customer number and other core fields; S2-2, data format standardization, convert the format of the cleaned data and perform standardization processing, specifically: Standardize time fields and convert all date and time fields into ISO format; Amount units are standardized, and all currency fields are converted to the base currency set by the system, with precision unified to two decimal places; Unify the precision of numerical fields, unify the format of numerical fields such as sales, tax, exchange rate, etc. to standard precision to maintain comparability between fields; Semantic field normalization: mapping enumeration values for fields such as supplier name, payment method, and invoice type to eliminate semantic inconsistencies caused by spelling and aliases.

[0024] S2-3, missing value processing, uses different methods to fill missing values in key fields. For key fields such as amount, quantity, tax rate, etc., the historical mean backfill algorithm is used. If the field is a category type, such as "invoice type", the "highest historical frequency value" is used for filling; Identify and eliminate outliers. Use the standard abnormality detection algorithm for fields such as amount and date to calculate the Z-score, as shown in the formula: in, Normal score for standard difference, is the current financial data value to be tested, is the historical average, is the standard deviation; if |Z| exceeds the set threshold, it is marked as abnormal and corrected or eliminated according to business rules; S2-4, multi-source data integration, based on the unified field standards, associates and merges the same business entities from different systems, as follows: Business primary key binding, performing association alignment based on primary key fields such as customer ID, contract number, invoice number, transaction serial number, etc. Cross-system field mapping: Build a field mapping table to consolidate fields with the same meaning but different names in different systems, such as "Customer Number" and "User ID"; Entity-level integration output generates unified financial data records for each merged entity, including income, expenditure, tax information, transaction status, etc., forming an analysis-oriented "wide table" structure of financial data.

[0025] Step S3, intelligent semantic classification, uses natural language processing (NLP) and knowledge graph technology to perform text semantic understanding, automatic transaction classification, field standardization, and label generation on the cleaned and integrated financial data, forming a "semantic financial data set" for intelligent analysis; S3-1, text feature extraction, performs text structured processing on various text fields, such as transaction notes, invoice content, and contract summaries, and extracts semantic vector features: Text preprocessing, performs basic processing such as word segmentation, stop word removal, and part-of-speech tagging on financial text; word vector generation, uses TF-IDF vector construction for short texts, as shown in the formula: in, is the TF-IDF vector, is the frequency of word t in document d, is the number of documents in which word t appears, and N is the total number of documents. For long texts with strong contextual relevance, the pre-trained BERT model is used to output semantic sentence vectors. S3-2, transaction classification and account grouping, inputs text vectors into the classification model to achieve automatic transaction classification and accounting account labeling, using a multi-layer perceptron (MLP) to classify transaction text; semantic transaction label generation, for example, classifying "Alipay advertising" as "marketing expenses" and "XX logistics fees" as "transportation expenses"; Accounting account mapping: automatically binds semantic tags to specific account codes based on the standard accounting chart configured by the enterprise; The classification results are written back into the data table, forming a semantic structure of "transaction data + transaction type + accounting subject"; S3-3, Entity Recognition and Standardization, identifies and standardizes proprietary entities such as company names, suppliers, and category names in text. A Conditional Random Field (CRF) model combined with a dictionary is used to identify entities such as company names, banks, platforms, and product names that appear in the text. Standardized code binding is used to uniformly bind different descriptions of the same manufacturer to a standardized supplier code. Rule base updates, with automatic incremental learning of custom dictionaries and mapping rules, to adapt to new businesses and new expression variants; S3-4, semantic tag management, unifies the processing results of S3-1 to S3-3 and forms structured tags: generates several semantic tags for each transaction record, binds the semantic tags to the original records to form a multi-dimensional tag structure, and writes all structured outputs into the "semantic tag database". This semantic information provides semantically rich support for subsequent indicator calculation and risk analysis.

