A system and method for automatic extraction and reconciliation processing of financial data

Abstract

The present invention relates to the technical field of financial data processing, and specifically provides a system and method for automated extraction and reconciliation processing of financial data, including: a financial data acquisition module for acquiring financial image data, audio data, and bank API data. The financial data processing module uses OCR recognition, Whisper voice conversion, and bank API paging query technology to standardize the data and extract financial data through an anomaly detection mechanism. The financial data sorting module classifies and stores the data in different financial databases based on format verification and priority marking, using dynamic routing and four-way shunt strategies. The financial data reconciliation module verifies and processes differences in the data through cross-database reconciliation rules to generate a financial analysis report. The financial data display module monitors the financial analysis results using a data snapshot and rollback mechanism, achieving automated extraction, reconciliation, and visualization of financial data, and improving the accuracy and compliance of financial data.

Description

Technical Field

[0001] The present invention relates to the technical field of financial data processing, and in particular to a system and method for automatic extraction and cross-checking of financial data. Background Art

[0002] With the accelerated development of enterprise informatization, financial processing is transforming from traditional manual operations to intelligent and automated operations. The sources of financial data are becoming increasingly diversified, and their processing efficiency and accuracy have a direct impact on the quality of corporate financial management. The existing financial system has improved the efficiency of data entry and the level of reconciliation automation to a certain extent, but there are still many challenges in the automated extraction and cross-checking of financial data, which makes it difficult to meet the growing needs of enterprises for intelligent financial management.

[0003] First, the data entry efficiency is insufficient, which affects the accuracy of the data. Although the existing financial system can use OCR technology to extract invoice information, the recognition accuracy is limited due to factors such as seal occlusion and difficulty in handwriting recognition. In addition, the existing OCR technology mainly relies on character extraction and lacks the ability to understand contextual semantics, and cannot automatically verify the authenticity and integrity of the data.

[0004] Secondly, the accuracy of cross-checking is insufficient, which is prone to misreporting or omissions. These financial systems usually use a single database to store financial data, lack the ability to integrate heterogeneous data, and find it difficult to efficiently integrate invoice data, bank statements and other financial records from OCR recognition, which may lead to data isolation and affect the efficiency and accuracy of reconciliation.

[0005] Finally, the lack of traceability of financial data after correction affects management optimization. When financial data is abnormal and adjusted, the existing system lacks a complete audit trail, making it difficult to trace the cause of the abnormality. At the same time, due to the lack of data snapshot storage, it is difficult for the system to trace back to the original state when processing financial data anomalies.

[0006] In order to improve the accuracy and compliance of financial data, a system and method for automatic extraction and cross-checking of financial data are proposed. Summary of the invention

[0007] The purpose of the present invention is to provide a system and method for automatic extraction and cross-checking of financial data, including: a financial data acquisition module for acquiring financial image data, audio data and bank API data. The financial data processing module uses OCR recognition, Whisper voice conversion and bank API paging query technology to standardize the data and extract financial data through anomaly detection mechanism. The financial data sorting module uses dynamic routing and four-way diversion strategies based on format verification and priority marking to classify and store data in different financial databases. The financial data cross-checking module verifies and processes the data through cross-database cross-checking rules to generate a financial analysis report. The financial data display module uses data snapshots and rollback mechanisms to monitor the results of financial analysis, realizes the automatic extraction, cross-checking and visualization of financial data, and improves the accuracy and compliance of financial data.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A system for automatic extraction and cross-checking of financial data, comprising:

[0010] Financial data acquisition module, used to acquire financial image data, audio data and bank API data;

[0011] The financial data processing module includes: an OCR unit, which is used to perform seal processing, text recognition and semantic correction on the financial image data, and extract first financial data through a first abnormal mechanism; a speech unit, which is used to convert the audio data into financial text data through a Whisper model, and obtain second financial data according to confidence level classification; an API unit, which is used to obtain transaction flow from the bank API data using a paging query technology, standardize the transaction flow, and extract third financial data using a second abnormal mechanism;

[0012] A financial data sorting module is used to perform format verification and priority marking on the data packets of the financial data processing module; classify the data packets through dynamic routing, and adopt four-way diversion to store them in different types of financial databases;

[0013] A financial data cross-checking module is used to verify and process the differences of the financial database based on the cross-checking rules and generate a financial analysis report;

[0014] The financial data display module is used to monitor the financial analysis report using a data snapshot and rollback mechanism.

