A power grid contract order reimbursement management method and system

Through edge data platform and natural language processing technology, the problems of low approval efficiency and difficulty in data integration in power grid contract order reimbursement management are solved, automated approval and real-time data synchronization are realized, and the intelligence and compliance of power grid contract order management are improved. .

CN118735472BActive Publication Date: 2024-11-29湖南省湘电试验研究院有限公司 +3
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411232561.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-11-29
Estimated Expiration
2044-09-04

AI Technical Summary

Technical Problem

The traditional power grid contract order reimbursement management model has problems such as low approval efficiency, difficult data dispersion and difficulty in integrating, and difficulty in cross-system collaboration. It is impossible to achieve real-time data synchronization, automated approval processes, precise authority control and intelligent data analysis.

Method used

A method and system for reimbursement of power grid contract orders is proposed to collect and store data through edge data platform, use signature identification model to perform data verification, use decision tree and large language model to determine process templates, realize automated approval and data synchronization, and Data query is carried out through natural language processing technology.

Benefits of technology

Improve the operational efficiency and approval speed of power grid contract order reimbursement, improve management reliability and intelligence level, ensure the transparency and compliance of approval, reduce the burden on management personnel, and achieve seamless connection and Real-time data synchronization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118735472B_ABST
    Figure CN118735472B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for managing power grid contract order reimbursement, the method comprising: collecting the business data of contract order reimbursement to be approved and performing signature verification, if the verification fails, prompting the data items that do not meet the verification rules and re-collecting the business data of contract order reimbursement to be approved; if the verification passes, matching the process template and removing the repeated approval nodes; sending the approval task of the process template to the corresponding approver, recording the approval data and storing it in the edge data platform; writing the approved contract order reimbursement business data in the edge data platform to the target source end system, or obtaining the contract order reimbursement business data from other source end systems and storing it in the edge data platform; parsing the natural language query requirements and obtaining the query results from the contract order reimbursement business data of the edge data platform. The present invention not only improves the operational efficiency and approval speed of power grid contract order reimbursement, but also improves the reliability and intelligence level of the entire power grid contract order management.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to information management technology, and in particular to a power grid contract order reimbursement management method and system. Background Art

[0002] Grid contract order reimbursement refers to the process of reimbursing orders and expenses related to grid contracts. This process involves submitting relevant orders, invoices and expense vouchers. After going through the company's internal approval process, the finance department will review and issue reimbursements. This ensures that the company pays the relevant expenses in accordance with the contract and maintains compliance.

[0003] As the scale of operation of power grid enterprises expands and the complexity of their business increases, the traditional contract order reimbursement management model faces problems such as low approval efficiency, scattered data that is difficult to integrate, and difficulty in cross-system collaboration. The existing technology lacks an integrated management solution for the entire life cycle of power grid contract order reimbursement, and cannot achieve functions such as real-time data synchronization, automated approval process, precise authority control, and intelligent data analysis. Summary of the invention

[0004] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides a power grid contract order reimbursement management method and system, which not only improves the operational efficiency and approval speed of power grid contract order reimbursement, but also improves the reliability and intelligence level of the entire power grid contract order management.

[0005] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0006] A power grid contract order reimbursement management method comprises the following steps:

[0007] After the initiator starts the grid contract order reimbursement process, collect the contract order reimbursement business data to be approved;

[0008] Perform signature verification on the contract order reimbursement business data to be approved. If the contract order reimbursement business data to be approved fails to pass the signature verification, prompt the data items that do not meet the verification rules and re-collect the contract order reimbursement business data to be approved; if the contract order reimbursement business data to be approved passes the signature verification, store the contract order reimbursement business data to be approved in the edge data platform;

[0009] Obtain the contract order reimbursement business data to be approved from the edge data platform, determine the applicable process template according to the preset matching rules, and remove the repeated approval nodes of the process template, wherein the process template includes the corresponding approval level and approval personnel;

[0010] According to the process template, the approval tasks are sent to the corresponding level approvers, and the approval data of each approval stage is recorded and stored in the edge data platform;

[0011] Read the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source systems, and write it into the target source system, or obtain the contract order reimbursement business data that needs to be synchronized from other source systems and store it in the edge data platform;

[0012] Use preset management reports to manage the contract order reimbursement business data of the edge data platform;

[0013] Parse natural language query requirements and obtain query results from the contract order reimbursement business data of the edge data platform.

[0014] Furthermore, the contract order reimbursement business data to be approved includes structured and unstructured information. When the signature verification of the contract order reimbursement business data to be approved is performed, the following steps are included:

[0015] Convert the unstructured information in the contract order reimbursement business data to be approved into image data;

[0016] Input the image data into a trained signature recognition model to obtain a predicted signature border, each predicted signature border being associated with a confidence score and a probability distribution of a category;

[0017] Through the specified confidence threshold, high-confidence prediction results are screened out from the predicted signature border, and redundant detection results are eliminated from the high-confidence prediction results to obtain the final result; if the final result is not empty, the contract order reimbursement business data to be approved has passed the signature verification; if the final result is empty, the contract order reimbursement business data to be approved has not passed the signature verification.

[0018] Furthermore, before inputting the image data into the trained signature recognition model, the step of training the signature recognition model is also included, specifically including:

[0019] A specified image recognition model is selected, a detail capture layer is added before the output layer of the image recognition model, and the last pooling layer of the image recognition model is replaced with a convolution layer with a step size of 1 to obtain a signature recognition model, wherein the detail capture layer includes three convolution layers, each of which is followed by a ReLU activation function and a normalization layer, the first convolution layer uses a 1×1 filter, and the second and third convolution layers both use 3×3 filters;

[0020] Use the preset signature dataset to train the signature recognition model.

[0021] Further, determining the applicable process template according to the preset matching rule and removing the repeated approval nodes of the process template includes:

[0022] A decision tree algorithm is used to classify and calculate the factors specified in the contract order reimbursement business data to be approved according to dynamically configurable business rules to obtain a first process template;

[0023] Using the large language model, the first process template is analyzed according to its process steps, and the approval nodes of the duplicate approvers are removed to form a second process template without duplicate approvers. Using the large language model to obtain the second process template specifically includes:

[0024] Prompt word part A is constructed according to the process nodes and approvers in the first process template, and prompt word part B is constructed according to the instruction to remove duplicate approval nodes. Prompt word part A is used to describe the original complex activities, and prompt word part B is used to describe the optimized results. Prompt word part A and prompt word part B are fused and used as the input of the large language model.

[0025] Furthermore, constructing the prompt word part A and constructing the prompt word part B both include: selecting a specified number of carriers, each carrier represents a sense organ, concretizing each sense organ description through a specific carrier scenario, and different sense organs form prompt words corresponding to the carriers.

