A subsidy auditing method, system and related device
The subsidy review system, which deeply integrates RPA and AI, solves the problems of low efficiency and low accuracy of traditional manual review, and achieves efficient and accurate automated review, reducing labor costs and ensuring the traceability of issues.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIONPAY MERCHANT SERVICES CO LTD
- Filing Date
- 2026-04-14
- Publication Date
- 2026-07-10
AI Technical Summary
Traditional manual review methods are inefficient and inaccurate, leading to a backlog of work orders, a poor user experience, and problems such as recognition errors and non-standard records.
The subsidy review system, which deeply integrates RPA and AI, achieves automated review through terminal initial inspection, server management, RPA client review, and AI server recognition. It accurately identifies document information and completes complex logical judgments, balancing efficiency and accuracy. Furthermore, it ensures traceability through mechanisms that automatically correct minor errors and return cases with major issues.
It achieves efficient and accurate automated review, reduces labor costs, improves review efficiency and accuracy, ensures traceability of issues, and reduces upgrade risks through non-intrusive deployment.
Smart Images
Figure CN122367384A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a subsidy review method, system and related equipment. Background Technology
[0002] With the implementation of the trade-in policy and the surge in subsidy applications, finance departments at all levels and banks are facing enormous review pressure. Against this backdrop, the traditional manual review model has revealed several shortcomings.
[0003] First, the review process is cumbersome and inefficient. The car trade-in program involves reviewing a large amount of information, and each work order takes a long time to process. Insufficient daily processing capacity leads to a severe backlog of work orders, with initial reviews averaging several weeks, significantly impacting the user's subsidy application experience. Second, manual review suffers from a high rate of recognition errors and logical reasoning mistakes. Due to the dense information on vehicle registration certificates and invoices, manual review is prone to errors due to fatigue or negligence. Furthermore, when logically verifying the transaction date on the used car registration certificate and the compliance of the invoice, manual review often confuses the rules, thus reducing the accuracy of the review.
[0004] Furthermore, the tracing and recording of issues are not standardized. During manual review, issue fields are mostly recorded manually, which easily leads to omissions and unclear records, making it difficult to quickly locate the source of the problem during subsequent reviews and audits. To cope with peak work orders, the review team needs to temporarily expand its staff, which not only increases labor costs but also requires a significant investment of time in policy and rule training. At the same time, high staff turnover also leads to inconsistent review standards, further exacerbating the complexity and instability of the review work.
[0005] In summary, given the various shortcomings of the traditional manual review model, there is an urgent need for an efficient, accurate, and traceable automated review solution to meet the growing business demands. Summary of the Invention
[0006] In view of this, embodiments of the present invention provide a subsidy review method, system and related equipment to solve the problems of low efficiency and low accuracy of manual review.
[0007] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:
[0008] The first aspect of this invention discloses a subsidy review system, the system comprising: a terminal, a server, an RPA client, and an AI server;
[0009] The server is communicatively connected to the terminal, the RPA client, and the AI server, respectively.
[0010] The RPA client and the AI server are connected in communication.
[0011] The terminal is used to receive information to be reviewed and perform a preliminary inspection; the information to be reviewed that passes the preliminary inspection is encrypted and packaged into a work order to be reviewed and sent to the server.
[0012] The server is used to send the work order to be reviewed to the RPA client for review, and update the status of the work order to be reviewed based on the received review results;
[0013] The RPA client is used to obtain the work order to be reviewed, call the AI server to identify the information to be reviewed in the work order to be reviewed, and obtain the identification result; verify the identification result and generate the verification result; and write the verification result into the review record table to obtain the review result.
[0014] Preferably, the terminal includes a subsidy application module and a preliminary inspection module;
[0015] The subsidy application module is used to display prompts for basic information, document information, and document images, and to receive basic information, document information, and document images uploaded by users, and generate information to be reviewed.
[0016] The initial inspection module is used to perform an initial inspection on the information to be reviewed; when the initial inspection is passed, the information to be reviewed that has passed the initial inspection is encrypted and encapsulated into a standard format, the work order to be reviewed is generated and sent to the server.
