A method and system for analyzing and processing power business data

Abstract

The present application provides a method and system for analyzing and processing electric power business data. Among them, the present application receives business expansion application work orders containing structured text and unstructured images through the business hall terminal; parses the key entity fields of the text and extracts the anti-counterfeiting feature points of the image, maps them to the unified vector space through the multimodal alignment layer to generate a semantic image feature stream; loads them into the hardware acceleration chip to perform parallel matrix operations, identifies the logical binding relationship between entities and outputs a nested relationship topology diagram; calls the historical work order decision path to train the reinforcement learning agent to generate a control strategy; compares the similarity between the anti-counterfeiting feature points and the pre-stored template, and outputs the verification results containing missing identification and anti-counterfeiting anomalies; if the fields are complete and the anti-counterfeiting similarity is greater than or equal to the threshold, the work order is released to the business system, otherwise the work order is frozen and an alarm is triggered. The present application improves processing efficiency and blocks the risk of abnormal generation, realizing automated anti-counterfeiting verification and process control of electric power work orders.

Description

Technical Field

[0001] The present application relates to the technical field of power business management, and in particular to a method and system for analyzing and processing power business data. Background Art

[0002] The power industry's expansion and installation work order processing process requires the coordinated verification of structured text data and unstructured image data. Specifically, this requires automated verification of key fields in the work order text and the anti-counterfeiting features of the attached image, while also checking for the completeness of required fields and business logic conflicts. This scenario requires the system to significantly improve work order processing efficiency while ensuring business security, avoiding delays and omissions caused by manual verification.

[0003] Currently, mainstream solutions utilize a separate image-text processing engine. This engine first independently extracts key fields from the work order text using optical character recognition technology and stores them in a database. It then performs a similarity comparison on attached images based on pre-stored templates. The system then logically verifies the structured text fields based on a pre-set rule base, ultimately making a comprehensive judgment based on the independently processed image and text results.

[0004] However, this solution's separate processing of text and image data prevented the establishment of cross-modal relationships. For example, the spatial binding between the applicant's name field and the signature image area was missing, increasing the risk of anomalies. Static rule bases were unable to dynamically adapt to complex business scenarios, making it difficult to cover fuzzy rules in historical work orders, leading to missed logical conflicts. Image template comparisons were coarse-grained, ignoring the spatial consistency between text semantics and image areas, resulting in a high rate of anti-counterfeiting vulnerabilities. Summary of the Invention

[0005] The present application provides a method and system for analyzing and processing electric power business data, which is used to solve the problems of anti-counterfeiting loopholes and missed detection of logical conflicts in the prior art.

[0006] In a first aspect, the present application provides a method for analyzing and processing power business data, comprising:

[0007] Receive a business expansion application form submitted by a user through a business hall terminal device, wherein the business expansion application form includes a structured work form text and an unstructured attachment image;

[0008] Parsing the key entity fields in the work order text, simultaneously extracting the anti-counterfeiting feature points of the attached image, and mapping the text entities and image features into a unified vector space through a multimodal alignment layer to generate a semantic image feature stream;

[0009] Loading the semantic image feature stream into a hardware acceleration chip, performing parallel matrix operations to identify the logical binding relationship between the key entity fields, and outputting a nested relationship topology graph;

[0010] Call the decision path sequence in the historical work order database, train the reinforcement learning agent through time series modeling, and generate a control strategy based on the state-action reward function;

[0011] According to the nested relationship topology diagram and the control strategy, the number of missing required fields is verified and key entity logic conflicts are detected. The similarity between the anti-counterfeiting feature points and the pre-stored reference template is compared, and a verification result including missing field identifiers and anti-counterfeiting abnormal areas is output;

[0012] A judgment operation is triggered based on the verification result. If the field is complete and the anti-counterfeiting similarity is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system. If the field is missing or the anti-counterfeiting similarity is less than the preset threshold, the blocking instruction is executed, the work order flow is frozen and an alarm signal is triggered to the business hall terminal.

[0013] Optionally, the key entity fields in the work order text are parsed, anti-counterfeiting feature points of the attachment image are extracted simultaneously, and the text entities and image features are mapped to a unified vector space through a multimodal alignment layer to generate a semantic image feature stream, including:

[0014] Scan the work order text to locate the type and content value of key entity fields, scan the attachment image, and extract the position coordinates and contour shape of anti-counterfeiting feature points;

[0015] Input the type and content value of the key entity field and the position coordinates and contour shape of the anti-counterfeiting feature point into the multimodal alignment layer, and generate a pairing relationship according to spatial coordinate mapping;

[0016] Based on the pairing relationship, corresponding fused feature vectors are generated respectively, and all the fused feature vectors are arranged in the order of submission time of the business expansion application work order to generate a semantic image feature stream.

[0017] Optionally, the type and content value of the key entity field and the position coordinates and contour shape of the anti-counterfeiting feature point are input into a multimodal alignment layer, and a pairing relationship is generated according to spatial coordinate mapping, including:

[0018] Establish the position coordinates of the key entity fields, take the upper left corner of the work order as the origin and the lower right corner as the maximum coordinate point, divide the text area by line, and assign position coordinates to each key entity field;

[0019] Extracting the contour shape of the anti-counterfeiting feature point, calculating the geometric center point of the contour shape, extending from the center point to the contour boundary, and generating position coordinates;

[0020] If the overlapping area of ​​the position coordinate region of the key entity field and the position coordinate region of the anti-counterfeiting feature point is greater than or equal to a minimum overlapping threshold, an initial pairing relationship set is generated;

[0021] For many-to-one conflicts, the pairing relationship with the largest overlapping area is retained. The unpaired key entity fields are matched to the anti-counterfeiting feature point with the closest Euclidean distance to the center point, marked as a low-confidence pairing, and the low-confidence pairing is eliminated in the initial pairing relationship set to obtain a valid pairing relationship.

[0022] Optionally, call the decision path sequence in the historical work order database, train the reinforcement learning agent through time series modeling, and generate a control strategy based on the state-action reward function, including:

[0023] Call the historical work order database, match similar work orders according to the work order type identifier, extract the decision path sequence corresponding to the work order, and split the decision path sequence into discrete state sequences and action sequences in chronological order;

[0024] Constructing a long short-term memory network as a temporal modeling layer, inputting the state sequence and outputting the probability distribution of the actions that can be executed in each state;

[0025] Based on the state-action reward function, a dynamic reward value is assigned to each action in the action sequence, positive rewards are given to compliant actions, negative penalties are given to illegal actions, and delayed actions are linearly penalized according to the delay duration;

[0026] The selection probability distribution is weightedly integrated with the dynamic reward value to generate a value evaluation result for each action. The actions with the highest value evaluation results are screened and arranged in chronological order as a priority action sequence, which is output as the control strategy for the current work order.

[0027] Optionally, based on the nested relationship topology and the control strategy, the number of missing required fields is verified and key entity logic conflicts are detected, the similarity between the anti-counterfeiting feature points and the pre-stored reference template is compared, and a verification result including missing field identifiers and anti-counterfeiting abnormal areas is output, including:

[0028] Traversing all entity nodes in the nested relationship topology graph, scanning predefined mandatory field labels, counting the number of missing fields and generating a missing field identifier list;

[0029] Calling the priority action sequence in the control strategy, locating the node pairs pointed to by conflicting connections in the nested relationship topology graph, and marking the coordinates and conflict types of the node pairs;

[0030] Superimposing the anti-counterfeiting feature points with the pre-stored reference template at the pixel level, calculating the contour overlap as a similarity quantification value, and marking the coordinates of the contour difference area;

[0031] The missing field identification list, the coordinates and conflict types of the node pairs, the similarity quantization values ​​and the coordinates of the contour difference areas are integrated, and a verification result including the missing field identification and the anti-counterfeiting abnormal area is output.

[0032] Optionally, the semantic image feature stream is loaded into a hardware acceleration chip, parallel matrix operations are performed to identify the logical binding relationship between the key entity fields, and a nested relationship topology diagram is output, including:

[0033] Loading the semantic image feature stream into a hardware acceleration chip, and generating a logical association strength relationship matrix between each field and other fields based on the type and content value of the key entity fields;

[0034] Synchronously calculating cross-logic results of a plurality of the logic association strength relationship matrices on the hardware acceleration chip, and marking strong logical connections and conflicting connections in the cross-logic results;

[0035] With the key entity fields as nodes and the strong logical connections and conflicting connections as lines, a tree-like relationship network is constructed as a nested relationship topology diagram.