[0026] Step S4: Financial indicator generation: Based on structured transaction data and semantic annotation results, accounting indicators, business operation ratios, trend statistics, and custom business indicators are automatically generated to comprehensively quantify the company's operations and financial performance; S4-1, basic indicator calculation, based on the collected accounting items, summarize each type of transaction within the item and calculate basic indicators, including total revenue, total cost, net profit, and non-operating income. All indicators are dynamically generated on a daily / monthly / quarterly basis; S4-2, Ratios and Comprehensive Indicators: Based on accounting and operational indicators, construct financial ratios that reflect the health of the company. Financial ratios include classic financial ratios and comprehensive business indicators. Classic financial ratios include debt-to-asset ratio, current ratio, and return on equity; comprehensive business indicators include customer conversion rate, platform commission rate, and return rate. S4-3, time series indicators, generates trend indicators based on historical data, including year-on-year growth rate, month-on-month growth rate, and rolling average. The year-on-year growth rate can be expressed as ; S4-4, Custom Analysis Indicators, calculates user-specific and advertising-related metrics based on the business characteristics of internet platforms. User-specific metrics include average revenue per user and customer acquisition cost; advertising-related metrics include advertising ROI and inventory turnover. All generated financial indicators are stored in a database and annotated with timestamps to provide input for predictive modeling and subsequent analysis.

[0027] Step S5: Based on historical financial indicators and integrating external influencing factors, a multi-model architecture is established to predict key financial variables, providing forward-looking operational decision support for the enterprise; S5-1, feature construction and selection, the financial indicators and other relevant features generated above are combined as model input features, and the input features are composed of core features and external features; The core features are historical indicators, such as net profit, accounts receivable, and inventory levels in the past 12 months; External characteristics include macroeconomic data, promotion intensity, and industry indices; Feature engineering is used to extract time features and derived features to improve model performance. Time features include extracting monthly, quarterly, and holiday periodic variables. Derivative features include constructing sliding averages, year-on-year, and month-on-month features. S5-2, model training, select the appropriate algorithm according to the prediction requirements of different indicators. Linear modeling example, use the multivariate linear regression model to predict a financial indicator Y, as shown in the formula: in, is the target financial indicator value predicted by the model, is the intercept term of the regression model, is the regression coefficient of the i-th feature, is the i-th input eigenvalue, is the number of feature dimensions of the input; For indicators with time correlation, the neural network-based time series model LSTM is used. The LSTM model is used to learn long-term dependencies, and its input is a continuous financial indicator vector sequence within a time window. S5-3, prediction output. After the trained model is deployed, the new period features are input into the prediction engine, and the indicator prediction values for the next 1 to 3 periods are output. At the same time, the prediction confidence interval is calculated, and the upper / lower limits are calculated using the 95% confidence interval to reflect the uncertainty of the prediction.

[0028] S5-4, model iteration and update. With the continuous arrival of new data, the model is regularly retrained or updated. If the error is greater than the set threshold, the system triggers automatic model retraining. All model versions and iteration records are stored in the "model version control library" to support retrospective comparison and switch deployment.

[0029] Step S6: Based on structured data, prediction results and historical patterns, potential financial anomalies and risky behaviors are identified, and intelligent early warning is achieved through scoring modeling and rule engines.

[0030] S6-1, Outlier Detection, performs periodic anomaly monitoring on key indicators using a sliding statistics method: constructing a sliding window mean and standard deviation for each indicator to determine whether its mean exceeds 3σ. In addition, using isolation forests and rarity scores based on sample segmentation, unsupervised anomaly detection is performed. S6-2, risk score calculation, in order to quantify the risk level, a scoring model is used to assign scores to key indicators and abnormal events. Specifically, abnormal events and characteristic variables are input into the logistic regression risk scorer, as shown in the formula: in, Indicates the probability value of the transaction being "high risk", is the model bias term, is the weight coefficient of the j-th risk factor, is the j-th risk characteristic value, is the number of risk factors, It is a natural exponential function; the risk level is determined by comparing the risk probability or score with the threshold; S6-3, early warning rule execution, compares detected anomalies and high-risk items with pre-set rules. If losses continuously exceed the threshold or cash flow turns negative, an alarm is triggered and the system automatically generates an early warning event. Each early warning event includes: indicator name, abnormal magnitude, timestamp, associated predicted value, and recommended treatment measures, which are sent to operations personnel or management; S6-4, decision-making recommendation output. Based on the analysis results and risk level, the system intelligently generates decision-making recommendations. For example, in the case of predicted cash flow shortages, it is recommended to control costs or raise funds; for discovered abnormal transactions, it is recommended to audit or review; all risk warning information and recommendations are included in the subsequent report.