[0017] Preferably, the initial inspection module is further used for:
[0018] When the initial inspection fails, a pop-up message is generated and output based on the pending review information of the failed initial inspection.
[0019] Preferably, the server includes a work order management module, an access control module, and a result receiving and status update module;
[0020] The work order management module is used to receive work orders to be reviewed sent by the terminal and classify and store them; for each work order to be reviewed, a work order number is generated.
[0021] The permission control module is used to assign review permissions to the RPA client and assign the work order number to the RPA client according to the review permissions, so that the RPA client can obtain the work order to be reviewed according to the work order number;
[0022] The result receiving and status update module is used to receive the review results sent by the RPA client and update the status of the work order to be reviewed.
[0023] Preferably, the result receiving and status update module is further configured to:
[0024] The audit process is recorded based on the audit results, and the audit process includes the audit time and the audit results.
[0025] Preferably, the RPA client includes an acquisition module, a verification module, and a submission module;
[0026] The acquisition module is used to acquire the work order to be reviewed and send the information to be reviewed in the work order to the AI server for recognition through the AI server interface;
[0027] The verification module is used to receive the recognition results sent by the AI server; verify the basic information, document information, invoices and logical rules in the recognition results, and generate verification results.
[0028] The submission module is used to write the verification result into the audit record table to obtain the audit result, and send the audit result to the server.
[0029] Preferably, the AI server includes a capacity scheduling module and a recognition result processing module;
[0030] The capacity scheduling module is used to receive the information to be reviewed in the work order to be reviewed; and according to the type of the information to be reviewed, call the corresponding model to identify and judge the information to be reviewed, and obtain the target information.
[0031] The recognition result processing module is used to structure and organize the target information to obtain the recognition result.
[0032] The second aspect of this invention discloses a subsidy review method, applied to the subsidy review system disclosed in the first aspect of this invention, the method comprising:
[0033] The terminal receives the information to be reviewed and performs a preliminary inspection; the information that passes the preliminary inspection is encrypted and packaged into a work order to be reviewed and sent to the server;
[0034] The server sends the work order to be reviewed to the RPA client for review, and updates the status of the work order to be reviewed based on the received review results;
[0035] The RPA client obtains the work order to be reviewed, calls the AI server to identify the information to be reviewed in the work order, and obtains the recognition result.
[0036] The identification result is verified to generate an inspection result; the inspection result is written into the audit record table to obtain the audit result.
[0037] The third aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the subsidy review method disclosed in the second aspect of the present invention.
[0038] The fourth aspect of this invention discloses an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the subsidy review method disclosed in the second aspect of this invention.
[0039] This invention provides a subsidy review method, system, and related equipment based on the above embodiments. The system includes a terminal, a server, an RPA client, and an AI server; the server is communicatively connected to the terminal, the RPA client, and the AI server, and the RPA client is communicatively connected to the AI server. The terminal performs a preliminary check on the information to be reviewed, encrypts and encapsulates it into a work order, and sends it to the server; the server sends the work order to the RPA client for review and updates the work order status according to the review results; the RPA client calls the AI server to identify the information, verify it, and write it into the review record table to generate the review results. This invention deeply integrates RPA and AI models to achieve automated review, accurately identify document information, and complete complex logical judgments, balancing efficiency and accuracy. It ensures traceability through a classification mechanism that automatically corrects minor errors and returns major issues, and it connects to existing platforms in a non-intrusive manner, enabling low-risk and rapid deployment. Attached Figure Description
[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0041] Figure 1 A schematic diagram of the framework of a subsidy review system provided in an embodiment of the present invention;
[0042] Figure 2 This is another schematic diagram of a subsidy review system provided in an embodiment of the present invention;
[0043] Figure 3 A flowchart of a subsidy review method provided in an embodiment of the present invention;
[0044] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0046] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0047] As the background technology shows, the manual review process is cumbersome and inefficient, resulting in a backlog of work orders and a poor user experience. At the same time, the high recognition error, non-standard records, and high staff turnover make the review work complicated and costly.