[0036] Optionally, a judgment operation is triggered based on the verification result. If the fields are complete and the anti-counterfeiting similarity is greater than or equal to a preset threshold, a work order release instruction is executed to push the work order to the power business execution system. If the fields are missing or the anti-counterfeiting similarity is less than a preset threshold, a blocking instruction is executed to freeze the work order flow and trigger an alarm signal to the business hall terminal, including:

[0037] Reading the missing quantity and the contour similarity quantization value in the verification result to trigger a determination operation;

[0038] If the missing number is zero and the contour similarity quantization value is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system;

[0039] If the number of missing items is greater than zero or the contour similarity quantization value is less than the preset threshold, a blocking instruction is executed to freeze the write permission of the work order database, and based on the position coordinates of the key entity fields corresponding to the missing number in the nested relationship topology diagram, combined with the pixel coordinates of the contour difference area in the anti-counterfeiting feature map, an alarm message is generated and sent to the business hall terminal.

[0040] In a second aspect, the present application provides a power business data analysis and processing system, comprising:

[0041] A receiving module, configured to receive a business expansion application form submitted by a user through a business hall terminal device, wherein the business expansion application form includes a structured work form text and an unstructured attachment image;

[0042] A mapping module is used to parse the key entity fields in the work order text, simultaneously extract the anti-counterfeiting feature points of the attached image, and map the text entities and image features into a unified vector space through a multimodal alignment layer to generate a semantic image feature stream;

[0043] A recognition module is used to load the semantic image feature stream into a hardware acceleration chip, perform parallel matrix operations to identify the logical binding relationship between the key entity fields, and output a nested relationship topology diagram;

[0044] The generation module is used to call the decision path sequence in the historical work order database, train the reinforcement learning agent through time series modeling, and generate the control strategy based on the state-action reward function;

[0045] A verification module is used to verify the number of missing required fields and detect key entity logic conflicts based on the nested relationship topology diagram and the control strategy, compare the similarity of the anti-counterfeiting feature points with the pre-stored reference template, and output a verification result including missing field identifiers and anti-counterfeiting abnormal areas;

[0046] The judgment module is used to trigger a judgment operation based on the verification result. If the field is complete and the anti-counterfeiting similarity is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system. If the field is missing or the anti-counterfeiting similarity is less than the preset threshold, the blocking instruction is executed, the work order flow is frozen and an alarm signal is triggered to the business hall terminal.

[0047] In a third aspect, the present application provides a computing device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for analyzing and processing electric power business data as described in the first aspect above.

[0048] In a fourth aspect, the present application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a method for analyzing and processing electric power business data as described in the first aspect.

[0049] This application receives the text and image data of business expansion application work orders through the business hall terminal to realize multi-source input integration; parses the key entity fields of the text and simultaneously extracts the anti-counterfeiting feature points of the image, generates a semantic feature stream through multimodal alignment layer mapping, eliminates the problem of image and text separation and forcibly associates text and image features; loads the semantic feature stream to the hardware acceleration chip to perform parallel operations, and quickly outputs a nested relationship topology map to identify logical bindings and conflicts between entities; calls historical work order decision paths to train reinforcement learning agents, generates dynamic control strategies to replace static rule bases, and adaptively optimizes the approval process; verifies the integrity of required fields, logical conflicts and anti-counterfeiting feature similarities based on topology maps and control strategies, and outputs closed-loop verification results to realize multi-dimensional risk screening; automatically triggers work order release or freeze instructions based on the verification results, completes risk control interception and alarm, and at the same time improves the abnormal generation and recognition capabilities through cross-modal binding of semantic feature streams.

[0050] Furthermore, by scanning the work order text to locate the type and content value of key entity fields, the attached image is simultaneously scanned to extract the position coordinates and contour shape of the anti-counterfeiting feature points. These two are then input into the multimodal alignment layer to perform spatial coordinate mapping operations to generate a pairing relationship. Based on this pairing relationship, fused feature vectors are generated and arranged in chronological order of work order submission to construct a semantic image feature stream. This process forcibly associates the spatial positions of text entities and image features, such as binding the customer name field to the signature area coordinates, completely eliminating anti-counterfeiting vulnerabilities caused by the separation of text and images. Furthermore, the time-series feature stream fully preserves business processing dependencies, providing a structured cross-modal input foundation for subsequent hardware-accelerated logical relationship recognition, thereby significantly improving verification accuracy and efficiency.

[0051] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] Figure 1 A flowchart of a method for analyzing and processing power business data provided by the present application is shown;

[0054] Figure 2 A schematic diagram of the structure of a power business data analysis and processing system provided by the present application is shown;

[0055] Figure 3 A schematic structural diagram of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0057] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0058] Researchers have found that in the process of processing business expansion and installation work orders in the power business, the existing solutions cannot establish cross-modal associations due to the separation of text and image data. For example, the spatial binding between the applicant's name field and the signature image area is missing, which increases the risk of anomaly generation. At the same time, the static rule base cannot dynamically adapt to complex business scenarios, and the fuzzy rules in historical work orders are difficult to cover, resulting in missed logical conflicts. In addition, the image template comparison granularity is coarse, ignoring the spatial consistency of text semantics and image areas, resulting in a high rate of anti-counterfeiting vulnerabilities. These defects seriously reduce the accuracy and security of work order processing and increase business delays and risks. Therefore, there is an urgent need for an efficient and intelligent method for analyzing and processing power business data.

[0059] In response to the above problems, the present invention proposes a method for analyzing and processing power business data, the core of which lies in the deep integration of multimodal data and the collaborative optimization of dynamic intelligent decision-making. Specifically, the business expansion application work order submitted by the user is received through the terminal equipment of the business hall, the key entity fields of the work order text are parsed and the anti-counterfeiting feature points of the attached image are extracted, and the multimodal alignment layer is used to map the two to a unified vector space to generate a semantic image feature stream; it is loaded into the hardware acceleration chip to perform parallel matrix operations, identify the logical binding relationship between entities and output a nested relationship topology diagram; the historical work order database is called, and the reinforcement learning agent is trained through time series modeling to generate an adaptive control strategy; based on this, the required fields are verified to be missing, logical conflicts and anti-counterfeiting similarities, the verification results are output and the work order flow control is triggered. This method effectively solves the problem of missing cross-modal associations caused by the split processing mentioned in the background technology, and realizes the spatial binding of text and image features through the multimodal alignment layer, which significantly reduces the risk of abnormal generation; the reinforcement learning agent dynamically adapts to complex scenes, covers fuzzy rules to eliminate logical conflict omissions; the refined anti-counterfeiting feature comparison is combined with semantic consistency verification to improve anti-counterfeiting accuracy and reduce the vulnerability rate.

[0060] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0061] Figure 1 A flowchart of a method for analyzing and processing power business data is provided for an embodiment of the present application. Figure 1 As shown, the method includes:

[0062] 101. Receive a business expansion application form submitted by a user through a business hall terminal device, wherein the business expansion application form includes a structured work form text and an unstructured attachment image;

[0063] In the above steps, the business hall terminal equipment refers to the equipment units set up at the power supply service outlets for user self-service operation or staff processing, such as computer terminals or touch screen self-service machines, which are connected to the background system for data collection; the business expansion application work order is a standardized application form document filled out by the user when applying for power expansion or other services, including necessary information and attachments; the structured work order text refers to the core text part of the work order, which contains preset fields such as customer name, contact information and power demand, etc. These data are stored in a fixed table format for easy machine reading; unstructured attachment images refer to image files submitted by users, such as scanned copies of user identifiers or real-life photos of the site. These images do not have a unified data format and are stored in the form of original images.

[0064] In an embodiment of the present application, the business expansion application work order submitted by the user is first received by using the business hall terminal device. The process includes the user inputting the structured work order text through the terminal interface, then uploading the unstructured image attachment, and finally the device confirming the receipt and integrating the stored work order data. The specific process is as follows: the user inputs the structured work order text data through the human-computer interaction interface of the business hall terminal device, and the device uses form input technology such as web forms or dedicated software interfaces to process these inputs. The system automatically verifies and standardizes the data format to ensure that it meets the preset requirements; then the user triggers the attachment upload function on the device interface, such as clicking the upload button, and uses the terminal's built-in camera or connected scanner to capture the unstructured image attachment. The device uses the file upload technology to receive the image format such as JPEG or PNG file and temporarily saves it to a temporary folder; finally, the terminal device integrates the structured text data and the image file into a complete work order, and uses data integration methods such as simple string splicing or file links to store it in the local cache or network server to ensure that the work order can be processed in subsequent steps.