[0031] Step S7, report generation, organizes the analysis results and generates an interactive report with both pictures and text to support management's daily business decisions and audit compliance needs.

[0032] S7-1, Report content arrangement, organizes key financial indicators, historical trends, forecast results and risk warnings into a report structure according to predefined templates, including summary, financial overview, forecast analysis, risk warning and other sections; S7-2, Visual Presentation, uses graphic elements generated by chart components to visualize important data. Line charts show revenue growth trends, pie charts show cost structures, heat maps indicate risk levels, and donut charts / funnel charts show user conversion and return on investment analysis. S7-3, Natural Language Summary, uses natural language generation (NLG) technology to automatically write a text description based on the analysis results, summarizing the key findings in the report. For example, "In the first quarter of 2025, the company's operating income reached 238 million yuan, a year-on-year increase of 12.5%. The main growth came from the expansion of third-party platform business. The net profit for the next quarter is expected to be approximately 5.2 million yuan, which is expected to decline slightly. Attention should be paid to the continued increase in the proportion of marketing expenses." S7-4, Output and Distribution, exports generated reports to PDF, PPT, or web page formats and distributes them to relevant decision makers or system archives based on their permissions. The report generation process can be automatically executed on a regular basis or manually triggered as needed, ensuring that management receives accurate financial analysis results in a timely manner.

[0033] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. An AI-based automated financial information processing method, characterized in that: The following steps are involved: Step S1: By accessing the enterprise's multi-source financial data, the system automatically collects and stores structured and unstructured data. The automatic collection covers four aspects: connection method, identification method, legality verification, and storage strategy. Step S2: Systematically preprocess the collected multi-source heterogeneous financial data, including removing duplicate records, standardizing fields, handling missing and outlier values, and merging multiple sources, to build a unified, high-quality financial data set. Step S3: Based on natural language processing and knowledge graph technology, the cleaned and integrated financial data is subjected to text semantic understanding, automatic transaction classification, field standardization, and label generation to form a "semantic financial data set" for intelligent analysis. Step S4: Based on the structured transaction data and semantic annotation results, automatically generate accounting indicators, business operation ratios, trend statistics, and custom business indicators to comprehensively quantify the enterprise's operations and financial performance; Step S5: Based on historical financial indicators and integrating external influencing factors, a multi-model architecture is established to predict key financial variables, providing forward-looking operational decision support for the enterprise; Step S6: Based on structured data, prediction results, and historical patterns, identify potential financial anomalies and risky behaviors, and implement intelligent early warning through scoring modeling and rule engines; Step S7: Organize the analysis results and generate an interactive report with both pictures and text to support management's daily business decisions and audit compliance needs.