[0026] Furthermore, when reading the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source end systems and writing it to the target source end system, the following steps are included:

[0027] Obtain synchronization requests from other source systems and insert the corresponding contract order reimbursement business data into the bottom of the data queue in a predetermined order;

[0028] According to the first-in-first-out principle, the preset source system write interface is called to write the contract order reimbursement business data in the data queue into the corresponding source system in sequence.

[0029] Further, when calling the preset source end system writing interface to write the contract order reimbursement business data in the data queue into the corresponding source end system in sequence, it includes:

[0030] Obtain the response code of the source end system corresponding to the current contract order reimbursement business data in the data queue, and determine whether the writing is successful according to the response code. If the writing is successful, write the next contract order reimbursement business data in the data queue into the corresponding source end system;

[0031] If the writing fails, wait for a preset time and then call the source system writing interface again to write the current contract order reimbursement business data in the data queue into the corresponding source system;

[0032] If the number of unsuccessful write attempts exceeds the preset value, the next contract order reimbursement business data in the data queue will be written to the corresponding source system and a prompt message will be generated.

[0033] Furthermore, the source system writing interface is constructed using RPA technology, including one or more of a fusion platform writing interface and an image 2.0 system writing interface.

[0034] Furthermore, when parsing the query requirements of natural language and obtaining the query results from the contract order reimbursement business data of the edge data platform, the following steps are included:

[0035] All contract order reimbursement business data in the edge data platform are sorted, a redundant table structure definition of contract order reimbursement is constructed, and the redundant table structure definition is converted into a natural language expression form, and a prompt word C part is constructed according to the redundant table structure definition in the natural language expression form, and the prompt word C part is used to describe complex data;

[0036] Acquire a data query requirement in natural language, construct a prompt word D part according to the data query requirement in natural language, wherein the prompt word D part is used to describe a precise query problem, and concatenate the prompt word C part and the prompt word D part to generate a standard prompt word;

[0037] Input standard prompt words into the language model for logical decision-making and reasoning, thereby generating standard SQL query statements that conform to the natural language structure;

[0038] Query the edge data platform according to the standard SQL query statement to obtain the query result data.

[0039] The present invention also proposes a power grid contract order reimbursement management system, including a user end, a service end, a cross-system integration module and an edge data platform, wherein:

[0040] The user terminal comprises:

[0041] The user data collection module is used to collect the business data of the contract order reimbursement to be approved after the initiator starts the power grid contract order reimbursement process. If the business data of the contract order reimbursement to be approved passes the signature verification, the business data of the contract order reimbursement to be approved is stored in the edge data platform;

[0042] The process approval module is used to obtain the contract order reimbursement business data to be approved through the edge data platform, determine the applicable process template according to the preset matching rules, and remove the duplicate approval nodes of the process template. The process template includes the corresponding approval level and approval personnel; it is also used to record the approval data of each approval stage and store it in the edge data platform;

[0043] The statistical analysis module is used to manage the contract order reimbursement business data of the edge data platform using preset management reports; it is also used to parse natural language query requirements and obtain query results from the contract order reimbursement business data of the edge data platform;

[0044] The server includes:

[0045] The automatic approval auxiliary module is used to perform signature verification on the contract order reimbursement business data to be approved. If the contract order reimbursement business data to be approved fails to pass the signature verification, the data items that do not meet the verification rules are pointed out;

[0046] The message reminder module is used to push status changes and process progress information to the initiator and approver according to the process template, form approval tasks and send them to the approvers at the corresponding level;

[0047] The cross-system integration module is used to read the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source systems, and write it into the target source system; or obtain the contract order reimbursement business data that needs to be synchronized from other source systems and store it in the edge data platform.

[0048] Compared with the prior art, the advantages of the present invention are:

[0049] 1. Based on the internal control system of the power grid and multi-level classifier technology, the applicable process template is determined according to the actual reimbursement business data. Each process template contains the approval level and the approver corresponding to the template, so as to automatically generate accurate approval levels and intelligently assign the responsibilities of the approvers, avoiding errors and delays that may be caused by manual intervention. This feature greatly improves the standardization and execution efficiency of the approval process, making the approval work more transparent and fair, while also reducing the workload of managers and ensuring that contract order reimbursement matters are handled in a timely and reasonable manner;

[0050] 2. Using advanced natural language processing technology, it can understand and analyze the statistical analysis requirements in natural language proposed by users, and quickly generate customized query results. It simplifies the user's query operation and reduces the requirements for professional knowledge, so that all types of users can easily and quickly obtain the required report data, providing strong support for decision-making;

[0051] 3. Intelligent detection of signature information in unstructured information through image recognition algorithms effectively prevents approval problems caused by missing signatures. This function not only automatically advances the complete approval process, but also quickly provides feedback to the initiator when problems are found, ensuring the authenticity and integrity of the documents and significantly improving the overall compliance of contract order reimbursement business;

[0052] 4. It realizes seamless connection and real-time data synchronization of different source systems. By eliminating information islands, the tedious process of manual operation and data migration is greatly reduced, and the speed and accuracy of data flow are significantly improved, thereby enhancing the overall effectiveness and operating efficiency of power grid contract order reimbursement management, laying a solid foundation for the company's refined management and decision-making support. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0054] Figure 2 is an overall system block diagram of an embodiment of the present invention;

[0055] Figure 3 A diagram showing the steps of initiating contract order reimbursement by the system according to an embodiment of the present invention;

[0056] Figure 4 A diagram of the steps for the system of an embodiment of the present invention to obtain query results for natural language query requirements. DETAILED DESCRIPTION

[0057] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0058] Embodiment 1:

[0059] In order to solve the problems of approval level determination, repeated filling, cross-system integration, information islands, and custom query difficulties in the contract order reimbursement process of traditional power grid enterprises, this embodiment proposes a power grid contract order reimbursement management method, which is applied to the scenario of multiple source systems. By introducing functions such as automated approval assistance and intelligent question and answer query, the contract order management efficiency and data utilization value are greatly improved.

[0060] In this embodiment, the source end system includes at least a user end, a server end, an edge data platform, and a cross-system integration module, wherein:

[0061] The cross-system integration module is mainly responsible for cross-system data transmission operations, ensuring seamless data docking and synchronization between various source systems. This module reads the data queues in the edge data platform that have been approved and need to be synchronized to other source systems, and writes data to the target source system by managing the horizontal domain boundaries of the information zone, or obtains data from other source systems and stores it on the edge data platform.

[0062] The edge data platform serves as a data hub and a core data storage, aggregation and processing unit. It is responsible for the deep integration of various database information resources that the system relies on, and collects and manages in real time all business-related data submitted by the initiator through the user end, the approval data submitted by the approver, and the log data generated during the operation of other modules, while providing data services to the functions of each module.