[0015] Furthermore, the implementation process of the OCR unit includes:

[0016] Performing geometric correction, denoising and color enhancement on the financial image data to generate pre-processed financial image data;

[0017] Using the PP-OCRv3 detection model to locate the text area of ​​the pre-processed financial image data, and performing seal interference detection, and if seal interference exists, repairing the text area;

[0018] Perform optical character recognition on the text area using the PP-OCRv3 recognition model to generate original financial text;

[0019] Performing context verification and logic authentication on the original financial text, and using a Transformer model to perform semantic correction, and outputting the first financial data;

[0020] The first exception mechanism is used to check the first financial data. If the text confidence is lower than the first threshold and any one of the identification failures of the same field occurs N times in a row, a manual review is triggered.

[0021] Furthermore, the implementation process of the speech unit includes:

[0022] Convert the audio data into financial text data using the Whisper speech recognition model and generate recognition confidence;

[0023] Classifying the recognition confidence, including high recognition confidence and low recognition confidence;

[0024] If the recognition confidence is the high recognition confidence, performing intent recognition on the financial text data using RasaNLU, and extracting financial key parameters using regular expressions to generate the second financial data;

[0025] If the recognition confidence is the low recognition confidence, the financial text data is marked as voice to be confirmed, and manual review is triggered.

[0026] Furthermore, the API unit includes:

[0027] Adopting circular paging to obtain all the bank API data;

[0028] Setting a filtering timestamp to pull the newly added bank API data according to a fixed timestamp and synchronize the bank API data of the previous timestamp;

[0029] Convert the bank API data into a unified time zone, a unified exchange rate, and clean transaction counterparties to generate the third financial data;

[0030] The second exception mechanism is used to check the third financial data. If the hash value of the third financial data is inconsistent, the reading is automatically retried; if the number of retries exceeds M times, the third financial data is isolated and manual review is triggered.

[0031] Furthermore, the dynamic routing includes: if the data packet includes a financial transaction field, it is classified as financial data; if the data packet includes a user operation, it is classified as operation instruction data; a financial cross-reference relationship is established for the data packet to describe the correlation between the data packets, and is classified as business relationship data.

[0032] Furthermore, the cross-checking rules include:

[0033] Amount matching verification rules are used to calculate the transaction amount deviation value using the amount tolerance formula and perform matching;

[0034] Time window verification rules are used to calculate the transaction time deviation value using the time window formula and mark whether the time is compliant;

[0035] The transaction subject name similarity verification rule is used to calculate the text similarity of the transaction subject name using the natural language processing model and decide whether to update the supplier map.

[0036] A method for automatic extraction and cross-checking of financial data, comprising:

[0037] Access financial image data, audio data, and banking API data;

[0038] The financial image data is subjected to seal processing, text recognition and semantic correction, and the first financial data is extracted through the first abnormal mechanism; the audio data is converted into financial text data through the Whisper model, and the second financial data is obtained according to the confidence level classification; the bank API data is subjected to paging query technology to obtain transaction flow, the transaction flow is subjected to standardization processing, and the third financial data is extracted using the second abnormal mechanism;

[0039] Performing format verification and priority marking on the first financial data, the second financial data, and the third financial data; classifying the data packets through dynamic routing, and adopting four-way diversion to store them in different types of financial databases;

[0040] Verify and process differences in the financial database based on cross-checking rules to generate a financial analysis report;

[0041] A data snapshot and rollback mechanism is used to monitor the financial analysis report.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1. The present invention integrates a financial data extraction method that integrates OCR recognition, voice analysis, and bank API data acquisition. By performing seal processing, text recognition, and semantic correction on image data, the recognition accuracy of bill information is improved, and misrecognition caused by seal occlusion or handwriting is effectively reduced. For voice data, financial operations can be accurately analyzed through intent recognition and key parameter extraction, reducing financial operation errors caused by voice misrecognition. In addition, standardized conversion and anomaly detection are performed on bank data to ensure that the transaction flow data format is unified and the consistency of data is improved.

[0044] 2. The present invention proposes a financial data classification method based on dynamic routing and four-way diversion, which performs format verification and priority marking on the acquired financial data packets to ensure that data with higher priority can be classified faster. According to the data content characteristics, the data is automatically classified into financial transaction data, operation instruction data, business relationship data and abnormal data, and stored in different types of databases. This method enhances the integration capability of heterogeneous data and realizes the precise association of cross-modal data, thereby improving the matching rate and query efficiency of financial data.

