Method for generating a security control file based on a picture recognition model training technique

Abstract

The application discloses a method for generating a safe control file based on a picture recognition model training technology, a machine vision system is built, a picture recognition model suitable for a power scene is selected through a multi-stage verification process; an intelligent agent is developed, three collaborative intelligent agents are constructed based on a power industry special large model, a picture recognition model and an optical character recognition technology, picture information extraction, text recognition, information completion, document generation and compliance audit are realized; picture collection and model training; intelligent service interface empowerment; display and output. The application has the beneficial effects that the problems of low efficiency and repeated input of manual input are solved, the labor cost is saved by about 50%, the staff is relieved from tedious information input and checking work, focuses on core business and improves the smoothness of management process; with the help of factory station name standardization completion, intelligent audit and other functions, the standard of filling is unified, and the phenomenon that the standardization of work tickets is uneven due to the operation habits and understanding differences of personnel is reduced.

Description

Technical Field

[0001] This invention relates to the field of power industry safety management document processing technology, specifically a method for generating safety management documents based on image recognition model training technology. Background Technology

[0002] Traditional power marketing and metering verification work orders and related documents rely on manual processing, severely hindering the efficiency and quality of operations. The work model for high-frequency operations suffers from numerous problems: on-site survey documents require staff to manually input extensive details such as plant / station names, room names, locations, and equipment names after the site visit; important information needs to be repeatedly entered in multiple different business scenarios, including work orders, survey records, and daily shift meetings, resulting in cumbersome information transmission; and differences in staff work habits, information delivery, and semantic understanding make it difficult to fully standardize the filling process.

[0003] The cumbersome and repetitive nature of traditional work processes significantly increases the burden on staff, and operational safety risks arising from document issues are commonplace. This not only threatens the safe and orderly conduct of power operations but also impacts overall production and operational efficiency and stability. Therefore, integrating technologies such as artificial intelligence into business processes to build smarter and more precise solutions is a core measure to address current production and operational pain points, ensure efficient and high-quality business operations, and is an inevitable choice for strengthening power operation safety, reducing costs, increasing efficiency, and achieving digital transformation for enterprises.

[0004] Therefore, it is necessary to propose a method for generating security management documents based on image recognition model training technology to address the aforementioned technical issues. Summary of the Invention

[0005] The purpose of this invention is to provide a method for generating secure control documents based on image recognition model training technology, thereby solving the aforementioned technical problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for generating security management files based on image recognition model training technology includes the following steps: S1. Machine vision system construction: Select an image recognition model suitable for the power scenario through a multi-stage verification process; S2. Image acquisition and model training: Collect and label power field image samples, and train the model to improve the ability to identify key elements. S3, Intelligent Agent Development: Based on the power industry-specific large model, image recognition model and optical character recognition (OCR) technology, a three-collaborative intelligent agent is built to realize image information extraction, text recognition, information completion, document generation and compliance review; S4, empowered by intelligent service interfaces, enables intelligent processing of various business scenarios through standardized service interfaces; S5. Display and Output: Performs logical verification and rule matching on the processed data, presents it in a standardized format, and supports downloading.

[0007] The multi-stage verification process described in step S1 includes: S11. Use 5 sets of images for rapid verification and eliminate models with an accuracy rate of less than 90%; S12. Use 20 sets of images for full verification, identify weaknesses through confusion matrix analysis and give a comprehensive score; S13. Conduct scene adaptation tests on the two models with the highest scores and select the model that balances performance, efficiency and scene compatibility.

[0008] The image extraction and recognition algorithm of the image recognition model in step S3 includes: S21. Image preprocessing: Gaussian filtering is used for noise reduction, Hough transform correction, and adaptive histogram equalization to enhance contrast. S22. Region of Interest (ROI) Extraction: Initial localization is achieved based on HSV color model segmentation and morphological operations, followed by precise cropping using bounding box regression algorithm; S23. Feature Extraction: Multi-dimensional image features are extracted using an improved CNN model combined with an attention mechanism; S24. Target Recognition and Classification: Input the fully connected layer and classify it using a customized classification label library for the power scenario. Mark suspicious areas for low-confidence results and trigger secondary verification.