[0063] like Figure 1 As shown, the method of this embodiment includes the following steps:

[0064] After the initiator starts the grid contract order reimbursement process, the user side collects the contract order reimbursement business data to be approved;

[0065] The server performs signature verification on the contract order reimbursement business data to be approved;

[0066] If the contract order reimbursement business data to be approved fails to pass the signature verification, the user end prompts the data items that do not meet the verification rules and re-collects the contract order reimbursement business data to be approved; if the contract order reimbursement business data to be approved passes the signature verification, the user end stores the contract order reimbursement business data to be approved in the edge data platform;

[0067] The user end obtains the contract order reimbursement business data to be approved from the edge data platform, determines the applicable process template according to the preset matching rules, and removes the repeated approval nodes of the process template, wherein the process template includes the corresponding approval level and approval personnel;

[0068] The server sends the approval tasks to the corresponding level approvers according to the process template, and the user side records the approval data of each approval stage and stores it in the edge data platform;

[0069] The cross-system integration module reads the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source systems, and writes it into the target source system, or obtains the contract order reimbursement business data that needs to be synchronized from other source systems and stores it in the edge data platform;

[0070] The user end uses preset management reports to manage the contract order reimbursement business data of the edge data platform. When a natural language query requirement is obtained, the natural language query requirement is parsed and the query result is obtained from the contract order reimbursement business data of the edge data platform.

[0071] Through the above steps, this embodiment realizes automation and intelligence in all links throughout the reimbursement process, from data input and verification to automated management of approval process, and then to integration with other systems and data analysis. Each link realizes efficient data processing and circulation through the synergy of carefully designed modules, which not only improves operational efficiency and approval speed, but also enhances the reliability and intelligence level of the entire power grid contract order management.

[0072] Each step is described in detail below.

[0073] In this embodiment, the contract order reimbursement business data to be approved includes structured and unstructured information. When the user terminal collects the contract order reimbursement business data to be approved, it specifically collects the required structured and unstructured information through a preset selection, entry and upload mechanism.

[0074] In this embodiment, the selection, input and upload mechanism includes: the initiator selects the associated annual budget code or name and the associated contract code or name data, enters the reimbursement reason, invoice amount, payment amount and whether data synchronization is required, and uploads the invoice and project process data with signature and seal;

[0075] The annual budget and related contract data selected by the initiator are obtained from the edge data platform and displayed, such as the annual budget amount, account code, control level, competent department, person in charge, contract amount, etc. At the same time, the remaining allowable reimbursement amount of the related contract is obtained through calculation and compared with the invoice amount and payment amount entered by the initiator, and the confirmation of the initiator is awaited.

[0076] When the initiator performs a confirmation operation, such as clicking the Submit function button, the server adopts the pre-configured data verification rules and uses the image recognition algorithm to perform signature verification on the contract order reimbursement business data to be approved, including the following steps:

[0077] Convert the unstructured information in the contract order reimbursement business data to be approved into image data. Specifically, convert the unstructured information into base64 encoding and communicate with the signature recognition interface. The signature recognition interface receives the base64 encoded data and restores it into PNG format image data.

[0078] Input the image data into a trained signature recognition model to obtain a predicted signature border, each predicted signature border being associated with a confidence score and a probability distribution of a category;

[0079] Through the specified confidence threshold, high-confidence prediction results are screened out from the predicted signature borders, and the non-maximum suppression (NMS) algorithm is used to eliminate redundant detection results for the high-confidence prediction results to obtain the final result. The final result can be used to locate the actual coordinate position information of all signatures in the image.

[0080] If the final result is not empty, that is, the actual coordinate position information of the signature exists in the final result, indicating that the document has been signed, then the contract order reimbursement business data to be approved has passed the signature verification and the approval process can be advanced to the next node; if the final result is empty, that is, there is no actual coordinate position information of the signature, indicating that the signature is missing, then the contract order reimbursement business data to be approved has not passed the signature verification, and the user end can generate a prompt: "The system has not detected the signature, please upload the stamped version of the file", and automatically return the process to the initiator of the contract order reimbursement business.

[0081] In this embodiment, the signature recognition model is obtained by modifying the basic image recognition model. An additional detail capture layer is introduced to increase the level of feature capture for the basic image recognition model. The detail capture layer can extract new features of the input image, so that the basic image recognition model can obtain more details and improve the ability to capture small target signature features. The specific steps of training the signature recognition model are as follows:

[0082] Step 1: Adjust the model architecture and add an additional 3 convolutional layers as detail capture layers near the input layer of the basic image recognition model. Each convolutional layer is followed by an activation function (ReLU) and a normalization layer (BatchNormalization). The first layer uses a 1×1 filter, and the second and third layers use 3×3 filters. Through this combined stacking, the model's ability to capture more details at high-resolution layers is enhanced, and the model's ability to perceive the original features of the signature is improved.

[0083] Step 2: Adjust the model architecture. Replace the last pooling layer of the basic image recognition model with a convolutional layer with a step size of 1. This allows feature learning to be performed while reducing the resolution. This allows the image recognition model to maintain a high-resolution feature map and retain more spatial detail information, thereby improving the accuracy of signature (small object) detection.

[0084] Step 3: After making the above adjustments, use the manually annotated signature data set to train the signature recognition model to obtain a trained signature recognition model.

[0085] In this embodiment, the basic image recognition model uses the YOLOV3 model. The modified YOLOV3 model is used as the signature recognition model to receive the restored PNG image as input and output a series of predicted signature bounding boxes. Each predicted signature bounding box is associated with a confidence score and a probability distribution of the category. The concept of anchor boxes is introduced in the bounding box prediction process. The signature recognition model predicts the offset (tx, ty, tw, th) of each anchor box relative to its coordinates, which is converted into the final predicted box coordinates through the following formula:

[0086] b_x =σ(tx) + c_x;

[0087] b_y =σ(ty) + c_y;

[0088] b_w =p_w × e^(tw);

[0089] b_h = p_h × e^(th);

[0090] Among them, tx, ty, tw, th are the offsets of each anchor box relative to its coordinates, σ() is a sigmoid function (a continuously differentiable nonlinear function), which compresses the input value to between (0,1), σ(tx) and σ(ty) represent the standardized output of the horizontal and vertical coordinate offsets of the center point of the prediction box relative to the center point of the anchor box, respectively; (c_x, c_y) are the original coordinates of the center point of the anchor box, that is, the coordinates obtained by adding the upper left corner point (x, y) of the anchor box to half the width and height of the anchor box; (b_x, b_y) are the coordinates of the center point of the prediction box, which are obtained by applying the sigmoid function to tx and ty and adding the center coordinates of the anchor box c_x and c_y; (b_w, b_h) are the width and height of the prediction box, where p_w and p_h are the original width and height of the anchor box, and tw and th are the logarithmic scaling factors of the predicted width and height relative to the width and height of the anchor box; e^(tw) and e^(th) The width and height offsets in the logarithmic space are respectively exponentially operated to enlarge or reduce the size of the anchor box to obtain the actual size of the predicted box.