[0065] In actual application, in a business hall in Location A, staff received business expansion application work orders submitted by users through the business hall terminal equipment. The work orders included structured work order text and unstructured attachment images. Specifically, during a working day, the terminal equipment received approximately 50 work orders. Each user was required to complete 15 structured fields when filling out the work order text on the device, such as user name, contact address, and service requirement details. At the same time, each work order also uploaded an unstructured attachment image, such as an on-site photo of the service point. During the operation, it took about 3 minutes for the terminal equipment to upload the work order data to the backend system, ensuring the smooth processing of the process. Through these processing methods, overall work efficiency has been improved, and service responsiveness has been enhanced.

[0066] In the above step 101, the business expansion application work order submitted by the user is uniformly received through the business hall terminal equipment. The work order contains structured work order text information and unstructured user attachment image data, realizing the digital collection of all elements of the user application information. The built-in intelligent processing module can simultaneously identify the structured demand information in the work order text and automatically process and parse the content in the associated attachment image, such as extracting the key identity information of the user's uploaded ID photo or identifying the equipment nameplate data of the on-site photo, effectively eliminating the information fragmentation and transcription risks in the previous manual processing link, and realizing the dual-modal fusion processing and intelligent verification of text requirements and image verification data, ensuring the completeness and accuracy of the declared data, significantly improving the automation level of work order entry and information extraction, providing high-quality and reliable support data for the back-end business approval process, accelerating the flow efficiency of the business expansion process, ensuring the standardization of business handling, and improving operational efficiency and service capabilities.

[0067] 102. Parse the key entity fields in the work order text, simultaneously extract the anti-counterfeiting feature points of the attachment image, and map the text entities and image features into a unified vector space through a multimodal alignment layer to generate a semantic image feature stream;

[0068] Optionally, the above step 102 may specifically include the following steps:

[0069] 1021. Scan the work order text, locate the type and content value of the key entity field, scan the attachment image, and extract the position coordinates and contour shape of the anti-counterfeiting feature points;

[0070] 1022. Input the type and content value of the key entity field and the position coordinates and contour shape of the anti-counterfeiting feature point into the multimodal alignment layer, and generate a pairing relationship according to spatial coordinate mapping;

[0071] Among them, step 1022 may specifically include the following processes: establishing the position coordinates of the key entity field, taking the upper left corner of the work order as the origin and the lower right corner as the maximum coordinate point, dividing the text area by line, and assigning position coordinates to each key entity field; extracting the contour shape of the anti-counterfeiting feature point, calculating the geometric center point of the contour shape, and extending the center point to the contour boundary to generate position coordinates; if the overlapping area of ​​the position coordinate area of ​​the key entity field and the position coordinate area of ​​the anti-counterfeiting feature point is greater than or equal to the minimum overlap threshold, generating an initial pairing relationship set; for many-to-one conflicts, retaining the pairing relationship with the largest overlapping area, matching the unpaired key entity field to the anti-counterfeiting feature point with the closest Euclidean distance to the center point, marking it as a low-confidence pairing, eliminating the low-confidence pairing in the initial pairing relationship set, and obtaining a valid pairing relationship.

[0072] 1023. Based on the pairing relationship, corresponding fused feature vectors are generated respectively, and all the fused feature vectors are arranged in the order of submission time of the business expansion application work order to generate a semantic image feature stream.

[0073] In the above steps, key entity fields refer to important data items contained in the work order text, such as core information such as customer name, contact information, and power demand value; anti-counterfeiting feature points refer to unique point features in the attached image used to verify the authenticity of the image, such as the edge position or shape center point of the marker; the multimodal alignment layer is an alignment mechanism used to coordinate different types of data such as text and images to the same mathematical representation framework; the vector space is a mathematical space that uses multidimensional array vectors to represent data for computer processing; the position coordinate is the specific numerical position of a point or area on the plane, such as the X-axis and Y-axis value pair, assuming that the upper left corner of the work order is the origin 0,0 and the lower right corner is the maximum value, such as 100,100, to form a coordinate system; the contour shape refers to the boundary lines around the anti-counterfeiting feature point, such as the geometric polygon description; the geometric center point refers to the center position of the contour shape calculated by calculation The average value of the shape boundary points is determined; the overlapping area refers to the area size of the overlapping parts of two areas, such as the text field area and the anti-counterfeiting point area, for example, in digital square units; the minimum overlap threshold is a preset parameter value as the minimum threshold for judging pairing, for example, 0.2 area units; the Euclidean distance is the straight-line distance between two points, calculated using the mathematical formula: the distance equals the square root of the first point's X coordinate minus the square of the second point's X coordinate plus the first point's Y coordinate minus the square of the second point's Y coordinate; the pairing relationship refers to the association and correspondence between the key entity field and the anti-counterfeiting feature point; the fused feature vector refers to a multidimensional data array generated by combining text information and image feature data, for example, consisting of a digital sequence; the semantic image feature stream refers to the data sequence stream formed by arranging all fused feature vectors in chronological order according to the time of work order submission, representing a complete information set.

[0074] In the embodiment of the present application, first, the work order text is scanned in step 1021 to locate the type and content value of the key entity field, and the attached image is scanned simultaneously to extract the position coordinates and contour shape of the anti-counterfeiting feature point. The specific process is that the device first uses text recognition technology to read the work order text, locates the key entity field, such as the customer name or contact number field, according to the preset dictionary, determines its type such as personal information text, and content value, and assigns a position coordinate area to each field. Assuming that the upper left corner of the work order is the origin 0, 0, and the maximum point in the lower right corner is 100, 100, the device divides the text area into rectangular blocks by line, and each field is assigned The coordinate values ​​of the upper left corner and the lower right corner, for example, the coordinates of the customer name field are 10, 10 for the upper left point and 30, 20 for the lower right point. At the same time, the device applies image processing technology to analyze the attached image, such as finding the boundary lines through image recognition methods, determining the contour shape of the anti-counterfeiting feature points such as polygonal boundaries, and then calculating the coordinates of the geometric center point of each contour. The center point is obtained by averaging the X-coordinate values ​​and Y-coordinate values ​​of all points on the boundary. For example, when the boundary points of a contour are 20, 20 and 40, 40, the center point is calculated to be 30, 30. Finally, the position coordinates and contour information are stored, and the device organizes the extracted data for use in subsequent steps.

[0075] Secondly, in step 1022, the type and content value of the key entity field and the position coordinates and contour shape of the anti-counterfeiting feature point are input into the multimodal alignment layer to generate a pairing relationship according to the spatial coordinate mapping. The specific process is that the multimodal alignment layer first establishes the position coordinate area origin for each key entity field in the upper left corner of the work order. Maximum point Separate rows into rectangular areas, such as the customer name field, and assign the upper left point lower right point At the same time, the geometric center point of the contour shape is calculated for each anti-counterfeiting feature point. The point coordinates are obtained by averaging the X and Y coordinate values ​​of its boundary points. For example, the contour boundary points 20, 20 and 40, 40 have a center point of 30, 30. Then the center point is extended to the contour boundary to generate a position coordinate area. It is assumed to be a rectangle with the upper left point 20, 20 and the lower right point 40, 40. Then, the overlapping area of ​​each field coordinate area and the anti-counterfeiting point coordinate area is calculated using the formula for overlap width: and overlap height: ,in and is the upper left coordinate of the field area, and is the lower right coordinate of the field area, similarly and is the upper left coordinate of the anti-counterfeiting point area, and is the lower right coordinate of the anti-counterfeiting point area, and the final overlapping area ,if If the overlap threshold is greater than or equal to 0.2, the pair is added to the initial pairing set; otherwise, it is skipped. After that, the many-to-one conflict is processed, and the pairing with the largest overlap area is retained. The Euclidean distance of the unpaired fields is calculated. in and is the coordinate of the center point of the field, and is the center coordinate of the anti-counterfeiting point, matches the anti-counterfeiting point with the minimum d but marks it as low confidence. Finally, eliminate all low-confidence pairings and remove such relationships to obtain the valid pairing device output result for subsequent use.