[0009] In image preprocessing, Hough transform is used to achieve geometric correction of tilted images by detecting straight edges such as device borders and sign outlines in the image; the adaptive histogram equalization algorithm is used to improve the clarity of image details in low-light and strong light reflection scenes.

[0010] In feature extraction, the improved CNN model captures features of different dimensions such as device texture and sign text outline through multi-scale convolutional kernels, and the attention mechanism is used to strengthen the weight of key features and suppress interference from irrelevant features.

[0011] The OCR acquisition algorithm in step S3 includes: S31. Text region detection: The CTPN algorithm is used to extract candidate boxes and merge and filter them; Text line segmentation: Text lines are segmented by projection method and tilted text is corrected a second time. S32. Character segmentation: Separating characters based on connected component analysis or adaptive threshold segmentation; Character recognition: Recognizing special characters using an LSTM+CTC model combined with a customized character set for the power industry. S33. Post-processing of recognition results: Spelling correction and standardization based on power industry dictionary.

[0012] The algorithm steps of the power industry-specific large model described in step S3 include: S41. Transform power industry standard documents and compliance verification criteria into a structured rule library that meets the requirements for outputting verification conditions for applicable rule ID scenarios. S42. Clean, standardize, and associate OCR text data and image recognition scene information with power industry entities; S43. Use a Transformer-based semantic matching algorithm to locate the target rule set that is suitable for the business scenario; S44. Complete missing fields based on pre-trained parameters and verify the logical consistency of multi-source data; integrate and output rule matching results, completion information and logical verification conclusions.

[0013] In step S3, the intelligent agent collaboratively revises the document content using three technologies, generates standardized documents based on power industry rules, strengthens business review rules, and simultaneously handles document data conflicts and format deviations.

[0014] The standardized service interfaces mentioned in step S4 include image text extraction service, plant and station information intelligent matching service, generated content manual revision service, on-site survey document generation service, intelligent invoicing service, intelligent invoice review service, daily station and shift meeting document generation service, and document template management service.

[0015] The display and output described in step S5 supports interactive display on PC and mobile devices. The output file format is Word, which includes power industry safety management documents such as on-site survey documents, work order documents, and daily shift meeting documents.

[0016] The present invention has the following beneficial effects: 1. It solves the problems of low efficiency and repetitive data entry in manual data entry, saving about 50% of labor costs, freeing staff from tedious data entry and verification work, allowing them to focus on core business and improve the smoothness of management processes; 2. By using functions such as standardized completion of plant and station names and intelligent review, the filling standards are unified, reducing the inconsistency in the standardization of work tickets caused by differences in personnel operating habits and understanding, reducing the error rate by 35% and reducing management difficulty; 3. Improve the accuracy of work orders and related documents, avoid operational safety risks caused by errors or omissions, reduce risks by about 10%, effectively avoid economic losses caused by safety accidents, and ensure the smooth progress of power operations; 4. Achieve digital and intelligent upgrades in the processing of work orders and related documents, providing a model for other units in the industry to learn from, and promoting the overall development of the power industry towards intelligence and efficiency. Attached Figure Description

[0017] Fig. 1 This is a flowchart of the method for the invention; Fig. 2 This is a schematic diagram of the machine vision system of the present invention; Fig. 3 This is a schematic diagram of the method flow of the present invention.

[0018] The diagram includes the following labels: 1. Resource Layer; 2. Model Layer; 3. Artificial Intelligence Platform Layer; 4. Service Layer; 5. Application Layer; 6. Presentation Layer; 7. Server; 8. Database; 9. Computing Power; 10. Network; 11. Security Measures; 12. Guangming Big Data Model; 13. Image Recognition Model; 14. Security Control Document Generation and Review Intelligent Agent; 15. Plant / Station Information Intelligent Matching Service; 16. Intelligent Invoicing Service; 17. Daily Station / Shift Meeting Document Generation Service; 18. On-site Survey Document Generation Service; 19. Image Text Extraction Service; 20. Intelligent Ticket Review Service; 21. Manual Revision Service for Generated Content; 22. Document Template Management Service; 23. On-site Survey Document Generation; 24. Security Control Document Template Management; 25. Work Order Document Generation; 26. Plant / Station Information Management; 27. Daily Station / Shift Meeting Document Generation; 28. PC Terminal; 29. ​​Mobile Terminal. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0020] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0021] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0022] This embodiment provides a technical solution: like Figs. 1 to 3 As shown, the method for generating security management documents based on image recognition model training technology includes the following steps: S1. Machine vision system construction: Select an image recognition model suitable for the power scenario through a multi-stage verification process; S2. Image acquisition and model training: Collect and label power field image samples, and train the model to improve the ability to identify key elements. S3, Intelligent Agent Development: Based on the power industry-specific large model, image recognition model and optical character recognition (OCR) technology, a three-collaborative intelligent agent is built to realize image information extraction, text recognition, information completion, document generation and compliance review; S4, empowered by intelligent service interfaces, enables intelligent processing of various business scenarios through standardized service interfaces; S5. Display and Output: Performs logical verification and rule matching on the processed data, presents it in a standardized format, and supports downloading.