2. The AI-based automated financial information processing method according to claim 1, characterized in that: In step S1, the following sub-steps are also included: S1-1: Identify and connect to the enterprise's multi-source financial data with a formatted structure. Formatted structures refer to data types with fixed field formats that can be directly parsed. These include: e-commerce platform transaction flows, business vouchers generated by ERP systems, third-party payment platform logs, account tables generated by financial accounting systems, reports exported from tax filing systems, and bank statement reconciliation files. The above data is acquired in real time and in batches through the RESTful API interface. The data acquisition process retains the original format of the data and adds metadata tags, including the source system, acquisition time and business module. S1-2: For unstructured financial documents in paper and image formats, including invoices, receipts, contract scans, and expense reports, data structuring is accomplished through the following process: Call the high-precision OCR engine to recognize image data and use the predefined template extraction algorithm to map the recognized information in the image into standard fields; Among them, standard fields include: invoice code, invoice date, amount, tax rate, supplier name and taxpayer identification number; S1-3, during the collection process, perform the following verification operations on all structured and structured converted data: Permission and encryption protection: Determine whether access to a certain type of data source is allowed based on user identity and permission control logic; HTTPS / SSL encryption is used during data transmission; Authenticity verification means, performing data signature verification and certificate validity check on the received transaction data; Structural integrity assessment: using field integrity statistics, primary key duplication rates, and field value validity verification to determine whether data is missing or abnormal. S1-4: Implement hierarchical management of collected data. All unprocessed raw data, including OCR original images and structured source tables, are stored in the raw data warehouse to retain audit traceability evidence. Structured data that has undergone OCR conversion and field extraction, as well as formatted multi-source financial data collected by S1-1, are written into the intermediate storage area according to the three-dimensional modeling of "business entity-time-data type." 3. The AI-based automated financial information processing method according to claim 1, characterized in that: In step S2, the following sub-steps are also included: S2-1, based on a set of fields with unique identification capabilities, duplicate records are identified using the following method: Field matching and deduplication: extracting single fields and field combinations as unique identifiers, determining conflicts between records with the same identifier value, and retaining only the latest version of the data. Hash fingerprint comparison: Generate hash values for key fields to detect duplicate records with consistent semantics but different data sources; Logically invalid record filtering removes records with a zero amount, "invalid" status, or missing customer number core fields. S2-2, data format standardization, convert the format of the cleaned data and perform standardization processing, specifically: Standardize time fields and convert all date and time fields into ISO format; Amount units are standardized, and all currency fields are converted to the base currency set by the system, with precision unified to two decimal places; Unified the precision of numeric fields, and unified the format of sales, tax, and exchange rate numeric fields to standard precision; Semantic field normalization: mapping enumeration values for supplier name, payment method, and invoice type fields to eliminate semantic inconsistencies caused by spelling and aliases. S2-3, missing value processing, uses different methods to fill missing values in key fields. For key fields, the historical mean backfill algorithm is used, and for category fields, the "highest historical frequency value" is used. Identify and eliminate outliers. Use the standard deviation detection algorithm for the amount and date fields to calculate the Z-score, as shown in the formula: in, Normal score for standard difference, is the current financial data value to be tested, is the historical average, is the standard deviation; S2-4, multi-source data integration, based on the unified field standards, associates and merges the same business entities from different systems, as follows: Business primary key binding, based on the customer ID, contract number, invoice number, transaction serial number primary key fields, perform association alignment; Cross-system field mapping: building a field mapping table to unify fields with the same meaning but different names in different systems; Entity-level integration output generates unified financial data records for each merged entity, forming an analysis-oriented "wide table" structure of financial data.

4. The AI-based automated financial information processing method according to claim 1, characterized in that: In step S3, the following sub-steps are also included: S3-1, text feature extraction, for various text fields, text structure processing, extracting semantic vector features: text preprocessing, performing word segmentation, stop word removal, and part-of-speech tagging on financial texts; word vector generation, using TF-IDF vector construction for short texts, as shown in the formula: in, is the TF-IDF vector, is the frequency of word t in document d, is the number of documents in which word t appears, and N is the total number of documents. For long texts with strong contextual relevance, the pre-trained BERT model is used to output semantic sentence vectors. S3-2, transaction classification and account grouping, inputs text vectors into the classification model to achieve automatic transaction classification and accounting account labeling, uses a multi-layer perceptron to classify transaction texts; and generates semantic transaction labels. Accounting account mapping: automatically binds semantic tags to specific account codes based on the standard accounting chart configured by the enterprise; The classification results are written back into the data table, forming a semantic structure of "transaction data + transaction type + accounting subject"; S3-3, Entity Recognition and Standardization, identifies and standardizes proprietary entities such as company names, suppliers, and category names in the text: A conditional random field model combined with a dictionary is used to identify company names, banks, platforms, and product names that appear in the text; standard code binding is used to uniformly bind different descriptions of the same manufacturer to a standard supplier code; Rule base updates, with automatic incremental learning of custom dictionaries and mapping rules, to adapt to new businesses and new expression variants; S3-4, semantic tag management, unifies the processing results of S3-1 to S3-3 and forms structured tags: generates several semantic tags for each transaction record, and the semantic tags are bound to the original records to form a multi-dimensional tag structure.