[0076] Finally, in step 1023, a corresponding fused feature vector is generated based on the pairing relationship, and then all fused feature vectors are arranged in order of the work order submission time to generate a semantic image feature stream. The specific process is that the device extracts key entity field features such as text embedding vectors and anti-counterfeiting feature point features such as image feature vectors for each valid pairing, and uses feature fusion technology such as numerical concatenation method to connect the two into a new fused feature vector. For example, a text vector sequence 0.1, 0.2 and an image vector sequence 0.3, 0.4 are generated to form a multidimensional array of length 32. Then, the device collects all fused feature vectors and arranges them in ascending order based on the timestamp of the business expansion work order submission, such as the date format string 2023-05-01, to form a feature stream sequence data set. The device stores the sequence for subsequent analysis. For example, a user work order has a pairing relationship between the customer name field and the identifier A to generate a fused vector digital sequence 0.1, 0.2 with the timestamp 2023-05-01, and the power requirement field and the identifier B to generate a fused vector digital sequence 0.3, 0.4 with the timestamp 2023-05-01. After chronological arrangement, the semantic image feature stream is represented as a sequential array stream.

[0077] In actual applications, when a business hall processes work orders and parses a work order containing an application for a power capacity increase, the system first locates the spatial coordinates of key fields in the text: the account name field occupies the coordinate area (120,150)-(220,180), and the power address field is located in a rectangular box between (120,220)-(400,250). At the same time, anti-counterfeiting features are extracted from the two uploaded attachments: the circular watermark of the user identifier scan has a center point of (180,300) and a radius of 15 pixels, and the equipment nameplate stamp occupies the area between (480,80)-(530,110). When the multimodal alignment layer established the mapping, it detected that the "Power Capacity (800kVA)" text box (120, 380)-(300, 410) overlapped with the stamp area by 65%, automatically generating a high-confidence pairing. However, the "Emergency" field, shifted to (450, 350), had a Euclidean distance of 270 pixels when associated with the nearest stamp feature point, marking it as a low-confidence pairing. During conflict resolution, it was discovered that two fields mapped to the same watermark: the system calculated that the "User ID Number" field (120, 280)-(300, 310) overlapped with the watermark by 78%, preserving this pairing. The resulting 128-dimensional fusion vectors were then sorted by work order submission time (14:30:05, 14:30:08, and 14:30:12) to form a feature stream.

[0078] In the overall solution of step 102 above, by deeply extracting the anti-counterfeiting features of the key fields of the work order text and the attached images, innovative multimodal alignment technology is used to construct a unified vector space of text entities and image features. The two are accurately associated through coordinate positioning and area coverage mechanisms to generate a fusion vector, and the vectors are arranged in time sequence to form a semantic image feature stream. The focus is on breaking through the limitations of the traditional business of separate processing of text and image information, realizing the automatic verification and matching of key anti-counterfeiting features with declared data, greatly improving the accuracy of business document authenticity identification, and effectively preventing the risk of abnormal data generation. At the same time, the entire business chain is connected in the form of standardized feature streams, providing a multimodal associated time-series data base for intelligent approval, significantly accelerating the efficiency of business expansion business flow, and ensuring the safe and efficient operation of power marketing services.

[0079] 103. Load the semantic image feature stream into a hardware acceleration chip, perform parallel matrix operations to identify the logical binding relationship between the key entity fields, and output a nested relationship topology diagram;

[0080] Optionally, the above step 103 may specifically include the following steps:

[0081] 1031. Load the semantic image feature stream into a hardware acceleration chip, and generate a logical association strength relationship matrix between each field and other fields based on the type and content value of the key entity fields;

[0082] 1032. Synchronously calculate cross-logical results of the plurality of logic association strength relationship matrices on the hardware acceleration chip, and mark strong logical connections and conflicting connections in the cross-logical results;

[0083] 1033. Using the key entity fields as nodes and the strong logical connections and conflicting connections as lines, a tree-like relationship network is constructed as a nested relationship topology diagram.

[0084] In the above steps, a hardware acceleration chip refers to an integrated circuit module dedicated to high-speed computing, such as a GPU or FPGA, which can improve computing efficiency through parallel processing. A logical binding relationship refers to the association rules between key entity fields, for example, the name and contact information of the same customer should match each other. Parallel matrix operations refer to efficient computing methods that utilize hardware acceleration features to simultaneously process multiple sets of data relationships. The logical association strength relationship matrix is ​​a two-dimensional numerical table used to quantify the closeness of the connection between different fields, ranging from 0 to 1, with higher values ​​indicating stronger binding relationships. The cross-logical result is the comprehensive relationship data obtained by superimposing and analyzing multiple matrices. A strong logical connection refers to a reliable relationship that meets the preset binding rules, such as the matching of a customer name and a user identifier. A conflicting connection refers to a contradictory relationship, such as the appearance of two different customer addresses in the same work order. A node refers to a graphical unit representing a key entity field in a nested relationship topology diagram. A link is a line segment connecting nodes, used to represent the relationship between fields. A tree relationship network is a tree-like data structure that expands layer by layer from the root node. The nested relationship topology diagram is the final visual relationship map containing hierarchical relationships.

[0085] In the embodiment of the present application, the semantic image feature stream is loaded into the hardware acceleration chip through step 1031, and a logical association strength relationship matrix between each field and other fields is generated based on the type and content value of the key entity field. The specific process is that the hardware acceleration chip receives the semantic image feature stream data stream generated in the previous step, which contains the fusion feature vector of the time series. The chip groups and processes by field type. For example, all customer name field features are grouped together, and the association strength value is generated for each group of fields through the similarity calculation function. The formula is ,in and It is the characteristic vector of different fields, with a numerical range between 0 and 1. The strength values ​​of all field combinations are then filled into a two-dimensional matrix table. The rows and columns are sorted by field name to form a complete relationship matrix. For example, when a work order contains three fields: customer name, user identification symbol, and electricity address, a 3 by 3 matrix is ​​generated, where the association strength between name and user identification symbol is 0.9, name and address is 0.3, and user identification symbol and address is 0.8. The device outputs this matrix for subsequent use.

[0086] Secondly, through step 1032, the cross-logical results of multiple logical association strength relationship matrices are synchronously calculated on the hardware acceleration chip, and strong logical connections and conflicting connections in the cross-logical results are marked. The specific process is that the hardware chip starts the parallel computing unit, processes the relationship matrices of multiple work orders at the same time, and applies preset rules to each matrix position for detection. Among them, if the association strength exceeds the strong connection threshold, such as 0.8, it is marked as a green strong logical connection. If the association strength is less than 0.3 or the same field appears in multiple conflicting groups, such as different addresses, it is marked as a red conflicting connection. At the same time, the chip merges multiple matrix results to detect continuous associations. For example, if customer A's name is associated with the same user identification symbol in multiple work orders, the binding relationship is strengthened, and finally a marked cross-logical result set is output. For example, if the association strength between the customer name and the user identification symbol in a work order is 0.95, it is marked as a strong connection. However, if two different addresses appear in the same work order, such as address 1 with a strength of 0.4 and address 2 with a strength of 0.9, address 1 is marked as a conflicting connection. The device records all marked results.

[0087] Finally, through step 1033, a tree-like relationship network is constructed as a nested relationship topology diagram using key entity fields as nodes and strong logical connections and conflicting connections as lines. The specific process is that the device reads the cross-logical results, converts each key entity field into an independent node, and uses a graphic coordinate positioning method to distribute them in the coordinate system. For example, the core field customer name is placed at the center coordinate 0, 0 level 0. According to the strong logical connection mark, the associated nodes are connected with solid lines. For example, the customer name node extends the line level 1 to the user identifier symbol node. The conflict connection mark is a dotted line. For example, the address 1 node is disconnected from the main tree. Through recursive traversal, all valid nodes are gradually expanded and connected to form a tree structure. Finally, a standard graphic file is generated and stored as a nested relationship topology diagram. For example, when constructing the topology diagram, the root node is the customer name. The first level below it is connected with the user identifier symbol node with a solid line, and the second level is connected with the electricity address node with a solid line. When address 2 is detected as a conflict, a dotted line is forked at the address node to connect address 2. The device outputs a complete tree network diagram for the business system to call.