[0048] Therefore, embodiments of the present invention provide a subsidy review method, system, and related equipment. The system includes: a terminal, a server, an RPA client, and an AI server; wherein the server is communicatively connected to the terminal, the RPA client, and the AI server, and the RPA client is communicatively connected to the AI server. The terminal performs a preliminary inspection of the information to be reviewed, encrypts and encapsulates it into a work order, and sends it to the server; the server sends the work order to be reviewed to the RPA client for review, and updates the work order status according to the review results; the RPA client calls the AI server to identify the information, verify it, and write it into the review record table to generate the review results. The present invention deeply integrates RPA and AI models to achieve automated review, accurately identify document information, and complete complex logical judgments, balancing efficiency and accuracy. It ensures traceability through a classification mechanism that automatically corrects minor errors and returns major issues, and integrates with existing platforms in a non-intrusive manner to achieve low-risk and rapid deployment.
[0049] See Figure 1 The diagram shows a framework schematic of a subsidy review system provided by an embodiment of the present invention.
[0050] The system includes: terminal 1, server 2, RPA client 3, and AI server 4.
[0051] Specifically, server 2 is connected to terminal 1, RPA client 3 and AI server 4 respectively; RPA client 3 and AI server 4 are connected to each other.
[0052] Terminal 1 is used to receive information to be reviewed and perform initial checks; it encrypts and encapsulates the information that passes the initial check into a work order to be reviewed and sends it to server 2. Terminal 1 is, for example, a client application.
[0053] Server 2 is used to send work orders to be reviewed to RPA client 3 for review, and update the status of work orders to be reviewed based on the received review results. Server 2 is, for example, the review platform server.
[0054] RPA client 3 is used to obtain work orders to be reviewed, call AI server 4 to identify the information to be reviewed in the work orders, obtain the recognition result; verify the recognition result, generate the verification result; and write the verification result into the review record table to obtain the review result. AI server 4 includes, for example, an Optical Character Recognition (OCR) server (customized or non-customized) and a multimodal large model.
[0055] It should be noted that AI Server 4 employs two types of OCR models, which are adapted differently based on the information features of the recognition scene:
[0056] General-purpose OCR model (i.e., non-customized OCR model): Suitable for non-standardized image scenarios where the location and form of elements are not fixed. The model returns all text content in the image and its coordinate information for RPA client 3 to perform content matching and compliance verification.
[0057] Customized OCR model: Specifically optimized for car trade-in verification scenarios, focusing on accurate recognition of financial, payment, and vehicle verification-related documents (such as vehicle registration certificates, driver's licenses, bank cards, and various invoices). The model returns specified fields in JSON key-value pair format, enabling direct extraction of target information.
[0058] Model training and optimization: The customized model is trained on the de-identified payment dataset to ensure the recognition adaptability in financial and certificate scenarios; at the same time, the RPA review process forms a data feedback loop, using incorrect recognition cases found by manual review as negative examples, and continuously optimizing the model by combining it with correct samples to continuously improve recognition accuracy.
[0059] Understandably, the multimodal visual model used in AI Server 4 has been specifically fine-tuned for the car trade-in subsidy review scenario. The core capabilities and optimization mechanisms are as follows:
[0060] It is suitable for recognizing unstructured visual information and making complex logical judgments. It supports operations such as original / scanned document identification, invoice and official seal detection, fuzzy document text repair, and registration certificate transfer record location. It can also perform associated logical verification based on visual information, such as comparing the time difference between the registration date and the sale date, and matching vehicle types.
[0061] Based on the scenario of vehicle trade-in verification, a dedicated dataset of vehicle documents was used to fine-tune the basic model. The dataset includes original / scanned copies of vehicle registration certificates, driver's licenses, and invoices for new / used vehicles, as well as invoices with / without official seals and samples with varying degrees of ambiguity. All samples have been anonymized to ensure that the model is adapted to business requirements in terms of visual recognition and logical judgment.
[0062] A linkage feedback mechanism is established with the RPA review process. The RPA client 3 synchronizes the identification errors and logical misjudgment cases found in the review (such as misjudgment of original / scanned documents, failure to check official seals, and date comparison errors) to the AI server 4. The AI server 4 combines the above error samples with the corresponding correct samples to form an optimized dataset, and regularly adjusts the model parameters to continuously improve the recognition accuracy and logical judgment precision.