[0091] After the contract order reimbursement business data to be approved submitted by the initiator passes the signature verification, the user end stores all the data submitted by the initiator in the edge data platform through the secure network channel of the management information zone, and then automatically generates the approval level and the responsibilities of the approvers according to the reimbursement business. Specifically, the contract order reimbursement business data to be approved is obtained through the edge data platform, and the approval node judgment and distribution processing unit are used to apply numerical calculation and matching rules to determine the applicable process template FLMB. Each process template FLMB contains the approval level and approver corresponding to the template; the process template FLMB adopts a multi-level classifier technology. The classifier comprehensively adopts a method combining a decision tree algorithm with an open source large language model, and uses the decision tree classification decision and large model decision decomposition capabilities. According to dynamically configurable business rules, the contract order type, contract amount, invoicing amount, and control level are classified and calculated to make decisions, so as to accurately locate the process template FLMB corresponding to the contract order reimbursement, while reducing repeated approval processes.

[0092] Therefore, in this embodiment, when determining the applicable process template according to the preset matching rules, a classification algorithm based on a decision tree is first used to classify and calculate the factors specified by the contract order reimbursement business data to be approved according to the dynamically configurable business rules, including the following steps:

[0093] Root node decision: Determine whether there is a "controlled amount" for the reimbursement item. If not, no approval is required; if so, proceed to the next step of judgment;

[0094] Sub-node decision: If the "Control Amount" is zero, approval is required; if it is non-zero, further confirmation is required as to whether there is a relevant contract;

[0095] Grandchild node decision: If there is no relevant contract, the judgment condition is whether the invoice amount or payment amount reaches or exceeds the control amount. If it is met, approval is required; if there is a relevant contract, it is divided into two branches:

[0096] Great-grandchild node 1: If there is a related contract and the payment type is not "framework contract", the judgment condition is that the invoice amount is greater than zero or the payment amount is greater than the approval limit of the responsible leader, and the contract amount must be greater than or equal to the control amount. If all conditions are met, approval is required;

[0097] Great-grandson node 2: If there is a related contract and the payment type is "framework contract", the judgment condition is that the invoice amount is greater than zero or the payment amount is greater than or equal to the approval limit of the unit head, and the invoice amount or payment amount must be greater than or equal to the control amount. If all conditions are met, approval is required.

[0098] After using the above decision tree algorithm to calculate the initial process nodes and approvers in the process template (which can be called the first process template for distinction) that matches the contract order reimbursement business data to be approved, use the large language model to analyze the process template according to its process steps, remove the approval nodes of duplicate approvers, and form a process template that does not contain duplicate approvers (which can be called the second process template for distinction). Specifically, in this embodiment, prompt engineering technology is used to construct prompt word part A according to the calculated initial process nodes and approvers, and at the same time, according to the instruction of removing duplicate approval nodes of "analyzing according to the accepted prompt word part A step by step, removing the approval nodes of duplicate approvers, and forming a process template FLMB without duplicate approvers", prompt word part B is constructed. Parts AB are jointly input to the large language model, and the large language model returns the result in json format to generate the process template FLMB.

[0099] Furthermore, when constructing prompt words, unlike the traditional prompt word construction method (i.e., the method of using text information as prompt words without processing or simply formatting), this embodiment proposes a prompt word construction method of "multi-sensory cognitive construction method". When constructing prompt words for different business scenarios, this method selects a variety of easy-to-understand carriers, each representing a sense, and concretizes each sensory description through a specific carrier scenario. Each sensory description is closely related to the business scenario context. After different senses form the prompt words corresponding to the carrier, each sensory prompt word is merged to form the final prompt word. This method enriches the context of the large model. By closely connecting each sensory description with the corresponding life scene, the model can more accurately grasp and understand the specific task background. These multi-sensory descriptions carefully crafted for this specific scene are merged to form the final prompt word. This method strengthens the insight of the language model in the entire prompt generation intermediate process, thereby achieving the goal of more accurately interpreting and responding to prompt words.

[0100] In the business scenario of the contract order reimbursement approval process, the construction of prompt word A and prompt word B can be described using the following easy-to-understand senses in life:

[0101] Sensory description of prompt word A (original approval level and approver information):

[0102] Visual carrier:

[0103] Prompt word A can be likened to a complex maze game diagram, where each node in the game represents an approval level, and the game player represents the approver. This maze game diagram presents the path that the player takes to travel through the nodes.

[0104] Auditory carrier:

[0105] Prompt word A can be compared to a symphony, in which the performance of each instrumentalist is a link in the approval process, and the music pieces they play correspond to their role in the approval process.

[0106] Sensory description of prompt word B (information after removing duplicate approvers):

[0107] Visual carrier:

[0108] Hint B can be imagined as a clear picture of a jigsaw puzzle after the unnecessary pieces have been removed. The completed version of this jigsaw puzzle is more concise, each piece is necessary, and there are no redundant parts, which represents the efficiency of the process.

[0109] Auditory carrier:

[0110] Cue B can be imagined as a melody stripped down to unnecessary notes, with no extra repetitive sections to disturb the integrity of the melody.

[0111] Fusion of prompt word A and prompt word B:

[0112] The fusion of prompt word A and prompt word B is equivalent to combining the complex and simplified concepts of different sensory descriptions with descriptive language. It can be constructed like this:

[0113] Fusion prompts: "At the beginning, we were faced with a complex maze map (the visual carrier of prompt word A), where each node and player together weaved the whole picture of the approval process. Each node has its own unique note (the auditory carrier of prompt word A), which together played a vibrant symphony. Over time, redundant paths in the maze were intelligently removed (the visual carrier of prompt word B), leaving a more efficient and intuitive path map, just like an optimized piece of music, retaining only the most exciting part (the auditory carrier of prompt word B). This combination of clarity and efficiency reduces time consumption and improves the accuracy of approval."

[0114] In summary, prompt word A is used to describe the original complex activity, while prompt word B is used to describe the optimized result. Finally, by combining the visual and auditory sensory carriers of the two, a comprehensive and vivid large model input prompt word is formed to generate an efficient approval process. The constructed prompt word can fully trigger the model to understand the complexity of the original approval process, and while removing duplicate approvers, optimize the entire approval process to make it more efficient and streamlined.