[0023] The multi-stage verification process described in step S1 includes: S11. Use 5 sets of images for rapid verification and eliminate models with an accuracy rate of less than 90%; S12. Use 20 sets of images for full verification, identify weaknesses through confusion matrix analysis and give a comprehensive score; S13. Conduct scene adaptation tests on the two models with the highest scores and select the model that balances performance, efficiency and scene compatibility.

[0024] The image extraction and recognition algorithm of the image recognition model in step S3 includes: S21. Image preprocessing: Gaussian filtering is used for noise reduction, Hough transform correction, and adaptive histogram equalization to enhance contrast. S22. Region of Interest (ROI) Extraction: Initial localization is achieved based on HSV color model segmentation and morphological operations, followed by precise cropping using bounding box regression algorithm; S23. Feature Extraction: Multi-dimensional image features are extracted using an improved CNN model combined with an attention mechanism; S24. Target Recognition and Classification: Input the fully connected layer and classify it using a customized classification label library for the power scenario. Mark suspicious areas for low-confidence results and trigger secondary verification.

[0025] In image preprocessing, Hough transform is used to achieve geometric correction of tilted images by detecting straight edges such as device borders and sign outlines in the image; the adaptive histogram equalization algorithm is used to improve the clarity of image details in low-light and strong light reflection scenes.

[0026] In feature extraction, the improved CNN model captures features of different dimensions such as device texture and sign text outline through multi-scale convolutional kernels, and the attention mechanism is used to strengthen the weight of key features and suppress interference from irrelevant features.

[0027] The OCR acquisition algorithm in step S3 includes: S31. Text region detection: The CTPN algorithm is used to extract candidate boxes and merge and filter them; Text line segmentation: Text lines are segmented by projection method and tilted text is corrected a second time. S32. Character segmentation: Separating characters based on connected component analysis or adaptive threshold segmentation; Character recognition: Recognizing special characters using an LSTM+CTC model combined with a customized character set for the power industry. S33. Post-processing of recognition results: Spelling correction and standardization based on power industry dictionary.

[0028] The algorithm steps of the power industry-specific large model described in step S3 include: S41. Transform power industry standard documents and compliance verification criteria into a structured rule library that meets the requirements for outputting verification conditions for applicable rule ID scenarios. S42. Clean, standardize, and associate OCR text data and image recognition scene information with power industry entities; S43. Use a Transformer-based semantic matching algorithm to locate the target rule set that is suitable for the business scenario; S44. Complete missing fields based on pre-trained parameters and verify the logical consistency of multi-source data; integrate and output rule matching results, completion information and logical verification conclusions.

[0029] In step S3, the intelligent agent collaboratively revises the document content using three technologies, generates standardized documents based on power industry rules, strengthens business review rules, and simultaneously handles document data conflicts and format deviations.

[0030] The standardized service interfaces mentioned in step S4 include image text extraction service, plant and station information intelligent matching service, generated content manual revision service, on-site survey document generation service, intelligent invoicing service, intelligent invoice review service, daily station and shift meeting document generation service, and document template management service.

[0031] The display and output described in step S5 supports interactive display on PC and mobile devices. The output file format is Word, which includes power industry safety management documents such as on-site survey documents, work order documents, and daily shift meeting documents.