[0045] 3. The present invention adopts a cross-database data verification and exception handling mechanism based on cross-checking rules. Through amount matching, time window verification and transaction subject similarity analysis, the financial data is accurately checked, which can effectively identify amount errors, cross-period transactions and subject name matching anomalies, thereby improving the accuracy of financial data. Combined with the automatic correction and manual review mechanism, the exception handling process is optimized, and the data can be automatically adjusted within the acceptable error range, and a manual review channel is provided for high-risk anomalies. This mechanism effectively reduces reconciliation errors and improves the accuracy and compliance of financial data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of the structure of a system for automatic extraction and cross-checking of financial data provided by the present invention;

[0047] Figure 2 A flowchart of a method for automatic extraction and cross-checking of financial data provided by the present invention;

[0048] Figure 3 A schematic flow chart of the four-way flow splitting storage method provided by the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] See also Figures 1 to 3 The present invention provides a system and method for automatic extraction and cross-checking of financial data, and the technical solution is as follows:

[0051] Embodiment 1:

[0052] In the process of enterprise financial data management, accurate and efficient extraction and verification of financial data is the key to improving the level of financial management. However, the existing technology still has many deficiencies in the automatic extraction, data integration, anomaly detection and intelligent cross-checking of financial data, and it is difficult to meet the needs of enterprises for high concurrency, real-time and intelligent financial processing. Therefore, in response to these problems, the present invention provides a system for automatic extraction and cross-checking of financial data, aiming to improve the accuracy and compliance of financial processing, such as Figure 1 As shown, including:

[0053] Financial data acquisition module, used to acquire financial image data, audio data and bank API data;

[0054] Specifically, first, use a mobile phone or scanner to take pictures of financial documents such as invoices, receipts, and contracts, and record metadata such as the image format, resolution, and upload time for subsequent processing. Next, obtain audio data, such as "Record this invoice No. INV20240201 as office expenses" or "Adjust the category of invoice No. INV20240201 to marketing promotion." Voice data is stored in formats such as WAV and MP3, and metadata such as audio duration, sampling rate, and upload time are recorded. Finally, the OAuth2.0 client credential mode is used to obtain an access token to ensure secure access to the bank's open platform API. When the token expires, the system will automatically refresh to avoid data acquisition interruptions caused by authentication failure. In addition, the system also uses a breakpoint resume mechanism to record the last synchronization timestamp so that it can continue from the breakpoint when the network is interrupted to ensure the integrity and continuity of data synchronization.

[0055] The financial data processing module includes: an OCR (Optical Character Recognition) unit, which is used to perform seal processing, text recognition and semantic correction on the financial image data, and extract first financial data through a first exception mechanism; a voice unit, which is used to convert the audio data into financial text data through a Whisper model, and obtain second financial data according to confidence level classification; an API (Application Programming Interface) unit, which is used to obtain transaction flow from the bank API data using a paging query technology, standardize the transaction flow, and extract third financial data using a second exception mechanism.

[0056] Furthermore, the implementation process of the OCR unit includes:

[0057] Performing geometric correction, denoising and color enhancement on the financial image data to generate pre-processed financial image data;

[0058] Using the PP-OCRv3 detection model to locate the text area of ​​the pre-processed financial image data, and performing seal interference detection, if seal interference exists, repairing the text area;

[0059] Perform optical character recognition on the text area using the PP-OCRv3 recognition model to generate original financial text;

[0060] Performing context verification and logic authentication on the original financial text, and using a Transformer model to perform semantic correction, and outputting the first financial data;

[0061] The first exception mechanism is used to check the first financial data. If the text confidence is lower than the first threshold and any one of the identification failures of the same field occurs N times in a row, a manual review is triggered.

[0062] Specifically, taking the invoice as an example, we first use OpenCV to accurately identify the boundary of the bill, and then perform a perspective transformation to make it appear in a standard rectangular format. Then, we perform a non-local mean filter on the image, which effectively reduces background noise and improves text contrast. For bills taken under low-light conditions, we use gamma correction technology to significantly improve the overall brightness of the image. Finally, we further optimize the text clarity through CLAHE (contrast-limited adaptive histogram equalization), making the text sharper during the OCR recognition process, thereby obtaining pre-processed financial image data, which effectively improves the accuracy of OCR recognition.

[0063] Then, the PP-OCRv3 detection model is used to accurately segment the text area based on the DBNet algorithm to extract possible character areas. The DBNet algorithm achieves accurate positioning of the text area by calculating the text area probability map. At the same time, the YOLOv5 target detection model is used to identify the red seal area to determine whether it causes occlusion of the text. If the seal affects key fields (such as amount and invoice number), the seal repair mechanism is triggered. OpenCV's cv2.inpaint can be used for image repair to fill the background color of the seal-occluded area. If the repaired area cannot restore the complete text, the affected fields are manually reviewed and marked.