5. The AI-based automated financial information processing method according to claim 1, characterized in that: In step S4, the following sub-steps are also included: S4-1, basic indicator calculation, based on the collected accounting items, summarize each type of transaction within the item and calculate basic indicators, including total revenue, total cost, net profit, and non-operating income. All indicators are dynamically generated on a daily / monthly / quarterly basis; S4-2, Ratios and Comprehensive Indicators: Based on accounting and operational indicators, construct financial ratios that reflect the health of the company. Financial ratios include classic financial ratios and comprehensive business indicators. Classic financial ratios include debt-to-asset ratio, current ratio, and return on equity; comprehensive business indicators include customer conversion rate, platform commission rate, and return rate. S4-3, time series indicators, generates trend indicators based on historical data, including year-on-year growth rate, month-on-month growth rate, and rolling average, where the year-on-year growth rate is expressed as ； S4-4, custom analysis indicators, calculates user dimension indicators and delivery indicators based on the business characteristics of the Internet platform. User dimension indicators include: average revenue per user and customer acquisition cost; delivery indicators include: advertising ROI and inventory turnover rate; all generated financial indicators are stored in the database.

6. The AI-based automated financial information processing method according to claim 1, characterized in that: In step S5, the following sub-steps are also included: S5-1, feature construction and selection, the financial indicators and other relevant features generated above are combined as model input features, and the input features are composed of core features and external features; Core characteristics are historical indicators; external characteristics include macroeconomic data, promotion intensity and industry index; Feature engineering is used to extract time features and derived features to improve model performance. Time features include extracting monthly, quarterly, and holiday periodic variables. Derivative features include constructing sliding averages, year-on-year, and month-on-month features. S5-2, model training, select the appropriate algorithm according to the prediction requirements of different indicators. Linear modeling example, use the multivariate linear regression model to predict a financial indicator Y, as shown in the formula: in, is the target financial indicator value predicted by the model, is the intercept term of the regression model, is the regression coefficient of the i-th feature, is the i-th input eigenvalue, is the number of feature dimensions of the input; For indicators with time correlation, the neural network-based time series model LSTM is used. The LSTM model is used to learn long-term dependencies, and its input is a continuous financial indicator vector sequence within a time window. S5-3, prediction output. After the trained model is deployed, the new period features are input into the prediction engine, and the indicator forecast values for the next 1 to 3 periods are output. At the same time, the prediction confidence interval is calculated, and the upper and lower bounds are calculated using the 95% confidence interval to reflect the uncertainty of the prediction. S5-4, model iterative update. As new data continues to arrive, the model is retrained and updated regularly. If the error is greater than the set threshold, the system triggers automatic retraining of the model.

7. The AI-based automated financial information processing method according to claim 1, characterized in that: In step S6, the following sub-steps are also included: S6-1, outlier detection, performs periodic anomaly monitoring on key indicators, sliding statistics method: constructs a sliding window mean and standard deviation for each indicator to determine whether its mean exceeds 3σ; In addition, we use isolation forest to perform unsupervised anomaly detection based on the rarity score of sample segmentation; S6-2, risk score calculation, in order to quantify the risk level, a scoring model is used to assign scores to key indicators and abnormal events. Specifically, abnormal events and characteristic variables are input into the logistic regression risk scorer, as shown in the formula: in, Indicates the probability value of the transaction being "high risk", is the model bias term, is the weight coefficient of the j-th risk factor, is the j-th risk characteristic value, is the number of risk factors, It is a natural exponential function; the risk level is determined based on the risk probability and the comparison between the score and the threshold; S6-3, early warning rule execution, comparing detected anomalies and high-risk items with pre-set rules. If losses continuously exceed the threshold, an alarm is triggered and the system automatically generates an early warning event. Each early warning event includes: indicator name, abnormal magnitude, timestamp, associated predicted value, and recommended treatment measures, which are sent to the operator; S6-4, decision recommendation output, based on the analysis results and risk level, the system intelligently generates decision recommendations.

8. The AI-based automated financial information processing method according to claim 1, characterized in that: In step S7, the following sub-steps are also included: S7-1, Report content arrangement, organizes key financial indicators, historical trends, forecast results and risk warning content into a report structure according to predefined templates, including summary, financial overview, forecast analysis, and risk warning; S7-2, Visual Presentation, uses graphic elements generated by chart components to visualize important data. Line charts show revenue growth trends, pie charts show cost structures, heat maps indicate risk levels, and donut charts and funnel charts show user conversion and return on investment analysis. S7-3, natural language summary, uses natural language generation technology to automatically write text descriptions based on the analysis results to summarize the key findings in the report; S7-4, output and distribution, exports the generated report into PDF, PPT, and web page formats, and distributes it to relevant decision makers or system archives according to their permissions. The report generation process is automatically executed regularly.

Images