[0032] The present invention has the following beneficial effects: 1. It solves the problems of low efficiency and repetitive data entry in manual data entry, saving about 50% of labor costs, freeing staff from tedious data entry and verification work, allowing them to focus on core business and improve the smoothness of management processes; 2. By using functions such as standardized completion of plant and station names and intelligent review, the filling standards are unified, reducing the inconsistency in the standardization of work tickets caused by differences in personnel operating habits and understanding, reducing the error rate by 35% and reducing management difficulty; 3. Improve the accuracy of work orders and related documents, avoid operational safety risks caused by errors or omissions, reduce risks by about 10%, effectively avoid economic losses caused by safety accidents, and ensure the smooth progress of power operations; 4. Achieve digital and intelligent upgrades in the processing of work orders and related documents, providing a model for other units in the industry to learn from, and promoting the overall development of the power industry towards intelligence and efficiency.

[0033] like Fig. 2 As shown, the technical architecture of the machine vision system consists of resource layer 1, model layer 2, artificial intelligence platform layer 3, service layer 4, application layer 5, and presentation layer 6.

[0034] Resource Layer 1: This layer provides hardware and platform support for the entire system, including: Server 7 (for model inference, training, and general business logic processing), Database 8 (for storing structured data), Computing Power 9 (for supporting the computation and operation of models and agents), Network 10 (to ensure the secure transmission and interaction of power system data), and Security Measures 11 (to resist attacks and ensure the stable operation of the power system).

[0035] Model Layer 2: With technology empowerment as the core, it relies on the Guangming Big Model 12 to build a basic intelligent capability foundation, while integrating the image recognition model 13 to form a model combination covering general intelligence and special processing capabilities, providing intelligent service support for upper-layer services.

[0036] AI Platform Layer 3: Added 14 intelligent agents for generating and reviewing security management documents, which handles intelligent splicing of key elements (plant, room, cabinet, equipment), intelligent review, intelligent generation of on-site survey documents, intelligent generation of work order documents, and intelligent generation of station / shift meeting documents.

[0037] Service Layer 4: Integrates security management document generation and review related services, and completes interface-enabled intelligent matching, including: image text extraction service 19, plant and station information intelligent matching service 15, generated content manual revision service 21, on-site survey document generation service 18, intelligent invoicing service 16, intelligent invoice review service 20, daily station and shift meeting document generation service 17, and document template management service 22. Through standardized service interfaces, it realizes intelligent processing of various business scenarios.

[0038] Application Layer 5: Focusing on application modules such as on-site survey document generation 23, work order document generation 25, daily shift meeting document generation 27, safety control document template management 24, and plant and station information management 26, the capabilities of the service layer are transformed into functions that directly support business operations, thereby improving business processing efficiency and standardization.

[0039] Presentation Layer 6: Provides interactive display platforms for system business applications on both PC and mobile devices. It is responsible for presenting various business functions and data, including: PC 28 and mobile 29.

[0040] like Fig. 3 The diagram shown is a flowchart of a machine vision system. Input and Trigger: When a user enters plant information, such as the substation name, into the web interface, the request is sent to the artificial intelligence platform layer, where the intelligent agent that generates and reviews security management documents completes the intelligent assembly of key elements.

[0041] Intelligent completion: Based on the input plant information, the system matches the complete plant name in the standardized plant name maintenance database and completes the plant name in a standardized manner.

[0042] Precise Extraction and Element Combination: Upload images in common formats such as JPEG and PNG. Using large-scale models and OCR technology, accurately extract room / space names, cabinet names, and equipment names from the images. Combine the extracted key text elements in the order of plant name + room / space name + cabinet name + equipment name, and output the text. The text output structure is in the order of plant name, room name, cabinet name, and equipment name. Example: 220kV Huangshi Power Plant: #210 Unit Central Control Room: #210 Unit Electronic Room: #210 Unit Energy Meter Panel, #210 Unit Main Transformer High Voltage Side Main Meter and Auxiliary Meter.

[0043] Manual revision and intelligent generation: Based on the combined element information, the element information is manually revised and the corresponding template is selected according to the actual business needs. The large model + manual revision provide dual protection to ensure the accuracy of the survey information. The element information and template requirements are input into the intelligent agent for generating and reviewing security control documents. Through semantic parsing and invoicing business rules, the agent intelligently fills in the work tasks, planned work time, work conditions, precautions and other contents in the document, assisting staff to quickly complete the invoice filling and generate documents in a standard format, thereby improving work efficiency.