[0064] The PP-OCRv3 recognition model is used to perform character recognition in the text area, and context verification and logic authentication are performed. For example, the format of the amount field is ensured to be correct. At the same time, the Transformer model based on the BERT algorithm is used to correct the text, such as correcting the misrecognized "5Z80.00" to "5280.00".

[0065] The first exception handling mechanism is set as follows: the confidence of the PP-OCRv3 recognition model and the confidence of the BERT algorithm are weighted and summed to obtain the comprehensive confidence. The first threshold is set to 0.7. If the comprehensive confidence is lower than the threshold, or the same field fails to be recognized three times in a row, the data is marked as pending review and stored in the manual review queue. If the manual review passes, the data enters the financial data sorting module; if the error rate is found to be high, the training data of the PP-OCRv3 recognition model or the Transformer model is adjusted.

[0066] By performing Gamma correction and CLAHE preprocessing on financial image data, the image quality is effectively optimized and the accuracy of OCR recognition is improved. On this basis, using PP-OCRv3 and YOLOv5 technology, the text area can be accurately located, and the seal interference can be identified and repaired, further improving the accuracy of text detection. At the same time, context correction is performed on the text area to reduce misrecognition and enhance the ability to detect anomalies. In addition, by setting the comprehensive confidence, the misjudgment caused by relying solely on the confidence of text detection is avoided, thereby improving the accuracy of automatic recognition of financial data.

[0067] Furthermore, the implementation process of the speech unit includes:

[0068] Convert the audio data into financial text data using the Whisper speech recognition model and generate recognition confidence;

[0069] Classifying the recognition confidence, including high recognition confidence and low recognition confidence;

[0070] If the recognition confidence is the high recognition confidence, performing intent recognition on the financial text data using RasaNLU, and extracting financial key parameters through regular expressions to generate the second financial data;

[0071] If the recognition confidence is the low recognition confidence, the financial text data is marked as voice to be confirmed, and manual review is triggered.

[0072] Specifically, the Whisper speech recognition model is used to transcribe the audio data, and the output includes financial text data and recognition confidence. For speech with a heavy accent or high ambient noise, noise reduction processing (such as spectral subtraction) can be used to improve the recognition accuracy of Whisper. The recognition confidence threshold is set to 0.85 to distinguish between high recognition confidence and low recognition confidence data. For financial text data with high recognition confidence, RasaNLU is first used for semantic analysis to identify user intent (such as adjusting the invoice category), and then regular expressions and lexical analysis techniques are used to extract key information such as invoice number, amount and category to ensure data accuracy and structured storage. For financial text data with low recognition confidence, it is stored in the manual review queue. If the recognition confidence is between 0.7 and 0.85, a voice questioning mechanism can be introduced, such as "Please confirm whether to record the invoice as office expenses?", asking the staff to confirm the key information to improve data quality. If the recognition confidence is less than 0.7, it is directly handed over to manual review. After the review is passed, the system records the corrections and classifies them as financial text data with high recognition confidence.

[0073] The accuracy of speech-to-text conversion is improved through the Whisper speech recognition model, speech noise reduction, and confidence grading. Combined with RasaNLU and regular matching technology, the automated processing capabilities of financial operations are further improved. For low-confidence speech data, the system automatically triggers a manual review mechanism to ensure the accuracy and reliability of the data. Finally, the system outputs structured financial data, providing strong support for the cross-checking of financial data.

[0074] Furthermore, the API unit includes:

[0075] Adopting circular paging to obtain all the bank API data;

[0076] Setting a filtering timestamp to pull the newly added bank API data according to a fixed timestamp and synchronize the bank API data of the previous timestamp;

[0077] Convert the bank API data into a unified time zone, a unified exchange rate, and clean transaction counterparties to generate the third financial data;

[0078] The second exception mechanism is used to check the third financial data. If the hash value of the third financial data is inconsistent, the reading is automatically retried; if the number of retries exceeds M times, the third financial data is isolated and manual review is triggered.

[0079] Specifically, the bank API returns a maximum of 500 transaction records each time, so it is necessary to obtain them through a paging mechanism until all data is synchronized. Use the "last_page" flag to determine whether it is the last page to ensure the integrity of financial data and avoid omissions. Bank transaction records will add new data every day. To prevent duplicate pulls, set a filtering timestamp to only obtain new data after the last synchronization. For example, pull new data at 4 a.m. every day and perform incremental synchronization operations.

[0080] Bank APIs may return data with different time zones, currencies, and non-standard counterparty names. These data need to be standardized to ensure consistency with local financial data. Unify the time zone to the local time zone, and unify different currencies to the local currency through exchange rate conversion. For non-standard counterparty names returned by bank APIs, they need to be cleaned and standardized, such as converting "ICBC Beijing Branch" to "ICBC Beijing Branch" to obtain standardized third-party financial data.