[0115] After the user matches the contract order reimbursement business data to be approved with a suitable process template, the server will push the status change and process progress information to the initiator and the approver according to the process in the process template, forming an approval to-do task and sending it to the approvers at the corresponding level. The approver enters the review interface through the to-do items, verifies the rationality of the contract order reimbursement business and other factors, and after selecting pass, reject or transfer, the process automatically flows to the next level of approval until the entire approval level is completed. In this process, the user will record the opinions and corresponding timestamps of each approval stage, and store the approval data on the edge data platform through the secure network channel of the management information zone. When the approval is completed, the corresponding contract order reimbursement business data is marked as approved contract order reimbursement business data.

[0116] After the approval is completed, the cross-system integration module reads the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source systems, and writes it to the target source system, including the following steps:

[0117] Obtain synchronization requests from other source systems and insert the corresponding contract order reimbursement business data into the bottom of the data queue in a predetermined order;

[0118] According to the first-in-first-out principle, the preset source system write interface is called to write the contract order reimbursement business data in the data queue into the corresponding source system in sequence, including:

[0119] Obtain the response code of the source end system corresponding to the current contract order reimbursement business data in the data queue, and determine whether the writing is successful according to the response code. If the writing is successful, write the next contract order reimbursement business data in the data queue into the corresponding source end system;

[0120] If the writing fails, wait for a preset time and then call the source system writing interface again to write the current contract order reimbursement business data in the data queue into the corresponding source system;

[0121] If the number of unsuccessful write attempts exceeds the preset value, the next contract order reimbursement business data in the data queue will be written to the corresponding source system and a prompt message will be generated.

[0122] In this embodiment, the source system write interface is constructed using RPA (Robotic Process Automation) technology, including one or more of the fusion platform write interface and the image 2.0 system write interface. Calling the fusion platform write interface can write structured data into the fusion platform, and calling the image 2.0 system write interface can write unstructured data into the image 2.0 system.

[0123] In this embodiment, when parsing the query requirements of natural language and obtaining the query results from the contract order reimbursement business data of the edge data platform, the following steps are included:

[0124] All contract order reimbursement business data in the edge data platform are sorted out to construct a redundant table structure definition for contract order reimbursement, and the redundant table structure definition is converted into a natural language expression to obtain the prompt word C part. Specifically, various detailed data of contract order reimbursement are extracted from the edge data platform, and the data are summarized, sorted and counted. For each contract order reimbursement business, a business redundant table is constructed using SQL standard statements to compress the multi-dimensional data of a contract order reimbursement business into one, and a redundant table structure definition for contract order reimbursement is constructed, and the table structure definition is converted into a natural language expression to construct the prompt word C part;

[0125] Obtaining data query requirements in natural language, thereby constructing the prompt word part D, and concatenating the prompt word part C and the prompt word part D to form a standard prompt word;

[0126] Similarly, when forming standard prompt words, the aforementioned "multi-sensory cognitive construction method" is also used. For prompt words C (reimbursement record redundancy table) and prompt word D (user's query question for reimbursement records), the following sensory description can be used to construct:

[0127] Sensory description of prompt word C:

[0128] Visual carrier:

[0129] Prompt word C can be compared to a detailed transaction history, including the destination, time, responsible person and detailed content of each sum of money. Visually, it is like rows of dense and detailed data columns, forming dense text and numbers on paper.

[0130] Tactile carrier:

[0131] Prompt word C can be compared to a thick account book, the thickness and weight of which are related to the rich information it contains.

[0132] Sensory description of prompt word D:

[0133] Auditory carrier:

[0134] Cue word D can be thought of as a precise and unambiguous question.

[0135] Visual carrier:

[0136] The prompt word D can be imagined as an efficient magnifying glass, which focuses on filtering out key information from the messy and disordered data.

[0137] The fusion of prompt word C and prompt word D:

[0138] When constructing the fused prompt word, the visual and tactile carriers can be used to describe the rich and complex data of the prompt word C, while the auditory and visual carriers can be used to describe the precise query problem of the prompt word D. For example:

[0139] Fusion prompt: "The thick account book in your hand (tactile carrier of prompt word C) records countless expense details. Each line of data is clearly marked with the expense number, time, person in charge, project, and amount (visual carrier of prompt word C). Now you have a clear question: How much is the total amount of all the expenses this year? (auditory carrier of prompt word D) Focus the magnifying glass on the account book and search only for this clear answer, skip all irrelevant details, and focus directly on the total amount (visual carrier of prompt word D)."

[0140] Through this fusion, prompt words C and D work together on the large language model, and the model can deeply understand complex data and use it to answer specific and concise questions. This process is closer to the actual needs of users and can quickly find core information in big data.

[0141] Input the standard prompt words into the preset language model for logical decision-making and reasoning, thereby generating a standard SQL query statement that conforms to the natural language structure;

[0142] Query the edge data platform according to the standard SQL query statement to obtain the query result data.

[0143] Embodiment 2:

[0144] This embodiment proposes a power grid contract order reimbursement management system, such as Figure 2 As shown, it includes a user end, a service end, a cross-system integration module and an edge data platform, and also includes a contract order reimbursement integration platform, which realizes communication connection with the user end, the service end, the cross-system integration module and the edge data platform respectively, wherein:

[0145] As the communication hub of the system, the contract order reimbursement integrated platform effectively connects and coordinates multiple functional modules such as the user end, service end, cross-system integration module and edge data platform, ensuring the efficiency and consistency of information transmission between modules. At the same time, it realizes communication with other source systems through a secure and reliable management information regional network interface;

[0146] The user end includes:

[0147] The user data collection module is used to collect the business data of the contract order reimbursement to be approved after the initiator starts the power grid contract order reimbursement process. If the business data of the contract order reimbursement to be approved passes the signature verification, the business data of the contract order reimbursement to be approved is stored in the edge data platform;

[0148] The user data collection module is designed to collect the original data related to contract order reimbursement. When the initiator starts the reimbursement process, the required structured and unstructured information is collected through the selection, entry and upload mechanism, and the data is stored in the edge data platform through the secure network channel of the management information zone;

[0149] The process approval module is used to obtain the contract order reimbursement business data to be approved through the edge data platform, determine the applicable process template according to the preset matching rules, and remove the duplicate approval nodes of the process template. The process template includes the corresponding approval level and approval personnel; it is also used to record the approval data of each approval stage and store it in the edge data platform;

[0150] The process approval module is set up according to the internal control system of the power grid, and is used by approval personnel at all levels to approve contract order reimbursement matters step by step. This module automatically generates approval levels and assigns approval personnel responsibilities based on reimbursement business data, forms approval to-do tasks and sends them to the corresponding level approvers. Approving personnel enter the review interface through the to-do items, and check the rationality of the contract order reimbursement business and other factors. After selecting approval, rejection or transfer, the process automatically flows to the next level of approval until the entire approval level is completed. During this process, the module records the opinions of each approval stage and its corresponding timestamps in detail, and stores the approval data on the edge data platform through the secure network channel of the management information area;