[0044] Manual supplementation + intelligent review: For personalized information not covered in the intelligently generated content, users can manually fill in the information, including specific personnel information such as the person in charge (guardian), the work group, and the work team members (excluding the person in charge), to ensure the completeness of the document content.

[0045] After uploading the completed document, the security management document generation and review intelligent agent automatically reviews the completeness, logical consistency, and compliance of the document content based on the review rules corresponding to the work ticket. After the review is completed, it outputs a list of review items and corresponding issues.

[0046] Revision and Output: Users can directly edit, supplement, or correct problematic ticket information in the system. After modification, the system supports resubmission for review to ensure that the ticket information meets the requirements. After the review is approved, a Word document can be downloaded and exported.

[0047] Workflow: By leveraging OCR technology to extract structured text information from images with high precision, and relying on the deep semantic understanding capabilities of a power industry-specific large-scale model, a multi-level element system for power plant, substation, room, and cabinet equipment is effectively constructed. In terms of decision generation: real-time retrieval of the power industry's professional rule base provides accurate normative basis for document generation. Based on multi-source information, deep reasoning and semantic analysis are performed to intelligently generate professional content such as work tasks and safety measures that conform to business logic, ensuring that the output documents possess both high standardization and scenario adaptability. Regarding compliance closed-loop: the power industry-specific large-scale model strengthens the review rules based on power business logic, performs multi-dimensional compliance checks on the generated documents, intelligently identifies data conflicts and format deviations, and supports manual intervention for revision. The final output is a standardized document that fully complies with power industry standards, forming an intelligent and high-quality closed loop of generation, review, and revision.

[0048] Example 1: Taking a 110kV substation as an example Step 1: Plant Input and Request Initiation (Corresponding Architecture Diagram: Artificial Intelligence Platform) When a user enters a description of the substation information in natural language on the web interface: "110kV substation", the web front-end program sends the text query to the back-end security management document generation and review intelligent agent via an HTTP request.

[0049] Step 2: Complete Plant / Site Information (Corresponding Architecture Diagram: Intelligent Matching of Plant / Site Information in the Service Layer) The security management document generation and review intelligent agent Query processing service receives the user's query.

[0050] Based on keyword similarity, a weighted matching is performed by combining the plant attribute tags (voltage level, region, plant type, etc.), and results with high similarity for the same attribute are displayed first.

[0051] Step 3: Key Element Extraction and Intelligent Stitching (Corresponding architecture diagram: Image recognition model at the model layer, intelligent stitching of key elements at the artificial intelligence platform layer) Upload survey images, and use large-scale models and OCR technology to accurately extract key information such as room / space names, cabinet names, and equipment names from the images; remove interfering information and correct recognition errors to ensure that the extracted text content is accurate and to provide reliable text data support for subsequent information analysis and application.

[0052] The information is combined with the completed plant information and key elements extracted from the images, and then combined in the order of plant name + room / space name + cabinet name + equipment name. For example: Huangshi Power Plant 220kV AC: #210 Unit Central Control Room: #210 Unit Electronic Room: #210 Unit Gate Energy Meter Panel: #210 Unit Main Transformer High Voltage Side Main Meter and Auxiliary Meter.

[0053] Step 4: Revision and Generation (Corresponding architecture diagram: Manual revision of service layer content, automatic generation of AI platform layer documentation) Users can review the combined element information and check for any errors. If errors are found, they can be manually modified. Once modifications are complete or the user has no objection to the generated content, the user must confirm the current content to ensure that subsequent intelligent document generation is based on accurate information.

[0054] Users select document types from preset options based on their actual work needs. The standardized elements and key information such as the selected document type are then converted into query vectors and sent to the artificial intelligence platform for pre-generation.

[0055] Step 5: Personalized Information Supplementation and Intelligent Review (Corresponding Architecture Diagram: Service Layer Generated Content Manual Revision Service, Service Layer Intelligent Ticket Review Service) For personalized information not covered in the intelligently generated content, users can manually fill in the missing information, including specific personnel information such as the person in charge of the work (guardian), the work group, and the work team members (excluding the person in charge of the work), to ensure the completeness of the information.