[0081] When initially storing the third financial data, the hash value of each transaction data is calculated and stored. In each subsequent incremental synchronization, the hash value of the third financial data is recalculated as a benchmark for future data verification. The newly calculated hash value is compared with the existing hash value in the database:

[0082] If the data is repeated (the hash value is the same), skip it.

[0083] If the data changes (the hash value is different), it may mean that the data has been tampered with or transmitted incorrectly. In this case, automatic retry or manual review is required.

[0084] If it is new data (the hash value does not exist in the database), it is stored normally.

[0085] When the hash values ​​are different, if the newly calculated hash value does not match the original hash value, the system will automatically re-pull the data M times (for example, 3 times). If it still fails after M retries, the manual review mechanism will be triggered to ensure the accuracy and completeness of the data.

[0086] The integrity of financial data is ensured through paging and "last_page" flag judgment. Through timestamp filtering, only new data is pulled to avoid duplicate pulling, which improves synchronization efficiency. The consistency of financial data is ensured by unified processing of time zones, currencies and counterparty names. Through hash value calculation and comparison, data changes or errors are identified, and automatic retries or manual reviews are triggered to ensure data integrity and reliability. Automatic retries and manual reviews are used to ensure the accuracy of financial data extraction.

[0087] The financial data sorting module is used to perform format verification and priority marking on the data packets of the financial data processing module; classify the data packets through dynamic routing, and adopt four-way diversion to store them in different types of financial databases.

[0088] In the financial data processing module, three types of financial data are aggregated into the data lake to form data packets. The data packets are checked for integrity to ensure that they contain necessary fields (such as "trans_id", "date" and "amount"). If they are missing, they are marked as abnormal data. Data packets are prioritized according to transaction time: data packets with transaction time ≤ 24 hours are set to high priority; data packets with transaction time > 24 hours are set to normal priority; abnormal data are marked as low priority and require manual review.

[0089] Furthermore, the dynamic routing includes: if the data packet includes a financial transaction field, it is classified as financial data; if the data packet includes a user operation, it is classified as operation instruction data; a financial cross-reference relationship is established for the data packet to describe the correlation between the data packets, and is classified as business relationship data.

[0090] Among them, the first financial data (such as invoices) is classified as financial data because it has financial fields such as transaction amount, invoice number and supplier. The second financial data (such as voice operation) is classified as operation instruction data because it contains financial operation instructions entered by the user. The third financial data (such as bank API transaction flow) is also classified as financial data and is used for matching and reconciliation. The financial cross-check relationship is used to describe the correlation between invoices, bank flow and voice operation, and is classified as business relationship data, which is convenient for subsequent reconciliation verification, abnormal analysis and data tracing. As shown in Table 1, it is specifically expressed as follows: All financial data adopt a dynamic transaction ID mechanism, first assigning transaction IDs to the first financial data and the second financial data, and the third financial data enters the waiting matching pool, waiting for reconciliation matching. In the later cross-check processing, the transaction ID is matched according to the "amount, time and transaction subject", and the database is updated.

[0091] Table 1 Example of financial cross-checking relationship

[0092] Data Types Unique reference field Transaction ID allocation strategy First financial data (such as invoices) Invoice number, supplier, invoicing time A transaction ID is generated when an invoice enters the system and is initially matched based on "vendor and date". Secondary financial data (such as voice operation) Voice content Parse the voice operation to extract the invoice number involved and bind it to the corresponding transaction ID. Third, financial data (such as bank API transaction flow) Bank transaction serial number, transaction time, transaction counterparty The transaction ID is not assigned yet, and it enters the pending matching pool to wait for later cross-checking processing.

[0093] Through the dynamic routing mechanism, financial data is automatically classified and stored, and uniformly managed through transaction ID, which enables automatic anomaly identification and analysis, and improves the accuracy and intelligence of financial data management.

[0094] A financial data cross-checking module is used to verify and process the differences of the financial database based on the cross-checking rules and generate a financial analysis report;

[0095] Furthermore, the cross-checking rules include:

[0096] The amount matching verification rule is used to calculate the transaction amount deviation value using the amount tolerance formula and perform matching. It can be expressed as:

[0097] ;

[0098] ;

[0099] in, is the transaction amount deviation value, is the amount of the first type of financial data (such as invoice amount), The amount of the second type of financial data (such as bank statement amount), is the amount tolerance coefficient (the default setting is 0.5%). , the match is successful, otherwise, it is marked as an abnormal transaction. Then, it enters the difference processing.