[0063] It should be noted that RPA client 3, as the automated execution carrier, forms a two-way collaborative and data-closed-loop linkage mechanism with OCR model and multimodal visual large model: RPA initiates recognition or judgment requests to AI server 4 according to the review process, and performs operations such as data comparison, rule verification, and result submission after receiving the structured return results; at the same time, it synchronizes AI model error cases found in the review to AI server 4, providing real business samples for model training and optimization, and achieving a two-way improvement in execution efficiency and recognition accuracy.
[0064] The following is combined with Figure 2 The diagram shows another framework of a subsidy review system, with detailed descriptions of terminal 1, server 2, RPA client 3, and AI server 4.
[0065] It should be noted that Terminal 1 includes a subsidy application module and a preliminary inspection module.
[0066] Combination Figure 2 As shown, the subsidy application module and the preliminary inspection module are used to realize functions such as data collection, data storage, data display and data synchronization.
[0067] Specifically, the subsidy application module displays prompts for basic information, document information, and document images, and receives the basic information, document information, and document images uploaded by the user, generating information to be reviewed.
[0068] For example, in the subsidy application module, the interface automatically loads the special fields for the car trade-in program and displays them in categories of "basic information" and "document information." Basic information includes, but is not limited to, the applicant's name, ID number, bank card number, bank name, and contact information; document information is divided into "new car" and "used car," including fields such as the new and used vehicle registration certificate numbers, license plate number, vehicle identification number (VIN), engine number, vehicle registration certificate validity period, and invoice codes and numbers for both new and used cars.
[0069] Users can fill in the relevant information as prompted and upload corresponding document images, including but not limited to the front and back of their ID card, the front and back of their bank card, new vehicle registration certificate and driver's license, used vehicle registration certificate and driver's license, and new and used vehicle invoices (original or scanned copies). During the upload process, Terminal 1 will automatically prompt for image shooting specifications to ensure that the images meet the requirements, such as frontal alignment, no obstructions, and sufficient clarity.
[0070] Specifically, the initial inspection module is used to perform an initial inspection of the information to be reviewed; when the initial inspection is passed, the information to be reviewed that has passed the initial inspection is encrypted and packaged into a standard format, a work order to be reviewed is generated and sent to server 2.
[0071] The initial inspection module is also used to generate and output a pop-up message based on the pending review information when the initial inspection fails.
[0072] For example, during the initial review of information to be reviewed, checks include whether required fields are complete, whether the ID number is compliant in length, whether the bank card number format is correct, whether the image clarity meets OCR recognition requirements, and whether the image is obscured. If the initial review fails, a pop-up window will display specific correction suggestions, such as "ID number is less than 18 digits, please check and re-enter" or "Used car invoice image is blurry, please retake and upload."
[0073] Understandably, when the initial inspection is passed, the information to be reviewed (i.e., the information filled in and the image of the certificate) that passed the initial inspection is encrypted and packaged into a standard format, a work order to be reviewed is generated and sent to the relay server of the terminal via HTTPS protocol, and then the relay server synchronizes it to server 2.
[0074] In some specific embodiments, after server 2 receives the work order to be reviewed, the relay server receives the "work order submitted successfully + unique work order number" prompt returned by server 2. At the same time, the relay server generates a work order progress query QR code and returns the unique work order number and work order progress query QR code to terminal 1.
[0075] Understandably, users can use the "Review Progress" entry on terminal 1 to enter the work order number or scan the work order progress query QR code to view the review results (e.g., approved / returned / pending review) and detailed reasons (e.g., "the registration date of the used car registration certificate is less than 1 year different from the sale date, the review is returned") synchronized with server 2 in real time.
[0076] It should be noted that Server 2 includes a work order management module, an access control module, and a result receiving and status update module.
[0077] Combination Figure 2 As shown, the work order management module, access control module, and result receiving and status update module are used to implement functions such as data receiving, data display, and data auditing.