[0088] In practical applications, within a data processing center's computational workflow, the system loads a semantic image feature stream containing a timestamp sequence into a hardware acceleration chip, immediately initiating parallel matrix operations to identify logical binding relationships between key entity fields. For example, when processing a feature stream containing 12 fields, such as power capacity (800kVA) and electricity rate type (industrial), the hardware chip first generates a logical association strength matrix. The correlation strength between the power capacity field and the industrial type (industrial) peaks at 0.87, while the correlation strength with the application's urgency level is only 0.12. Generating this 32-by-32 matrix takes 0.6 milliseconds. During the synchronous cross-calculation phase, the chip identifies two core logical bindings: a strong logical connection between the user ID field (user ID type) and the image watermark feature, achieving a cross-validation confidence level exceeding 98%. Furthermore, it detects a hardware conflict between the construction requirement's slot spacing of 3 meters and the equipment model's base specification of 2.8 meters. By processing 128 sets of matrix cross operations in parallel, a 16-layer tree-like topology network is finally constructed with the user node as the root: key paths such as the electricity address to the regional power supply capacity are marked with solid lines for strong logical connections, and conflicting nodes between construction requirements and equipment parameters are marked with dotted lines for conflict connections. The overall nested relationship topology diagram completes the hardware operation output within 1.8 milliseconds.

[0089] In the overall solution for step 103 above, a hardware acceleration chip performs parallel matrix operations on the semantic image feature stream, innovatively analyzing the logical binding relationships between key entity fields: first, a logical association strength relationship matrix is ​​constructed based on field type and content value. Then, cross-logical results between multiple matrices are calculated in parallel, and strong logical connections and conflicting connections are automatically identified. Finally, a tree-like nested relationship topology diagram is automatically constructed, with fields as nodes and logical links as connections. The core breakthrough lies in the use of hardware-level parallel computing to analyze the logical dependencies of massive fields, eliminating business logic conflict blind spots during manual review and accurately marking data contradictions. The simultaneously generated nested relationship topology diagram makes business rules explicit, providing a structured basis for quantifiable logical judgments for the intelligent approval system, dynamically tracing the integrity of the logical chain of declared data, and realizing intelligent closed-loop management of the entire process of power marketing business, from data collection to logical verification.

[0090] 104. Call the decision path sequence in the historical work order database, train the reinforcement learning agent through time series modeling, and generate a control strategy based on the state-action reward function;

[0091] Optionally, step 104 may specifically include the following steps:

[0092] 1041. Call the historical work order database, match similar work orders according to the work order type identifier, extract the decision path sequence corresponding to the work order, and split the decision path sequence into discrete state sequences and action sequences in chronological order;

[0093] 1042. Construct a long short-term memory network as a temporal modeling layer, input the state sequence, and output a probability distribution of the actions that can be executed in each state;

[0094] 1043. Based on the state-action reward function, a dynamic reward value is assigned to each action in the action sequence, a positive reward is given to compliant actions, a negative penalty is given to illegal actions, and a linear penalty is given to delayed actions according to the delay duration;

[0095] 1044. The selection probability distribution and the dynamic reward value are weighted and integrated to generate a value evaluation result for each action. The actions with the highest value evaluation results are selected and arranged in chronological order as a priority action sequence as the control strategy output for the current work order.

[0096] In the above steps, the historical work order database refers to a storage system that stores past business expansion work order processing records, including work order types, operation records, and decision results. The decision path sequence refers to a chronological sequence of operation step records when processing a work order. Time series modeling refers to a learning method for analyzing data patterns in chronological order. The reinforcement learning agent refers to a computational model that learns optimal decision strategies through environmental interaction. The state-action reward function is a rule function used to calculate reward or penalty values ​​based on the current work order status and executed operations. Compliance actions refer to operational behaviors that comply with business regulations. Illegal actions refer to behaviors that violate business rules. Delayed actions refer to operations that are not completed within the specified time limit. The dynamic reward value is a value calculated based on operational compliance. The control strategy refers to the optimal sequence of operational steps recommended for the current work order. The long short-term memory network is a neural network model that can memorize long-term time series relationships. The selection probability distribution refers to the set of probability values ​​for each operational step being selected for execution. The state sequence is a set of state descriptions at different time points during the work order processing process. The action sequence is a set of operational steps. The value assessment result is a numerical value for the comprehensive score of the operational steps.

[0097] In an embodiment of the present application, step 1041 calls the historical work order library to match similar work orders and extracts the decision path sequence to split it into a state sequence and an action sequence. The specific process is that the device retrieves the same type of work order records in the historical database according to the current work order type identifier, such as the power expansion application number 1001, and extracts the complete decision path text of each work order, such as the timestamp sequence operation review submission review materials approval, and then cuts the text in chronological order, and uses the time point status description such as materials awaiting review as the state sequence element, and the execution operation such as submitting materials as the action sequence element, and finally organizes it into a corresponding array of state sequence state A state B state C and action sequence action X action Y action Z. For example, the state sequence of a work order history record is that the work order creation review is completed, and the action sequence is that the upload attachment is approved and archived.

[0098] Secondly, through step 1042, a long short-term memory network model is constructed to input a state sequence and output an action selection probability distribution. The specific process is that the device deploys the long short-term memory network module, the network input layer receives the state sequence encoding vector, for example, the vector value is used to represent the work order creation status, the hidden layer processes the time step relationship through the memory unit, and the output layer generates a probability distribution value. The formula is that the probability value is obtained by processing the activation function, where the sum of the probability values ​​corresponding to each action is 1. For example, after the input state sequence is created, reviewed, and processed, the network outputs the action sequence probability distribution with a file upload probability of 0.7, an approval probability of 0.2, and a modification probability of 0.1. The device records the probability distribution result.

[0099] Then, in step 1043, a dynamic reward value is assigned to the action based on the reward function. The specific process is to define the state-action reward function formula:

[0100] ;

[0101] in, For compliance bonus base, such as +2, The base value for violation penalty is -3, The penalty coefficient for the delay unit is -0.5, It is the number of delay time units. For example, if an action complies with the rules, it will get a +2 reward value, and if it violates the rules, it will get -3. If it is delayed for 3 hours, it will get -1.5. Then the reward value sequence of all actions in the action sequence is calculated and the device stores the calculation results.

[0102] Finally, the control strategy is generated by integrating the probability distribution and the reward value in step 1044. The specific process is to execute the value evaluation formula for each action. ,in is the selection probability, For example, if the probability of an action is 0.7 and the reward value is +2, the value assessment value is 1.4. All action values ​​are compared to select the action with the highest value. Finally, the actions with the highest value are arranged in chronological order to form a control strategy sequence output. For example, a priority action sequence is generated and submitted for approval and archived for storage as the current work order processing strategy.

[0103] In practical applications, a certain intelligent decision-making system, upon receiving the nested relationship topology of a current work order, immediately accesses a database of 3,562 similar work orders stored over the past three months. For example, when processing an industrial power capacity increase work order, the system matched 217 similar work orders and extracted the decision path sequence within each work order's approval process: the 42 operational steps of a reference work order were split into a state sequence and an action sequence by time. A long-short-term memory network (LSTM) was then constructed with 128 hidden units. When the current work order state was input, the output action probability distribution was based on the power supply capacity overrun state. The network recommended increasing the weight of power supply level operations to 67%, while routine review actions were only weighted at 23%. Furthermore, based on a dynamic reward mechanism, the system awarded a +2 reward for compliant actions, such as passing safety regulations acceptance, and deducted a 5-point penalty for non-compliant actions, such as skipping the insulation test step. Delayed actions, such as those not processed within three days, were linearly deducted at a rate of 0.1 point per minute. During the value assessment phase, when it is detected that the value of the alternative action of skipping environmental assessment has dropped to -3.2, the system automatically filters and generates a priority action sequence: the first priority triggers the dual-circuit power supply solution generation action 5 minutes after receiving the work order, and the second priority performs the equipment load-bearing verification action at the 18th minute. Finally, a control strategy containing 7 core steps is output for the execution engine to call.

[0104] In the overall solution of step 104 above, time series modeling training is performed by calling on massive decision path sequence data from the historical work order database: first, the state-action sequence chain of similar work orders is extracted, and a long short-term memory network is used to accurately fit the dynamic behavior patterns in the business scenario. Then, an innovative three-level state-action reward function is designed, which includes positive compliance incentives, negative violation penalties, and delayed linear deductions. By integrating the value assessment of historical actions with weighted rewards, the intelligent agent can autonomously generate a prioritized action sequence as a work order control strategy. Its core technical value lies in the deep integration of business rules and historical experience data, realizing intelligent deduction and self-optimization of the entire approval decision process: it not only accurately quantifies the compliance and timeliness benefits of each operation action, but also dynamically avoids experience blind spots and rule execution deviations in manual approval, thereby achieving a dual improvement in risk control intensity and business flow efficiency in the power industry expansion scenario, and building an intelligent decision-making center with continuous evolution capabilities for power marketing services.