[0151] The statistical analysis module is used to manage the contract order reimbursement business data of the edge data platform using preset management reports; it is also used to parse natural language query requirements and obtain query results from the contract order reimbursement business data of the edge data platform;

[0152] The statistical analysis module uses the data resources of the edge data platform to query and deeply analyze the contract order reimbursement data. The data calculated by the edge data platform will be displayed in a pre-set management report format, or the relevant data will be presented upon request by the report user. This module can perform multi-dimensional query and comprehensive analysis of reimbursement data, thereby generating statistical reports that are conducive to decision-making;

[0153] The statistical analysis module includes an intelligent query submodule, which is specially designed to receive and parse the statistical analysis query requirements of the initiator in the form of natural language, understand the user's intention through the open source large language model, and use the given table structure definition in the edge data platform to generate SQL (Structured Query Language, a database language with multiple functions such as data manipulation and data definition) query statements and query the relevant data and return it to the user, thereby realizing customized query results; the intelligent query submodule is equipped with a user interaction function unit, which is interconnected with the data acquisition unit, the decision analysis unit and the execution unit;

[0154] The server includes:

[0155] The automatic approval auxiliary module is used to perform signature verification on the contract order reimbursement business data to be approved. If the contract order reimbursement business data to be approved fails to pass the signature verification, the data items that do not meet the verification rules are pointed out;

[0156] The automatic approval auxiliary module provides intelligent assistance for the approval process through pre-configured data verification functions. The system uses flexibly configurable data verification rules to verify the unstructured information data submitted by the initiator in the user data collection module. Only when the data passes the verification can the initiator successfully submit and store the data in the edge data platform through the management information area; otherwise, the system will refuse to submit and point out the data items that do not meet the verification rules, and guide the initiator to correct the data;

[0157] The message reminder module is used to push status changes and process progress information to the initiator and approver according to the process template, form approval tasks and send them to the approvers at the corresponding level;

[0158] The message reminder module is responsible for sending instant notifications according to the generated reimbursement business approval level and the list of approvers when the contract order reimbursement business status changes. Whenever the reimbursement process status is updated, the system will automatically push the status change and process progress information to the initiator and approver;

[0159] The cross-system integration module is used to read the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source systems, and write it into the target source system; or obtain the contract order reimbursement business data that needs to be synchronized from other source systems and store it in the edge data platform. The cross-system integration module includes the cross-system task scheduling submodule, the cross-system data writing submodule, and the cross-system data acquisition submodule;

[0160] The cross-system task scheduling submodule adopts a unified asynchronous task queue architecture design. When receiving an external request, the system inserts the tasks in the request into the bottom of the queue in a predetermined order and executes the tasks in the queue in accordance with the first-in-first-out principle. Once the task is executed, the system returns a clear status code and response information in the form of response parameters to facilitate tracking of task execution status and results;

[0161] The cross-system data writing sub-module integrates multiple source system writing interfaces built using RPA (Robotic Process Automation) technology, including the fusion platform writing interface, the Image 2.0 system writing interface, etc.

[0162] The cross-system data acquisition submodule is also equipped with other source system data acquisition interfaces created by RPA technology, such as the economic and legal system 2.0 contract interface and the ERP system budget information acquisition interface.

[0163] RPA technology specifically refers to automated data writing programs written with the help of Python Selenium tool library and other related development tools. In this system, the cross-system data writing submodule effectively integrates multiple RPA writing programs into one, realizing real-time and active data writing operations on different source end systems. In this way, the system successfully breaks the information islands existing between multiple source end systems and promotes the smooth flow and exchange of data flows.

[0164] like Figure 3 As shown, the steps for initiating contract order reimbursement by the system of this embodiment are as follows:

[0165] Step 1: The initiator starts the contract order reimbursement process and collects the required structured and unstructured information through the user data collection module selection, entry and upload mechanism: specifically, the initiator selects the associated annual budget code or name and the associated contract code or name data, enters the reimbursement reason, invoice amount, payment amount and whether data synchronization is required, and uploads the invoice and project process data with signatures and seals; the user data collection module obtains the annual budget and associated contract data selected by the initiator from the edge data platform, such as the annual budget amount, account code, control level, competent department, person in charge, contract amount and other data and displays them on the page, and at the same time obtains the remaining allowable reimbursement amount of the associated contract through calculation and processing and compares it with the invoice amount and payment amount entered by the initiator;

[0166] Step 2: After the initiator selects, enters and uploads all data through the user data collection module, clicks the submit function button, and the automatic approval auxiliary module converts the unstructured information uploaded by the initiator into base64 encoding and communicates with the signature recognition interface in the signature automatic recognition function unit. The signature recognition interface receives the data converted into Base64 encoding and restores it to PNG format image data. Then, the pre-trained improved YOLOV3 model, that is, the signature recognition model trained in Example 1, receives this PNG image as input and outputs a series of predicted signature borders, each of which is associated with a confidence score and a probability distribution of the category. By setting the confidence threshold, high-confidence prediction results are screened out, and the non-maximum suppression (NMS) algorithm is used to eliminate redundant detection results for the high-confidence prediction results, and finally the actual coordinate position information of all signatures in the image is located. The location information is converted into JSON format and then passed back to the automatic approval auxiliary module. When the automatic approval auxiliary module receives the prediction result containing the actual coordinate location information of the signature returned by the signature recognition interface, it indicates that the document has been signed. At this time, the initiator's request is approved, and all the data submitted by the initiator is stored in the edge data platform through the secure network channel of the management information zone. On the contrary, if the signature is missing, it will communicate with the process approval module to generate a prompt: "The system has not detected the signature, please upload the stamped version of the file", and automatically return the process to the initiator of the contract order reimbursement business.

[0167] Step 3: After the initiator successfully submits the data, the process approval module obtains the contract order reimbursement data to be approved through the edge data platform, and uses the approval node judgment and distribution processing unit to apply numerical calculation and matching rules to determine the applicable process template FLMB. Each process template FLMB contains the approval level and approver corresponding to the template; the process template FLMB adopts a multi-level classifier technology. The classifier comprehensively adopts a decision tree algorithm combined with an open source large language model. First, the decision tree algorithm is used to classify and calculate the factors specified by the contract order reimbursement business data to be approved according to dynamically configurable business rules to obtain a process template matching the contract order reimbursement business data to be approved. Then, the process nodes and approvers in the process template matching the contract order reimbursement business data to be approved are used to construct the prompt word part A, and the duplicate approval node removal instruction of "analyzing the accepted prompt word part A step by step, removing the approval nodes of duplicate approvers, and forming a process template FLMB without duplicate approvers" is used to construct the prompt word part B. The prompt word parts A and B are input into the large language model together, and finally the process template without duplicate approvers output by the large language model is obtained. Therefore, by utilizing the decision tree classification decision and large model decision decomposition capabilities, the contract order type, contract amount, invoice amount and control level and other factors are classified and calculated according to dynamically configurable business rules to accurately locate the process template FLMB corresponding to the contract order reimbursement, while reducing repeated approval processes; after the process approval module confirms the process template FLMB to which the contract order reimbursement business belongs, it automatically pushes status changes and process progress information to the initiator and the approver through the message reminder module, forming an approval to-do task and sending it to the approver at the corresponding level. The approver enters the review interface through the to-do items, verifies the rationality of the contract order reimbursement business and other factors, and after selecting approval, rejection or transfer, the process automatically flows to the next level of approval until the entire approval level is completed. During this process, the process approval module records the opinions and corresponding timestamps of each approval stage in detail, and stores the approval data on the edge data platform through the secure network channel of the management information zone;