[0078] Specifically, the work order management module is used to receive work orders to be reviewed from the terminal and classify and store them; for each work order to be reviewed, a work order number is generated.
[0079] For example, work orders awaiting review are stored in categories such as "pending review, under review, and reviewed (approved / returned)". For each work order awaiting review, an internal work order number is generated on the platform and associated with the UnionPay Business work order number.
[0080] Specifically, the access control module is used to assign review permissions to RPA clients and assign work order numbers to RPA clients based on the review permissions, so that RPA clients can obtain work orders to be reviewed based on the work order numbers.
[0081] It should be noted that the access control module provides RPA client 3 with a dedicated login account, allowing it to retrieve work orders, view data, and submit review results. At the same time, it restricts data modification and deletion permissions to ensure data security.
[0082] Specifically, the result receiving and status update module is used to receive the review results sent by the RPA client and update the status of the work orders to be reviewed.
[0083] It should be noted that the status of a work order pending review includes pending review, approved, pending modification, returned, and no processing.
[0084] The result receiving and status update module is also used to: record the audit trajectory based on the audit results, the audit trajectory including the audit time and audit results.
[0085] It should be noted that RPA client 3 includes an acquisition module, a verification module, and a submission module.
[0086] Combination Figure 2As shown, the acquisition module, verification module, and submission module are used to implement system login, order capture, AI recognition, logic verification and result processing, and Excel entry.
[0087] Specifically, the acquisition module is used to acquire work orders to be reviewed and send the information to be reviewed in the work orders to the AI server for recognition through the AI server interface.
[0088] Understandably, when the reviewer starts the RPA client 3, they automatically log in to the server 2 using a preset account and password, and are redirected to the "Pending Review Work Orders" list page, which is sorted and displayed by work order submission time and supports filtering by license plate number and work order number.
[0089] It should be noted that the acquisition module operates based on pre-exported ledger information (i.e., work order information awaiting review, mainly including order numbers, usually exported by staff from the system). RPA client 3 uses the order numbers to search the system for orders awaiting review, then retrieves the work order details one by one (including structured data filled in by the user, supplementary information from the merchant, and images of various documents), and generates an Excel record table. The table header includes information such as work order number, license plate number, applicant's name, review items, AI recognition result, comparison result, problem description, correct result, review conclusion, and processing time.
[0090] Understandably, RPA client 3 calls the AI server interface to upload various document images to AI server 4 (OCR server (custom or non-custom) and multimodal large model) and initiates a recognition request.
[0091] For example, identification requirements can be specified according to document type: ID card: identify name, ID number, and validity period; bank card: identify card number and issuing bank; new and used vehicle registration certificate / vehicle license: identify license plate number, VIN code, engine number, registration date, owner's name, and transfer record (pages 3 and 4 of the used vehicle registration certificate); invoice: identify invoice code, number, invoice date, names of buyer and seller, vehicle type, transaction amount, name of used car market, and official seal.
[0092] Specifically, the verification module receives the recognition results sent by the AI server; it verifies the basic information, document information, invoices, and logical rules in the recognition results, and generates verification results.
[0093] In practical applications, data comparison and logical verification are performed in three categories of verification items:
[0094] Category 1: Basic Information Comparison and Verification. The verification module compares the AI-recognized ID card number and bank card number with the fields entered by the user on the platform. If they match completely, it is marked as "Comparison Passed". If there are 1-2 character errors (e.g., the last 'X' in the ID card number is incorrectly capitalized or a digit is missing from the bank card number), it is marked as "Comparison Abnormal", the correct result is recorded in Excel, the corresponding field on the platform is corrected, and the correction trace is retained.
[0095] The second category: Vehicle document information comparison and verification. The verification module compares the license plate number, VIN code, engine number, and registration date recognized by AI with the fields entered on the platform, and verifies the validity period of the vehicle registration certificate. If there is a single error in key fields such as the license plate number or VIN code (such as incorrect capitalization or confused numbers), it is marked as "comparison anomaly," the correct result is recorded in Excel, and the platform field is corrected; if the fields are completely inconsistent or the vehicle registration certificate is expired, it is marked as "verification failed."