[0100] As shown in Table 2, the tolerance is the tolerable amount deviation calculated based on the amount tolerance coefficient. The first invoice amount does not match the bank flow amount.

[0101] Table 2 Amount matching verification example

[0102] Invoice amount Bank transaction amount Transaction amount deviation Tolerance Match ¥5280.00 ¥5250.00 ¥30.00 ¥26.40 no ¥1200.00 ¥1198.00 ¥2.00 ¥6.00 yes

[0103] The time window verification rule is used to calculate the transaction time deviation value using the time window formula and mark whether the time is compliant. It can be expressed as:

[0104] ;

[0105] ;

[0106] in, is the transaction time deviation value, The first type of financial data time (such as invoice time), For the second type of financial data (such as bank flow time), The allowed time window (the default setting is 3 days). If the transaction time deviation value is within the time window, the match is successful. Otherwise, it is marked as an abnormal transaction and enters the difference processing.

[0107] The transaction subject name similarity verification rule is used to convert text into vectors using the natural language processing model, and use cosine similarity to calculate the similarity of two vectors, that is, to calculate the text similarity of the transaction subject name and decide whether to update the supplier map. The natural language processing model is represented as:

[0108] ;

[0109] in, is the similarity score, ranging from [0,1]. The higher the score, the higher the match. is the cosine similarity function, For the first category suppliers, For the second-category counterparty. For example, if the first-category supplier is "Shanghai XXXX Technology Co., Ltd.", the vector is "[0.2, 0.5, 0.8, ...]", and the second-category counterparty is "Shanghai XXXX Technology", the vector is "[0.1, 0.4, 0.7, ...]". At the same time, multi-layer filtering mechanisms such as industry classification comparison and enterprise credit code verification are introduced to prevent similar but different company names from being mismatched, ensuring the accuracy and reliability of supplier map updates.

[0110] Through the financial data verification method based on cross-checking rules, amount matching, time window verification and transaction subject similarity analysis are adopted to accurately complete cross-checking. At the same time, combined with the exception handling mechanism and automatic report generation, the level of intelligent financial management is improved to ensure the integrity, accuracy and traceability of financial data.

[0111] The financial data display module is used to monitor the financial analysis report using a data snapshot and rollback mechanism.

[0112] The present invention integrates OCR recognition, voice analysis and bank API data acquisition to achieve accurate extraction and standardized processing of multi-source financial data. Then, through dynamic routing and four-way diversion, data storage and classification management are optimized. Then, based on the cross-checking rules, financial data is automatically verified and anomaly detected, and combined with data snapshots and rollback mechanisms, the accuracy, traceability and intelligence of financial data analysis and management are ensured, thereby improving the efficiency and compliance of corporate financial management.

[0113] Embodiment 2:

[0114] A technology company is engaged in the manufacture of electronic components. It needs to process a large number of supplier purchase invoices, bank payment records and manual financial records every month. Due to the involvement of multiple suppliers and complex approval processes, the company faces challenges in financial data processing, such as scattered data sources, difficulty in identifying anomalies and insufficient data processing accuracy. In order to solve these problems, Figure 2 As shown, a method for automatic extraction and cross-checking of financial data is used, including:

[0115] Access financial image data, audio data, and banking API data;

[0116] The financial image data is subjected to seal processing, text recognition and semantic correction, and the first financial data is extracted through the first abnormal mechanism; the audio data is converted into financial text data through the Whisper model, and the second financial data is obtained according to the confidence level classification; the bank API data is subjected to paging query technology to obtain transaction flow, the transaction flow is subjected to standardization processing, and the third financial data is extracted using the second abnormal mechanism;

[0117] The first financial data, the second financial data and the third financial data are format checked and prioritized; the data packets are classified through dynamic routing, and stored in different types of financial databases using four-way diversion.

[0118] in, Figure 3 The flowchart of the four-way flow splitting storage method provided by the present invention is shown in FIG. Figure 3 As shown in the figure, data packets are obtained from the data lake. After classifying the data packets through dynamic routing, financial data is stored in MySQL in batches, operation instruction data is stored in MongoDB in real-time writing, business relationship data is stored in Neo4j in transactional Cypher statements, abnormal data is continuously streamed and stored in the Kafka dead letter queue, and the cause of the error is recorded.

[0119] The financial database is verified and differences are processed based on the cross-checking rules to generate a financial analysis report.