[0105] 105. Based on the nested relationship topology diagram and the control strategy, verify the number of missing required fields and detect key entity logic conflicts, compare the similarity between the anti-counterfeiting feature points and the pre-stored reference template, and output a verification result including missing field identifiers and anti-counterfeiting abnormal areas;

[0106] Optionally, the above step 105 may specifically include the following steps:

[0107] 1051. Traverse all entity nodes in the nested relationship topology graph, scan predefined required field labels, count the number of missing fields, and generate a missing field identifier list;

[0108] 1052. Call the priority action sequence in the control strategy, locate the node pairs with conflicting connections in the nested relationship topology graph, and mark the coordinates and conflict types of the node pairs;

[0109] 1053. Superimpose the anti-counterfeiting feature point with the pre-stored reference template at the pixel level, calculate the contour overlap as a similarity quantization value, and identify the coordinates of the contour difference area;

[0110] 1054. Integrate the missing field identifier list, the coordinates and conflict types of the node pairs, the similarity quantization values, and the coordinates of the contour difference areas, and output a verification result including the missing field identifiers and the anti-counterfeiting abnormality areas.

[0111] In the above steps, the nested relationship topology diagram is the tree-like relationship network diagram generated in the previous step, which contains key entity field nodes and their association relationships. The control strategy is the recommended sequence of operation steps for the current work order. The number of missing required fields refers to the number of fields that must be filled in but are not filled in the inspection work order. The key entity logical conflict refers to the contradictory association relationship between fields, such as inconsistent address information. The anti-counterfeiting feature point is the identification feature in the attached image used to verify the authenticity. The pre-stored reference template is the standard anti-counterfeiting image pre-stored by the system. The similarity quantization value calculates the degree of matching between the two images through numerical calculations. The missing field identification list records the list of missing field names. The anti-counterfeiting abnormal area is the coordinate of the contour area in the image that does not match the template. The verification result is a comprehensive report containing all inspection conclusions.

[0112] In an embodiment of the present application, step 1051 is used to traverse the nested relationship topology graph to count the number of missing required fields and generate an identification list. The specific process is that the device reads all entity nodes in the topology graph, such as customer name, user identification symbol, etc., compares the predefined required field label list, such as the items required to be filled in the standard form, detects the missing required fields in the node and counts them, and forms a text list of the missing field names, such as emergency contacts. For example, the work order contains 10 required fields, and 8 nodes in the topology graph detect missing electricity type and emergency contact fields, count 2, and generate a missing field identification list. The device stores the list.

[0113] Secondly, the control strategy is called through step 1052 to locate the conflicting nodes and mark the coordinates and types. The specific process is to trigger the conflict detection module to scan the topology map according to the priority action sequence of the control strategy, find all node pairs marked as conflicting connections, such as address 1 node and address 2 node, extract the plane coordinates of the nodes in the map, such as address 1 coordinate (20, 30) address 2 coordinate (50, 60), record the conflict type, such as address contradiction or information duplication, for example, if it is detected that the electricity type node in a work order has a double definition, the conflict type is recorded as repeated entry of node coordinates (10, 40) and (15, 45), and the device packages the coordinates and conflict type for record.

[0114] Next, the similarity between the anti-counterfeiting feature points and the reference template is calculated in step 1053 and the difference area is identified. The specific process is to overlay the anti-counterfeiting feature point image of the current work order on the pre-stored reference template and use image processing technology to calculate the contour overlap. The similarity formula is defined as the ratio of the overlapping pixel area to the total area of ​​the template. in Indicates the number of overlapping pixels, Represents the total number of pixels in the reference template. The coordinates of the difference area are generated by comparing the contour boundaries. For example, the upper left point of the mismatched area is (100, 200) and the lower right point is (150, 250). For example, if the total number of template pixels is 10,000 and the overlapping pixels are 8,000, the similarity is 80%. The coordinates of the difference area are (120, 180) to (130, 190). The device records the values ​​and coordinates.

[0115] Finally, all inspection results are integrated through step 1054 to output a verification report. The specific process is to merge the missing field identification list, such as the field list emergency contact electricity type, conflict node coordinates and types, such as address conflict coordinates (10,40) (15,45) and the similarity quantization value 80% and the difference area coordinates (120,180) to (130,190), and organize them into a structured data set to generate a verification result document for the business system to call. For example, the final output text content includes the required field missing identification emergency contact electricity type, anti-counterfeiting abnormal area coordinates X1=120Y1=180X2=130Y2=190, and the system automatically displays or stores the result.

[0116] In actual application, during the operation of a work order verification platform, the system performed verification operations based on the nested topology diagram and control strategy of industrial power capacity expansion work orders. While traversing the topology nodes, required fields were detected to be missing: the power supply capability certificate was missing, identified as FF005, and the emergency contact number was missing, identified as FF009, for a total of two required fields. The anti-counterfeiting verification process was simultaneously initiated, overlaying the pixel-by-pixel feature points of the current work order's document watermark with the reference template. The circular watermark's contour overlap in the coordinate area (230, 110)-(260, 140) was detected to be 65%, a 12.5% ​​difference from the reference value of 85%. The system identified the abnormal area (240, 130)-(250, 150) and marked it as AF001 anti-counterfeiting anomaly. The control strategy's seventh priority action was also invoked to locate a logic conflict: a dual-core graphics memory conflict existed between the equipment parameter node and the construction standard node. The system marked a C-shaped conflict line between the topology coordinates (120, 380) and (450, 80). The final verification result is integrated and output: the missing field identification list contains 2 red alert field numbers, the anti-counterfeiting abnormal area locks the 3×20 pixel rectangular frame coordinates, and the logical conflict coordinate pair is attached with a third-level emergency level identification. The entire verification process is completed within 28 milliseconds.

[0117] In the overall solution of step 105 above, a collaborative verification mechanism combining nested relationship topology maps and control strategies is used to innovatively implement multi-dimensional business logic verification: Based on topology map nodes, the system intelligently identifies missing required fields and precisely locates logical conflict points. Combined with a pre-stored benchmark template, the intelligent image comparison system accurately quantifies the similarity of anti-counterfeiting feature points. The output includes both structured field missing identifiers and visual coordinates of anti-counterfeiting anomaly areas, forming a three-in-one automated verification closed loop. The core technological breakthrough lies in the integration of the business rule logic tree and the intelligent decision-making action chain. This dual guarantee enables simultaneous and precise processing of required field verification, logical conflict detection, and anti-counterfeiting feature verification, significantly reducing the risk of manual verification blind spots. The output, locatable verification results, can directly drive the business system to automatically block abnormal work order processes or trigger manual review tasks, enabling a qualitative breakthrough in the dual dimensions of risk control and process efficiency improvement for the power marketing business, and building a high-precision, self-driven, and full-link intelligent quality inspection system for business expansion work orders.

[0118] 106. Based on the verification result, a judgment operation is triggered. If the field is complete and the anti-counterfeiting similarity is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system. If the field is missing or the anti-counterfeiting similarity is less than the preset threshold, the blocking instruction is executed to freeze the work order flow and trigger an alarm signal to the business hall terminal.

[0119] Optionally, step 106 may specifically include the following steps:

[0120] 1061. Read the missing quantity and the contour similarity quantization value in the verification result to trigger a determination operation;

[0121] 1062. If the missing quantity is zero and the contour similarity quantization value is greater than or equal to the preset threshold, executing the work order release instruction and pushing the work order to the power business execution system;

[0122] 1063. If the number of missing items is greater than zero or the quantized value of the contour similarity is less than a preset threshold, a blocking instruction is executed to freeze the write permission of the work order database, and based on the position coordinates of the key entity fields corresponding to the number of missing items in the nested relationship topology diagram, combined with the pixel coordinates of the contour difference area in the anti-counterfeiting feature map, an alarm message is generated and sent to the business hall terminal.

[0123] In the above steps, the verification result refers to the comprehensive inspection report generated in step 105, which includes data such as the number of missing fields and image similarity. The missing number refers to the numeric value of the required fields that are not filled in the verification result. The contour similarity quantification value refers to the numerical result of the degree of matching between the anti-counterfeiting feature map and the template, such as a percentage. The preset threshold is the minimum standard value pre-set by the system, such as 0.8 represents 80% matching. The work order release instruction refers to the instruction to approve the work order. The power business execution system refers to the operating platform for processing subsequent business processes of the work order. The blocking instruction refers to the command to prevent the execution of the work order. Freezing the work order flow refers to locking the write operation permission of the work order database to suspend the process. The alarm signal refers to the alarm information notifying the operator. The nested relationship topology diagram is the tree relationship diagram generated in step 103, which includes the coordinate positions of key fields. The contour difference area is the image mismatch coordinate range identified in step 105.