[0096] The third category: Invoice and logical rule verification. The core logical judgment is completed by an AI multimodal large model, and the RPA receives the judgment result and performs verification.
[0097] Used car registration certificate logic: If the owner of the transfer certificate is the applicant, then check whether the difference between the registration date and the sale date is ≥1 year; if the owner is not the applicant, then check the transfer information on pages 3 and 4 to confirm that the difference between the time the applicant acquired the vehicle and the time of sale is ≥1 year, and both times are clearly recorded. If the condition is met, it passes; otherwise, it is marked as "logic verification failed".
[0098] Used car invoice compliance: AI determines whether the invoice is an original or a scanned copy and whether it bears an official seal. RPA verifies the consistency between the buyer's and seller's names and the platform information, ensuring the vehicle type is a small or micro passenger car. It also checks the consistency between the used car market name and the system's recognition result. If the market name is empty, it verifies the consistency between the buyer's name on the invoice and the used car market name recognized by the system; if any one of these conditions is not met, it is marked as "Invoice verification failed."
[0099] New vehicle invoice verification: If the AI recognition result matches the fields entered on the platform, and the invoice is within the subsidy period, it will be marked as passed.
[0100] Specifically, the submission module is used to write the verification results into the audit record table to obtain the audit results, and then send the audit results to the server.
[0101] Understandably, based on the verification results, the submission module will perform the following categorization processing:
[0102] All validation items passed: RPA automatically clicked "Approved" and recorded "Approved" in Excel, and then submitted the information to Server 2.
[0103] If there are only 1-2 minor errors in basic information or vehicle document information (which have been corrected), and there are no logical validation or invoice compliance issues: RPA clicks "Approved", records the problem description and correct result in detail in Excel, and submits the information to Server 2 for reviewers to refer to.
[0104] If there are logical validation failures (e.g., the registration date and the sale date are less than 1 year apart) or invoice compliance issues (e.g., no official seal, not the original, vehicle type does not match): RPA will automatically select "return", record the specific return reason in Excel, fill in the return description on server 2, and submit it to server 2.
[0105] After the review is completed, RPA will automatically save the Excel record to the specified path, with the file named "Subsidy Review Issue Record_YYYYMMDD.xlsx".
[0106] It's worth noting that if unexpected situations occur during the review process, such as the reviewer prematurely closing the RPA client, the system will skip the already reviewed parts and directly continue reviewing the unreviewed orders upon the next re-execution, based on the log output (whether the RPA execution was successful). Finally, after all orders have been reviewed, the reviewer can open the log to view the review results, determine which orders passed, which failed and the reasons, and what precautions should be taken for the passed orders.
[0107] AI Server 4 includes a capacity scheduling module and a recognition result processing module.
[0108] Combination Figure 2 As shown, the capability scheduling module and the recognition result processing module are used to implement functions such as receiving requests, server calculation, obtaining results, and model optimization.
[0109] Specifically, the capacity scheduling module is used to receive the information to be reviewed from the work orders to be reviewed; based on the type of the information to be reviewed, it calls the corresponding model to identify and judge the information to be reviewed, and obtain the target information.
[0110] Understandably, the capability scheduling module receives the identification request from RPA client 3 and allocates the corresponding AI model based on the type of information to be reviewed.
[0111] OCR Text Recognition Model: This model is used to process standard character information in ID cards, bank cards, invoices, and registration certificates / vehicle licenses to ensure accurate extraction.
[0112] Multimodal Vision Large Model: This model is used to handle complex scene recognition, including fuzzy registration certificate text repair, used car invoice official seal detection, and registration certificate transfer record area localization. Simultaneously, the model performs logical judgments, such as comparing the registration date with the sale date, vehicle type recognition, and determining whether the original invoice is a scanned copy.
[0113] It should be noted that the logical judgments of the capacity scheduling module can be divided into two categories:
[0114] OCR type: This type only returns data, such as standard information like ID card numbers, bank card numbers, and invoice numbers. RPA logically compares the results returned by OCR with the content entered by the user on the webpage (e.g., consistency and compliance), and marks the results accordingly.