[0168] Step 4: After the approval is completed, the cross-system integration module reads the data queue in the edge data platform that has been approved and needs to be synchronized to other source systems, communicates with the cross-system data writing sub-module, and synchronously writes the business data into the target source system through the horizontal domain boundary of the management information zone, ensuring seamless data docking and synchronization with each source system;

[0169] The detailed process of the cross-system data writing submodule to synchronously write data into the target source system is as follows: The cross-system task scheduling submodule adopts a unified asynchronous task queue architecture design. When an external request is received, the system inserts the task in the request to the bottom of the queue in a predetermined order and executes the tasks in the queue according to the first-in-first-out principle. Once the task is completed, the system returns a clear status code and response information in the form of response parameters to facilitate tracking of task execution status and results; the cross-system data writing submodule integrates a variety of source system writing interfaces built using RPA (Robotic Process Automation) technology, including the fusion platform writing interface, the image 2.0 system writing interface, etc. The cross-system data writing submodule calls the fusion platform writing interface for structured data and writes the data to the fusion platform; at the same time, it calls the image 2.0 system writing interface to write unstructured data to the image 2.0 system; when the cross-system data writing submodule receives the system response code 200, it means that the writing is successful and the data writing is completed; when the cross-system data writing submodule receives the system response code 403, 500, 502, etc., it means that the service is interrupted or there is a problem with the interface. The cross-system data writing submodule will try to call the writing interface again after 1 minute; if it still fails after three calls, it will communicate with the message reminder module and send a to-do message to remind the system administrator;

[0170] Step 5: When the initiator or manager needs to conduct statistical analysis on the historical contract order reimbursement data, the data management is performed through the management report preset in the statistical analysis module. The preset management report data is displayed after statistical calculation using the edge data platform. If other source system data is required, the statistical analysis module communicates with the cross-system data acquisition submodule in the cross-system integration module to obtain the required data. The specific acquisition process is the same as step 4. If there is a need for personalized data management, it is managed through the intelligent question submodule of the statistical analysis module. The detailed process is as follows: Figure 4 As shown, the following steps are included:

[0171] S1: The data acquisition unit extracts various detailed data of contract order reimbursement from the edge data platform, summarizes, organizes and counts them, and for each contract order reimbursement business, uses SQL standard statements to build a business redundant table to compress the multi-dimensional data of a contract order reimbursement business into one, builds a redundant table structure definition of the contract order reimbursement, and converts the table structure definition into a natural language expression to construct the prompt word C part;

[0172] S2: Users use natural language to raise personalized data query requirements through the user interaction function unit in the intelligent question submodule, for example: "How much has been paid for the project "State Grid XXX" in 2023?";

[0173] S3: The decision analysis unit obtains the prompt word part C converted by the data acquisition unit, constructs the prompt word part D with the intelligent question natural language received by the user interaction function unit, and concatenates the prompt word part C and the prompt word part D to form a standard prompt word;

[0174] S4: The formed standard prompt words are passed to the built-in large language model (parameters are set to 14B) and logical decision-making and reasoning are performed to generate a standard SQL query statement that conforms to the natural language structure. The execution unit then connects to the edge data platform through communication to execute the standard SQL query statement and obtain the query result data;

[0175] S5: The execution unit feeds back the query result data to the user interaction function unit, and the latter displays the query result to the query user, and the query user obtains the relevant results.

[0176] In summary, the present invention proposes a method for managing power grid contract order reimbursement, and designs a software system for managing the entire process of power grid enterprise contract order reimbursement according to the proposed method, which has the following beneficial effects:

[0177] 1. Using the process approval module, based on the internal control system of the power grid and multi-level classifier technology, it can automatically generate accurate approval levels based on actual reimbursement business data, and intelligently assign the responsibilities of approval personnel, avoiding errors and delays that may be caused by manual intervention. This feature greatly improves the standardization and execution efficiency of the approval process, making the approval work more transparent and fair, while also reducing the workload of managers and ensuring that contract order reimbursement matters are handled in a timely and reasonable manner;

[0178] 2. The intelligent query submodule embedded in the statistical analysis module uses advanced natural language processing technology to understand and analyze the statistical analysis requirements in natural language form proposed by users, and quickly generate customized query results. The use of this module simplifies the user's query operation and reduces the requirements for professional knowledge, allowing all types of users to easily and quickly obtain the required report data, providing strong support for decision-making;

[0179] 3. The automatic signature recognition function unit in the automatic approval auxiliary module uses an image recognition algorithm to intelligently detect signature information in unstructured information, effectively preventing approval problems caused by missing or forged signatures. This function not only automatically advances the complete approval process, but also quickly provides feedback to the initiator when problems are found, ensuring the authenticity and integrity of the document and significantly improving the overall compliance of the contract order reimbursement business;

[0180] 4. The cross-system integration module uses RPA technology to integrate cross-system task scheduling, data writing and data acquisition sub-modules, and realizes seamless docking of different source systems and real-time data synchronization. By eliminating information islands, the tedious process of manual operation and data migration is greatly reduced, and the speed and accuracy of data flow are significantly improved, thereby enhancing the overall effectiveness and operating efficiency of the power grid contract order reimbursement management system, laying a solid foundation for the company's refined management and decision-making support.