[0124] In an embodiment of the present application, step 1061 is used to read the number of missing fields and the quantized value of the contour similarity in the verification result and trigger a judgment operation. The specific process is that the device extracts digital values ​​from the verification result document, including the total number of missing fields and the image similarity percentage, and then starts the judgment process to determine the next operation according to the preset rules. For example, the verification result report of a work order shows that the number of missing fields is 0, and the quantized value of the contour similarity is 0.95, which means 95% match. After reading, the device triggers the judgment mechanism to prepare to execute subsequent steps.

[0125] Then, step 1062 is used to check whether the missing quantity is zero and whether the quantized value of the profile similarity is greater than or equal to the preset threshold. If the conditions are met, the work order release instruction is executed and the work order is pushed. The specific process is that the device compares the read values. When the missing quantity is equal to 0 and the quantized value of the profile similarity is greater than or equal to the preset threshold, such as 0.8, a work order release instruction is generated. The instruction contains the unique identifier of the work order. The device pushes the work order data to the power business execution system for subsequent processing. For example, if the missing quantity of a work order is 0 and the profile similarity is 0.9, the preset threshold is 0.8, the device determines that the conditions are met and executes the release, and pushes it to the system to automatically start the power expansion arrangement.

[0126] Finally, when it is detected in step 1063 that the number of missing items is greater than zero or the quantized value of the contour similarity is less than the preset threshold, a blocking instruction is executed to freeze the work order and generate an alarm signal. The specific process is that if the device determines that the conditions are not met, a blocking instruction is generated to lock the write permission of the work order database to prevent data modification. At the same time, the missing field identifiers in the verification results are integrated, such as the field name and its coordinate position in the nested relationship topology map, such as X coordinate 10Y coordinate 20, and the starting coordinates and ending coordinates of the contour difference area, such as X1 coordinate 100Y1 coordinate 200 and X2 coordinate 150Y2 coordinate 250. This information is combined to generate a detailed alarm message text and send it to the business hall terminal for display or sound prompt. For example, a work order with a missing quantity of 2 contains the field emergency contact at coordinates 20, 30 in the topology map, and the contour difference coordinates start at 100, 200 and end at 150, 250. The device freezes the work order and sends an alarm to the terminal, such as the screen displays a prompt of the coordinates of the missing anti-counterfeiting abnormal area of ​​the emergency contact.

[0127] In actual applications, when determining the compliance of work orders at the business hall terminal, the system detects that the industrial capacity expansion work order is missing two required fields: power supply capacity certificate and emergency number, and the user identifier watermark has a contour overlap of only 65% ​​in the coordinate area (240,130)-(250,150), which is lower than the 85% threshold. The blocking mechanism is immediately triggered: the database write permission of the work order ID 2023-0765 is frozen within 0.3 seconds, and a high-priority alarm is sent to the terminal simultaneously - the missing field is located at the topology map coordinate (360, At the same time, a blue zebra line frame is superimposed on the document image to mark the difference area coordinates (240,130)-(250,150), and the terminal buzzer sounds immediately and generates a correction list with a deadline of 8 hours; on the contrary, when the resident electricity meter work order verification shows zero missing fields and the overlap of the steel stamp area (80,60)-(105,85) reaches 92%, the system pushes the work order ID 2023-0766 to the power execution business system through TCP443 port within 15 milliseconds, automatically triggering the meter installation process queue.

[0128] In the overall solution of step 106 above, by judging the similarity value between the field integrity index and the anti-counterfeiting feature, the work order release push or blocking freeze operation is intelligently executed. When it is detected that the field is zero missing and the anti-counterfeiting feature meets the preset security threshold, the work order is automatically pushed to the power business execution system. When a field is missing or an anti-counterfeiting anomaly is identified, the work order flow is immediately frozen, and a visual alarm message is generated based on the coordinates of the nested relationship topology map and the coordinates of the anti-counterfeiting difference area to accurately locate the problem point and push it to the business hall terminal. The core value lies in opening up the control link between the intelligent verification results and the business execution system, forming a millisecond-level blocking response capability for risky work orders and a non-sensing release mechanism for compliant work orders. At the same time, coordinate-level alarm information is used to guide manual precise review, achieving a breakthrough improvement in the security control intensity and business processing efficiency in the power business expansion process, and building a self-driven business assurance system covering the entire link of risk perception, intelligent decision-making, and precise execution.

[0129] The following is a specific embodiment of steps 101 to 106:

[0130] In an operational example at a power business data processing center, after receiving a user-submitted industrial power capacity increase work order at a business hall terminal, the system automatically parsed 15 structured fields within the work order text, including the user name and electricity address, and simultaneously extracted the circular anti-counterfeiting watermark features of the user identifier in the attached image. Using a multimodal alignment layer, the 800kVA text entity was spatially mapped to the watermark's center coordinates, 150 pixels on the horizontal axis and 400 pixels on the vertical axis. This generated a 128-dimensional semantic image feature stream sorted by the 14:30:05 timestamp. This feature stream was loaded into a hardware acceleration chip for parallel computation, outputting a 32x32 association matrix in 0.9 milliseconds. This identified the key logical conflict between the construction requirement slot spacing of 3 meters and the equipment parameter base spacing of 2.8 meters, and constructed a 16-layer tree topology diagram containing conflicting links between equipment parameter nodes and construction standard nodes. The system then retrieved decision paths from 217 similar work orders from a historical database and, through a reinforcement learning agent, generated a prioritized action sequence within the state sequence. Compliance actions, such as safety acceptance, were assigned a positive 2-point reward, triggering the generation of a dual-circuit power supply solution. During the verification phase, if it is detected that the power supply capability certificate field is missing and the user identifier watermark has an outline overlap of 65% in the coordinate area of ​​240 to 250 on the horizontal axis and 130 to 150 on the vertical axis, which is lower than the preset threshold of 85%, the work order flow will be frozen immediately, and an alarm will be sent to the terminal regarding the missing field positioning coordinates of 360 on the horizontal axis and 220 on the vertical axis in the topology map, as well as a blue frame coordinate marking of the difference area in the document image; and if zero missing fields are detected for the residential meter work order and the overlap of the steel stamp area of ​​80 to 105 on the horizontal axis and 60 to 85 on the vertical axis is 92%, it will be pushed to the business execution system through the TCP443 port in milliseconds.

[0131] Figure 2 A schematic diagram of a power business data analysis and processing system is provided for an embodiment of the present application. Figure 2As shown, the system includes:

[0132] The receiving module 21 is used to receive a business expansion installation work order submitted by a user through a business hall terminal device, wherein the business expansion installation work order includes a structured work order text and an unstructured attachment image;

[0133] A mapping module 22 is used to parse the key entity fields in the work order text, simultaneously extract the anti-counterfeiting feature points of the attachment image, and map the text entities and image features into a unified vector space through a multimodal alignment layer to generate a semantic image feature stream;

[0134] Identification module 23, used to load the semantic image feature stream into the hardware acceleration chip, perform parallel matrix operations to identify the logical binding relationship between the key entity fields, and output a nested relationship topology diagram;

[0135] A generation module 24 is used to call the decision path sequence in the historical work order database, train the reinforcement learning agent through time series modeling, and generate a control strategy based on the state-action reward function;

[0136] Verification module 25, for verifying the number of missing required fields and detecting key entity logic conflicts based on the nested relationship topology and the control strategy, comparing the similarity between the anti-counterfeiting feature points and the pre-stored reference template, and outputting a verification result including missing field identifiers and anti-counterfeiting abnormal areas;

[0137] The judgment module 26 is used to trigger a judgment operation based on the verification result. If the field is complete and the anti-counterfeiting similarity is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system. If the field is missing or the anti-counterfeiting similarity is less than the preset threshold, the blocking instruction is executed to freeze the work order flow and trigger an alarm signal to the business hall terminal.

[0138] Figure 2 The power business data analysis and processing system can perform Figure 1 The implementation principles and technical effects of the power business data analysis and processing method described in the illustrated embodiment will not be elaborated on here. The specific manner in which each module and unit performs operations in the power business data analysis and processing system in the above embodiment has been described in detail in the embodiment of the method and will not be elaborated on here.

[0139] In one possible design, Figure 2 The power business data analysis and processing system of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0140] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0141] The processing component 32 is used for the above Figure 1 The power business data analysis and processing method of the embodiment.