[0115] Multimodal large model class: This model can directly perform logical judgments, such as asking whether the invoice is the original. The AI can directly return the result of "yes" or "no", and RPA will then perform further processing.
[0116] Specifically, the recognition result processing module is used to structure and organize the target information to obtain the recognition result.
[0117] The last four steps are understandable. The recognition result processing module is responsible for structuring the recognized information and returning it in the format specified by RPA. The content includes the recognized fields, confidence levels, and screenshot annotations. Fields with a confidence level below 90% will be marked as "suspected error" to remind RPA to perform key verification.
[0118] The specific analysis is as follows: The OCR model automatically provides a confidence score when outputting results. This confidence score is based on the softmax maximum probability output of deep learning, requiring no manual intervention. For results from multimodal large models, the confidence score needs to be calculated and output simultaneously in the prompt words.
[0119] For example, when inquiring about whether an invoice bears an official seal, the large model should be asked to return the confidence level of that judgment. In this case, the confidence level is automatically assessed by the large model based on the complexity of the problem; the more certain the result, the higher the confidence level. In summary, the recognition result processing module ensures that all recognition information is returned in a structured format and appropriately labels the confidence levels for subsequent verification and processing.
[0120] In some specific embodiments, the RPA client 3 uploads AI recognition error cases (such as fuzzy character recognition errors or missing official seals) found during the review process to the AI server 4. The AI server 4 combines the error cases with the correct information and regularly optimizes the OCR model and multimodal large model parameters to improve recognition accuracy and logical judgment precision.
[0121] In this embodiment of the invention, through the deep integration of RPA automated execution and AI models, full-time automated review is achieved in the car trade-in subsidy review scenario, replacing repetitive manual operations. The AI server integrates OCR and multimodal large model to accurately identify document information and complete complex logical judgments, improving review efficiency while ensuring accuracy. The classification and processing mechanism of automatically correcting minor field errors and recording them to Excel, and directly returning major issues, balances review efficiency and review convenience, ensuring that issues are traceable. RPA non-intrusively connects to the existing subsidy platform without modifying the original system architecture, achieving low-risk and rapid deployment and effectively reducing upgrade costs. At the same time, an AI model iteration mechanism is established to continuously optimize the model using recognition error cases fed back by RPA, achieving dynamic improvement in recognition accuracy. The terminal refines the collection fields and local verification rules according to the car scenario, reducing data errors from the source and lowering subsequent rework costs.
[0122] A subsidy review system provided in the above embodiments of the present invention is applied, see [link to relevant documentation]. Figure 3 The diagram illustrates a flowchart of a subsidy review method provided by an embodiment of the present invention. The method includes:
[0123] Step S301: The terminal receives the information to be reviewed and performs a preliminary inspection; the information to be reviewed that passes the preliminary inspection is encrypted and packaged into a work order to be reviewed and sent to the server.
[0124] Step S302: The server sends the work order to be reviewed to the RPA client for review, and updates the status of the work order to be reviewed based on the received review results.
[0125] Step S303: The RPA client obtains the work order to be reviewed, calls the AI server to identify the information to be reviewed in the work order, and obtains the recognition result.
[0126] Step S304: Verify the recognition result and generate the verification result; write the verification result into the audit record table to obtain the audit result.
[0127] It should be noted that the specific implementation process of steps S301 to S304 is detailed in [link to documentation]. Figure 1 and Figure 2 The content shown will not be repeated here.
[0128] In this embodiment of the invention, the solution deeply integrates RPA and AI models to achieve automated review at all times, accurately identify document information and complete complex logical judgments, balancing efficiency and accuracy; traceability is ensured through a classification mechanism that automatically corrects minor errors and returns major issues; RPA is non-intrusive and integrates with existing platforms for low-risk and rapid deployment; recognition accuracy is continuously improved by relying on the AI model iteration mechanism; and data quality is ensured from the source by refining the collection and verification rules at the terminal.
[0129] Another embodiment of this application provides an electronic device, such as... Figure 4 As shown, it includes: memory 401 and processor 402.
[0130] Among them, memory 401 is used to store computer programs.