[0120] Among them, the difference processing includes automatic correction strategy and manual review processing. The automatic correction strategy automatically adjusts the error within the allowable range. For example, the invoice amount is ¥5290, and the bank statement is ¥5291, which is marked as an amount mismatch. However, if the amount matching verification rule is passed, the automatic correction strategy will be automatically enabled, and the invoice amount will be automatically adjusted based on the bank statement. The manual review processing is a serious anomaly and needs to be submitted for manual review. For example, the invoice amount is ¥5290, and the bank statement is ¥5320. The amount matching verification rule is not passed and it needs to be submitted to the financial staff for review.

[0121] A data snapshot and rollback mechanism is used to monitor the financial analysis report.

[0122] Specifically, in this embodiment, during the financial data cross-checking process, such as automatic correction of amounts and manual review of transactions, data changes may occur. However, if the adjustment is wrong, the original data cannot be restored, and the mismatched transactions are not discovered in time, which may affect the accuracy of financial analysis. Before the financial data is changed (such as automatic adjustment of amounts, update of transaction IDs, and execution of exception processing), the data is backed up first. The storage method and data snapshot are shown in Tables 3 and 4. Data from different sources are stored in snapshots, and the original values ​​and adjusted values ​​are recorded to ensure that they can be traced back during the reconciliation process. Snapshots are generated only before data changes to avoid storage redundancy.

[0123] Table 3 Data snapshot storage method example

[0124] Snapshot Type Data Source Invoice data snapshot MySQL (invoices table) Bank statement snapshot MySQL (bank_transactions table) Abnormal reconciliation record snapshot MongoDB (mismatch_logs collection)

[0125] Table 4 Data snapshot example

[0126] Transaction ID Invoice Number Original amount Adjusted amount Creation time TX202X0201 INV202X0201001 ¥5280.00 ¥5300.00 202X-02-02 10:00

[0127] Table 5 Financial analysis report example

[0128] Analytical Indicators Numeric Total transactions 1200 Matching success rate 98.3% Number of abnormal rollbacks 5 Main abnormality type Amount error 50%, time limit 25%, subject mismatch 25%

[0129] Table 6 Effect comparison

[0130] Comparison Dimensions A technology company's original method The present invention Data extraction accuracy 86.5% 96.8% (OCR, semantic correction, seal repair, etc.) Multi-source data integration capabilities Only supports OCR recognition Supports OCR, voice analysis and automatic extraction of bank API data Cross-check processing accuracy 94.2% (fixed rule matching) 97.4% (intelligent cross-checking rules and NLP matching transaction entities, etc.) Abnormal data rollback capability No rollback mechanism Using data snapshots and automatic rollback

[0131] Next, before the financial analysis report is generated, the integrity of the financial data is monitored. First, the current data is compared with the snapshot data, and the number of difference lines between the current data and the snapshot data is calculated. If the number of difference lines exceeds 5%, the data is marked as abnormal. Then, the changes in key fields are checked, and an early warning is triggered when the changes in transaction amount, transaction time, and transaction subject exceed 2%. When the change exceeds 5%, the generation of the financial analysis report is suspended, and the rollback process is entered to re-execute the financial cross-checking rules, thus ensuring the accuracy of financial data processing. The generation results of the financial analysis report are shown in Table 5, which shows the number of transactions, matching rate, rollback number, and exception type, which can support enterprises to optimize financial management strategies.

[0132] Compared with the original financial data extraction and cross-checking processing method of a certain technology company, as shown in Table 6, the OCR data extraction accuracy and cross-checking processing accuracy of the present invention are improved, and it supports multi-source data integration capabilities and abnormal data rollback capabilities, which can effectively improve the accuracy, traceability and intelligence of financial data analysis and management.