[0181] The above is only a preferred embodiment of the present invention, and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A power grid contract order reimbursement management method, characterized in that: The following steps are involved: After the initiator starts the grid contract order reimbursement process, collect the contract order reimbursement business data to be approved; Perform signature verification on the contract order reimbursement business data to be approved. If the contract order reimbursement business data to be approved fails to pass the signature verification, prompt the data items that do not meet the verification rules and re-collect the contract order reimbursement business data to be approved; if the contract order reimbursement business data to be approved passes the signature verification, store the contract order reimbursement business data to be approved in the edge data platform; Obtain the contract order reimbursement business data to be approved from the edge data platform, determine the applicable process template according to the preset matching rules, and remove the repeated approval nodes of the process template, wherein the process template includes the corresponding approval level and approval personnel; According to the process template, the approval tasks are sent to the corresponding level approvers, and the approval data of each approval stage is recorded and stored in the edge data platform; Read the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source systems, and write it into the target source system, or obtain the contract order reimbursement business data that needs to be synchronized from other source systems and store it in the edge data platform; Use preset management reports to manage the contract order reimbursement business data of the edge data platform; Parse natural language query requirements and obtain query results from the contract order reimbursement business data of the edge data platform; The contract order reimbursement business data to be approved includes structured and unstructured information. When the signature verification of the contract order reimbursement business data to be approved is performed, the following steps are included: Convert the unstructured information in the contract order reimbursement business data to be approved into image data; Input the image data into a trained signature recognition model to obtain a predicted signature border, each predicted signature border being associated with a confidence score and a probability distribution of a category; Through the specified confidence threshold, high-confidence prediction results are screened out from the predicted signature border, and redundant detection results are eliminated from the high-confidence prediction results to obtain the final result; if the final result is not empty, the contract order reimbursement business data to be approved has passed the signature verification; if the final result is empty, the contract order reimbursement business data to be approved has not passed the signature verification; Wherein, determining the applicable process template according to the preset matching rule and removing the repeated approval nodes of the process template includes: A decision tree algorithm is used to classify and calculate the factors specified in the contract order reimbursement business data to be approved according to dynamically configurable business rules to obtain a first process template; Using the large language model, the first process template is analyzed according to its process steps, and the approval nodes of the duplicate approvers are removed to form a second process template without duplicate approvers. Using the large language model to obtain the second process template specifically includes: Prompt word part A is constructed according to the process nodes and approvers in the first process template, and prompt word part B is constructed according to the instruction to remove duplicate approval nodes. Prompt word part A is used to describe the original complex activities, and prompt word part B is used to describe the optimized results. Prompt word part A and prompt word part B are fused and used as the input of the large language model.

2. The power grid contract order reimbursement management method according to claim 1, characterized in that: Before inputting the image data into the trained signature recognition model, the step of training the signature recognition model is also included, specifically including: Select a specified image recognition model, add a detail capture layer before the output layer of the image recognition model, replace the last pooling layer of the image recognition model with a convolution layer with a step size of 1, and obtain a signature recognition model; the detail capture layer includes 3 convolution layers, each of which is followed by a ReLU activation function and a normalization layer, and the first convolution layer uses 1 1 filter, the second and third convolutional layers both use 3 3 filters; Use the preset signature dataset to train the signature recognition model.

3. The power grid contract order reimbursement management method according to claim 1, characterized in that: The construction of prompt word part A and the construction of prompt word part B both include: selecting a specified number of carriers, each carrier represents a sense organ, concretizing each sensory description through a specific carrier scenario, and forming prompt words corresponding to different sense organs.

4. The power grid contract order reimbursement management method according to claim 1, characterized in that: Reading the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source systems, and writing it to the target source system, includes the following steps: Obtain synchronization requests from other source systems and insert the corresponding contract order reimbursement business data into the bottom of the data queue in a predetermined order; According to the first-in-first-out principle, the preset source system write interface is called to write the contract order reimbursement business data in the data queue into the corresponding source system in sequence.

5. The power grid contract order reimbursement management method according to claim 4, characterized in that: When calling the preset source end system write interface to write the contract order reimbursement business data in the data queue into the corresponding source end system in sequence, it includes: Obtain the response code of the source end system corresponding to the current contract order reimbursement business data in the data queue, and determine whether the writing is successful according to the response code. If the writing is successful, write the next contract order reimbursement business data in the data queue into the corresponding source end system; If the writing fails, wait for a preset time and then call the source system writing interface again to write the current contract order reimbursement business data in the data queue into the corresponding source system; If the number of unsuccessful write attempts exceeds the preset value, the next contract order reimbursement business data in the data queue will be written to the corresponding source system and a prompt message will be generated.

6. The power grid contract order reimbursement management method according to claim 5, characterized in that: The source system writing interface is constructed using RPA technology, including one or more of a fusion platform writing interface and an image 2.0 system writing interface.

7. The power grid contract order reimbursement management method according to claim 1, characterized in that: When parsing the query requirements of natural language and obtaining the query results from the contract order reimbursement business data of the edge data platform, the following steps are included: All contract order reimbursement business data in the edge data platform are sorted, a redundant table structure definition of contract order reimbursement is constructed, and the redundant table structure definition is converted into a natural language expression form, and a prompt word C part is constructed according to the redundant table structure definition in the natural language expression form, and the prompt word C part is used to describe complex data; Acquire a data query requirement in natural language, construct a prompt word D part according to the data query requirement in natural language, wherein the prompt word D part is used to describe a precise query problem, and concatenate the prompt word C part and the prompt word D part to generate a standard prompt word; Input standard prompt words into the language model for logical decision-making and reasoning, thereby generating standard SQL query statements that conform to the natural language structure; Query the edge data platform according to the standard SQL query statement to obtain the query result data.

8. A power grid contract order reimbursement management system, characterized in that: The management system adopts the power grid contract order reimbursement management method according to any one of claims 1 to 7, and the management system includes a user end, a service end, a cross-system integration module and an edge data platform, wherein: The user terminal comprises: The user data collection module is used to collect the business data of the contract order reimbursement to be approved after the initiator starts the power grid contract order reimbursement process. If the business data of the contract order reimbursement to be approved passes the signature verification, the business data of the contract order reimbursement to be approved is stored in the edge data platform; The process approval module is used to obtain the contract order reimbursement business data to be approved through the edge data platform, determine the applicable process template according to the preset matching rules, and remove the duplicate approval nodes of the process template. The process template includes the corresponding approval level and approval personnel; it is also used to record the approval data of each approval stage and store it in the edge data platform; The statistical analysis module is used to manage the contract order reimbursement business data of the edge data platform using preset management reports; it is also used to parse natural language query requirements and obtain query results from the contract order reimbursement business data of the edge data platform; The server includes: The automatic approval auxiliary module is used to perform signature verification on the contract order reimbursement business data to be approved. If the contract order reimbursement business data to be approved fails to pass the signature verification, the data items that do not meet the verification rules are pointed out; The message reminder module is used to push status changes and process progress information to the initiator and approver according to the process template, form approval tasks and send them to the approvers at the corresponding level; The cross-system integration module is used to read the contract order reimbursement business data that has been approved in the edge data platform and needs to be synchronized to other source systems, and write it into the target source system; or obtain the contract order reimbursement business data that needs to be synchronized from other source systems and store it in the edge data platform.

Citation Information

Patent Citations

  • Data processing method based on information identification technology and related device

    CN110264288A

  • Check seal verification method and device, electronic equipment and medium

    CN117523586A