[0142] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0143] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0144] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0145] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0146] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0147] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0148] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 A method for analyzing and processing electric power business data in the illustrated embodiment.

[0149] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0150] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0151] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer or server) to execute the methods described in each embodiment or certain portions of the embodiments.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for analyzing and processing power business data, characterized in that: include: Receive a business expansion application form submitted by a user through a business hall terminal device, wherein the business expansion application form includes a structured work form text and an unstructured attachment image; Parse the key entity fields in the work order text, extract the anti-counterfeiting feature points of the attachment image simultaneously, and map the text entities and image features into a unified vector space through a multimodal alignment layer to generate a semantic image feature stream, including: Scan the work order text to locate the type and content value of the key entity field, scan the attachment image, and extract the position coordinates and contour shape of the anti-counterfeiting feature point; input the type and content value of the key entity field and the position coordinates and contour shape of the anti-counterfeiting feature point into the multimodal alignment layer, and generate a pairing relationship based on spatial coordinate mapping; based on the pairing relationship, generate corresponding fused feature vectors respectively, arrange all the fused feature vectors in the order of submission time of the business expansion work order, and generate a semantic image feature stream; Loading the semantic image feature stream into a hardware acceleration chip, performing parallel matrix operations to identify the logical binding relationship between the key entity fields, and outputting a nested relationship topology diagram. The nested relationship topology diagram refers to a tree-like relationship network constructed with the key entity fields as nodes and strong logical connections and conflicting connections as lines. The strong logical connection refers to a reliable relationship that meets the preset binding rules, and the conflicting connection refers to a contradictory association. Recall the decision path sequence in the historical work order database. The decision path sequence refers to the chronological sequence of operation step records when processing the work order. The decision path sequence is split into a discrete state sequence and action sequence in chronological order. A long short-term memory network is constructed as the temporal modeling layer, which inputs the state sequence and outputs the selection probability distribution of executable actions in each state. Based on the state-action reward function, a dynamic reward value is assigned to each action in the action sequence. The reinforcement learning agent is trained through temporal modeling, and a control policy is generated based on the state-action reward function. The control policy refers to the optimal sequence of operation steps recommended for the current work order. According to the nested relationship topology diagram and the control strategy, the number of missing required fields is verified and key entity logic conflicts are detected. The similarity between the anti-counterfeiting feature points and the pre-stored reference template is compared, and a verification result including missing field identifiers and anti-counterfeiting abnormal areas is output; A judgment operation is triggered based on the verification result. If the field is complete and the anti-counterfeiting similarity is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system. If the field is missing or the anti-counterfeiting similarity is less than the preset threshold, the blocking instruction is executed, the work order flow is frozen and an alarm signal is triggered to the business hall terminal.

2. The method according to claim 1, characterized in that Inputting the type and content value of the key entity field and the position coordinates and contour shape of the anti-counterfeiting feature point into the multimodal alignment layer, and generating a pairing relationship according to spatial coordinate mapping, including: Establish the position coordinates of the key entity fields, take the upper left corner of the work order as the origin and the lower right corner as the maximum coordinate point, divide the text area by line, and assign position coordinates to each key entity field; Extracting the contour shape of the anti-counterfeiting feature point, calculating the geometric center point of the contour shape, extending from the center point to the contour boundary, and generating position coordinates; If the overlapping area of ​​the position coordinate region of the key entity field and the position coordinate region of the anti-counterfeiting feature point is greater than or equal to a minimum overlapping threshold, an initial pairing relationship set is generated; For many-to-one conflicts, the pairing relationship with the largest overlapping area is retained. The unpaired key entity fields are matched to the anti-counterfeiting feature point with the closest Euclidean distance to the center point, marked as a low-confidence pairing, and the low-confidence pairing is eliminated in the initial pairing relationship set to obtain a valid pairing relationship.

3. The method according to claim 1, characterized in that Call the decision path sequence in the historical work order database, train the reinforcement learning agent through time series modeling, and generate a control strategy based on the state-action reward function, including: Call the historical work order database, match similar work orders based on the work order type identifier, extract the decision path sequence corresponding to the work order, assign positive rewards to compliant actions, assign negative penalties to non-compliant actions, and linearly penalize delayed actions based on the delay duration; The selection probability distribution is weightedly integrated with the dynamic reward value to generate a value evaluation result for each action. The actions with the highest value evaluation results are screened and arranged in chronological order as a priority action sequence, which is output as the control strategy for the current work order.

4. The method according to claim 1, wherein Based on the nested relationship topology and the control strategy, the number of missing required fields is verified and key entity logic conflicts are detected. The similarity between the anti-counterfeiting feature points and the pre-stored reference template is compared, and the verification result including the missing field identifier and the anti-counterfeiting abnormal area is output, including: Traversing all entity nodes in the nested relationship topology graph, scanning predefined mandatory field labels, counting the number of missing fields and generating a missing field identifier list; Calling the priority action sequence in the control strategy, locating the node pairs pointed to by conflicting connections in the nested relationship topology graph, and marking the coordinates and conflict types of the node pairs; Superimposing the anti-counterfeiting feature points with the pre-stored reference template at the pixel level, calculating the contour overlap as a similarity quantification value, and marking the coordinates of the contour difference area; The missing field identification list, the coordinates and conflict types of the node pairs, the similarity quantization values ​​and the coordinates of the contour difference areas are integrated, and a verification result including the missing field identification and the anti-counterfeiting abnormal area is output.

5. The method according to claim 1, wherein The semantic image feature stream is loaded into the hardware acceleration chip, parallel matrix operations are performed to identify the logical binding relationship between the key entity fields, and a nested relationship topology diagram is output, including: Loading the semantic image feature stream into a hardware acceleration chip, and generating a logical association strength relationship matrix between each field and other fields based on the type and content value of the key entity fields; Synchronously calculating cross-logic results of a plurality of the logic association strength relationship matrices on the hardware acceleration chip, and marking strong logical connections and conflicting connections in the cross-logic results; With the key entity fields as nodes and the strong logical connections and conflicting connections as lines, a tree-like relationship network is constructed as a nested relationship topology diagram.

6. The method according to claim 4, characterized in that Based on the verification result, a judgment operation is triggered. If the fields are complete and the anti-counterfeiting similarity is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system. If the fields are missing or the anti-counterfeiting similarity is less than the preset threshold, the blocking instruction is executed to freeze the work order flow and trigger an alarm signal to the business hall terminal, including: Reading the missing quantity and the contour similarity quantization value in the verification result to trigger a determination operation; If the missing number is zero and the contour similarity quantization value is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system; If the number of missing items is greater than zero or the contour similarity quantization value is less than the preset threshold, a blocking instruction is executed to freeze the write permission of the work order database, and based on the position coordinates of the key entity fields corresponding to the missing number in the nested relationship topology diagram, combined with the pixel coordinates of the contour difference area in the anti-counterfeiting feature map, an alarm message is generated and sent to the business hall terminal.

7. A power business data analysis and processing system, configured to execute the power business data analysis and processing method according to any one of claims 1 to 6, characterized in that: include: A receiving module, configured to receive a business expansion application form submitted by a user through a business hall terminal device, wherein the business expansion application form includes a structured work form text and an unstructured attachment image; A mapping module is used to parse the key entity fields in the work order text, simultaneously extract the anti-counterfeiting feature points of the attached image, and map the text entities and image features into a unified vector space through a multimodal alignment layer to generate a semantic image feature stream; A recognition module is used to load the semantic image feature stream into a hardware acceleration chip, perform parallel matrix operations to identify the logical binding relationship between the key entity fields, and output a nested relationship topology diagram; The generation module is used to call the decision path sequence in the historical work order database, train the reinforcement learning agent through time series modeling, and generate the control strategy based on the state-action reward function; A verification module is used to verify the number of missing required fields and detect key entity logic conflicts based on the nested relationship topology diagram and the control strategy, compare the similarity of the anti-counterfeiting feature points with the pre-stored reference template, and output a verification result including missing field identifiers and anti-counterfeiting abnormal areas; The judgment module is used to trigger a judgment operation based on the verification result. If the field is complete and the anti-counterfeiting similarity is greater than or equal to the preset threshold, the work order release instruction is executed and the work order is pushed to the power business execution system. If the field is missing or the anti-counterfeiting similarity is less than the preset threshold, the blocking instruction is executed, the work order flow is frozen and an alarm signal is triggered to the business hall terminal.

8. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an electric power business data analysis and processing method as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the method for analyzing and processing electric power business data according to any one of claims 1 to 6 is implemented.

Images