[0131] The processor 402 is used to execute a computer program, which, when executed, is specifically used to implement a subsidy review method as provided in any of the above embodiments.
[0132] The electronic devices mentioned in this article can be servers, PCs, tablets, mobile phones, ECUs (Electronic Control Units), VCUs (Vehicle Control Units), MCUs (Micro Controller Units), HCUs (Hybrid Control Units), etc.
[0133] Another embodiment of this application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements a subsidy review method as provided in any of the above embodiments.
[0134] Computer-readable storage media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0135] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0136] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0137] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A subsidy verification system, characterized in that, The system includes: a terminal, a server, an RPA client, and an AI server; The server is communicatively connected to the terminal, the RPA client, and the AI server, respectively. The RPA client and the AI server are connected in communication. The terminal is used to receive information to be reviewed and perform a preliminary inspection; the information to be reviewed that passes the preliminary inspection is encrypted and packaged into a work order to be reviewed and sent to the server. The server is used to send the work order to be reviewed to the RPA client for review, and update the status of the work order to be reviewed based on the received review results; The RPA client is used to obtain the work order to be reviewed, call the AI server to identify the information to be reviewed in the work order to be reviewed, and obtain the identification result; verify the identification result and generate the verification result; and write the verification result into the review record table to obtain the review result.
2. The system according to claim 1, characterized in that, The terminal includes a subsidy application module and a preliminary inspection module; The subsidy application module is used to display prompts for basic information, document information, and document images, and to receive basic information, document information, and document images uploaded by users, and generate information to be reviewed. The initial inspection module is used to perform an initial inspection on the information to be reviewed; When the initial inspection is passed, the information to be reviewed is encrypted and packaged into a standard format, and the work order to be reviewed is generated and sent to the server.
3. The system according to claim 2, characterized in that, The initial inspection module is also used for: When the initial inspection fails, a pop-up message is generated and output based on the pending review information of the failed initial inspection.
4. The system according to claim 1, characterized in that, The server includes a work order management module, an access control module, and a result receiving and status update module. The work order management module is used to receive work orders to be reviewed sent by the terminal and classify and store them; for each work order to be reviewed, a work order number is generated. The permission control module is used to assign review permissions to the RPA client and assign the work order number to the RPA client according to the review permissions, so that the RPA client can obtain the work order to be reviewed according to the work order number; The result receiving and status update module is used to receive the review results sent by the RPA client and update the status of the work order to be reviewed.
5. The system according to claim 4, characterized in that, The result receiving and status update module is also used for: The audit process is recorded based on the audit results, and the audit process includes the audit time and the audit results.
6. The system according to claim 1, characterized in that, The RPA client includes an acquisition module, a verification module, and a submission module; The acquisition module is used to acquire the work order to be reviewed and send the information to be reviewed in the work order to the AI server for recognition through the AI server interface; The verification module is used to receive the recognition result sent by the AI server; The basic information, document information, invoices, and logical rules in the identification results are verified to generate verification results. The submission module is used to write the verification result into the audit record table to obtain the audit result, and send the audit result to the server.
7. The system according to claim 1, characterized in that, The AI server includes a capacity scheduling module and a recognition result processing module; The capacity scheduling module is used to receive the information to be reviewed in the work order to be reviewed; according to the type of the information to be reviewed, call the corresponding model to identify and judge the information to be reviewed, and obtain the target information; The recognition result processing module is used to structure and organize the target information to obtain the recognition result.
8. A subsidy verification method, characterized in that, Applied to the subsidy review system according to any one of claims 1 to 7, the method includes: The terminal receives the information to be reviewed and performs a preliminary inspection; the information that passes the preliminary inspection is encrypted and packaged into a work order to be reviewed and sent to the server; The server sends the work order to be reviewed to the RPA client for review, and updates the status of the work order to be reviewed based on the received review results; The RPA client obtains the work order to be reviewed, calls the AI server to identify the information to be reviewed in the work order, and obtains the recognition result. The identification result is verified to generate an inspection result; the inspection result is written into the audit record table to obtain the audit result.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the subsidy review method as described in claim 8.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the subsidy review method as described in claim 8.