[0133] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A system for automatic extraction and cross-checking of financial data, characterized in that: include: Financial data acquisition module, used to acquire financial image data, audio data and bank API data; The financial data processing module includes: an OCR unit, which is used to perform seal processing, text recognition and semantic correction on the financial image data, and extract the first financial data through the first abnormal mechanism; the implementation process of the OCR unit includes: Performing geometric correction, denoising and color enhancement on the financial image data to generate pre-processed financial image data; Using the PP-OCRv3 detection model to locate the text area of ​​the pre-processed financial image data, and performing seal interference detection, and if seal interference exists, repairing the text area; Perform optical character recognition on the text area using the PP-OCRv3 recognition model to generate original financial text; Performing context verification and logic authentication on the original financial text, and using a Transformer model to perform semantic correction, and outputting the first financial data; Using the first exception mechanism to check the first financial data, if the text confidence is lower than the first threshold and if the recognition of the same field fails N times in a row, triggering manual review; A speech unit is used to convert the audio data into financial text data through a Whisper model, and obtain second financial data according to confidence level. The implementation process of the speech unit includes: Convert the audio data into financial text data using the Whisper speech recognition model and generate recognition confidence; Classifying the recognition confidence, including high recognition confidence and low recognition confidence; If the recognition confidence is the high recognition confidence, performing intent recognition on the financial text data using RasaNLU, and extracting financial key parameters using regular expressions to generate the second financial data; If the recognition confidence is the low recognition confidence, marking the financial text data as voice to be confirmed and triggering manual review; The API unit is used to obtain transaction flow using paging query technology for the bank API data, standardize the transaction flow, and extract third financial data using the second exception mechanism; the API unit includes: Adopting circular paging to obtain all the bank API data; Setting a filtering timestamp to pull the newly added bank API data according to a fixed timestamp and synchronize the bank API data of the previous timestamp; Convert the bank API data into a unified time zone, a unified exchange rate, and clean transaction counterparties to generate the third financial data; Using the second exception mechanism to check the third financial data, if the hash value of the third financial data is inconsistent, automatically retry reading; if the number of retries exceeds M times, isolate the third financial data and trigger manual review; A financial data sorting module is used to perform format verification and priority marking on the data packets of the financial data processing module; classify the data packets through dynamic routing, and adopt four-way diversion to store them in different types of financial databases; A financial data cross-checking module is used to verify and process the differences of the financial database based on the cross-checking rules and generate a financial analysis report; The financial data display module is used to monitor the financial analysis report using a data snapshot and rollback mechanism.

2. A system for automatic extraction and cross-checking of financial data according to claim 1, characterized in that: The dynamic routing includes: if the data packet includes a financial transaction field, it is classified as financial data; if the data packet includes a user operation, it is classified as operation instruction data; a financial cross-reference relationship is established for the data packet to describe the correlation between the data packets, and is classified as business relationship data.

3. A system for automatic extraction and cross-checking of financial data according to claim 1, characterized in that: The cross-checking rules include: Amount matching verification rules are used to calculate the transaction amount deviation value using the amount tolerance formula and perform matching; Time window verification rules are used to calculate the transaction time deviation value using the time window formula and mark whether the time is compliant; The transaction subject name similarity verification rule is used to calculate the text similarity of the transaction subject name using the natural language processing model and decide whether to update the supplier map.

4. A method for automatic extraction and cross-checking of financial data, characterized in that: include: Access financial image data, audio data, and banking API data; The financial image data is subjected to seal processing, text recognition and semantic correction, and the first financial data is extracted through the first abnormal mechanism; the implementation process includes: Performing geometric correction, denoising and color enhancement on the financial image data to generate pre-processed financial image data; Using the PP-OCRv3 detection model to locate the text area of ​​the pre-processed financial image data, and performing seal interference detection, and if seal interference exists, repairing the text area; Perform optical character recognition on the text area using the PP-OCRv3 recognition model to generate original financial text; Performing context verification and logic authentication on the original financial text, and using a Transformer model to perform semantic correction, and outputting the first financial data; Using the first exception mechanism to check the first financial data, if the text confidence is lower than the first threshold and if the recognition of the same field fails N times in a row, triggering manual review; The audio data is converted into financial text data through the Whisper model, and second financial data is obtained according to confidence level classification. The implementation process includes: Convert the audio data into financial text data using the Whisper speech recognition model and generate recognition confidence; Classifying the recognition confidence, including high recognition confidence and low recognition confidence; If the recognition confidence is the high recognition confidence, performing intent recognition on the financial text data using RasaNLU, and extracting financial key parameters using regular expressions to generate the second financial data; If the recognition confidence is the low recognition confidence, marking the financial text data as voice to be confirmed and triggering manual review; For the bank API data, a paging query technology is used to obtain transaction flow, the transaction flow is standardized, and the third financial data is extracted using a second exception mechanism, including: Adopting circular paging to obtain all the bank API data; Setting a filtering timestamp to pull the newly added bank API data according to a fixed timestamp and synchronize the bank API data of the previous timestamp; Convert the bank API data into a unified time zone, a unified exchange rate, and clean transaction counterparties to generate the third financial data; Using the second exception mechanism to check the third financial data, if the hash value of the third financial data is inconsistent, automatically retry reading; if the number of retries exceeds M times, isolate the third financial data and trigger manual review; Performing format verification and priority marking on the first financial data, the second financial data, and the third financial data; classifying data packets through dynamic routing, and adopting four-way diversion to store them in different types of financial databases; Verify and process differences in the financial database based on cross-checking rules to generate a financial analysis report; A data snapshot and rollback mechanism is used to monitor the financial analysis report.

Images