[0056] Based on real-time retrieval of professional rule base rules for invoice verification, the system automatically performs comprehensive validation of all information contained on the invoice, specifically covering content completeness checks (such as whether required fields are missing), logical consistency checks (such as whether the date order is reasonable), and compliance checks (such as whether the invoice content is compliant), achieving automation and precision in invoice verification. Anomalies discovered during the intelligent invoice verification process are summarized to generate a structured problem list, facilitating quick and accurate problem verification for users.

[0057] Step 5: Verify Abnormal Content (Corresponding Architecture Diagram: Manual Revision of Content Generated by the Service Layer) According to the problem list, users can directly edit, supplement, or correct the problematic ticket information in the system. After the modification is completed, the system supports resubmission for review to ensure that the ticket information meets the requirements.

[0058] Step Six: Output Results (Corresponding Architecture Diagram: Application Layer, Presentation Layer) The revised information is regenerated at the application layer, and the final result is returned to the PC / mobile interface at the presentation layer, presented to the user in a clear and standardized format. It can be downloaded and exported to commonly used standard file formats (such as Word) for easy archiving, printing, and other subsequent operations.

[0059] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for generating security management documents based on image recognition model training technology, characterized in that: Includes the following steps: S1. Machine vision system construction: Select an image recognition model suitable for the power scenario through a multi-stage verification process; S2. Image acquisition and model training: Collect and label power field image samples, and train the model to improve the ability to identify key elements. S3, Intelligent Agent Development: Based on a large-scale model specific to the power industry, an image recognition model, and optical character recognition (OCR) technology, a three-collaborative intelligent agent is constructed to achieve image information extraction, text recognition, information completion, document generation, and compliance review. S4, empowered by intelligent service interfaces, enables intelligent processing of various business scenarios through standardized service interfaces; S5. Display and Output: Performs logical verification and rule matching on the processed data, presents it in a standardized format, and supports downloading.

2. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: The multi-stage verification process described in step S1 includes: S11. Use 5 sets of images for rapid verification and eliminate models with an accuracy rate of less than 90%; S12. Full verification using 20 sets of images; identification of weaknesses and comprehensive scoring through confusion matrix analysis. S13. Scene adaptation testing of the two highest-scoring models; selection of the model that balances performance and efficiency. A model that is compatible with various scenarios.

3. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: The image extraction and recognition algorithm of the image recognition model in step S3 includes: S21. Image preprocessing: Gaussian filtering is used for noise reduction, Hough transform correction, and adaptive histogram equalization to enhance contrast. S22. Region of Interest Extraction: Preliminary localization is based on HSV color model segmentation and morphological operations, followed by precise cropping using bounding box regression algorithm; S23. Feature Extraction: Multi-dimensional image features are extracted using an improved CNN model combined with an attention mechanism; S24. Target Recognition and Classification: Input the fully connected layer and classify it using a customized classification label library for the power scenario. Mark suspicious areas for low-confidence results and trigger secondary verification.

4. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: In the image preprocessing, the Hough transform achieves geometric correction of tilted images by detecting straight edges such as device borders and sign outlines in the image; the adaptive histogram equalization algorithm is used to improve the clarity of image details in low-light and strong light reflection scenes.

5. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: The improved CNN model in the feature extraction process captures features of different dimensions, such as device texture and sign text outline, through multi-scale convolutional kernels. The attention mechanism is used to strengthen the weight of key features and suppress interference from irrelevant features.

6. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: The OCR acquisition algorithm in step S3 includes: S31. Text region detection: The CTPN algorithm is used to extract candidate boxes and merge and filter them; Text line segmentation: Text lines are segmented by projection method and tilted text is corrected a second time. S32. Character segmentation: Separating characters based on connected component analysis or adaptive threshold segmentation; Character recognition: Recognizing special characters using an LSTM+CTC model combined with a customized character set for the power industry. S33. Post-processing of recognition results: Spelling correction and standardization based on power industry dictionary.

7. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: The algorithm steps of the power industry-specific large model described in step S3 include: S41. Transform power industry standard documents and compliance verification criteria into a structured rule library that meets the requirements for outputting verification conditions for applicable rule ID scenarios. S42. Clean, standardize, and associate OCR text data and image recognition scene information with power industry entities; S43. Use a Transformer-based semantic matching algorithm to locate the target rule set that is suitable for the business scenario; S44. Complete missing fields based on pre-trained parameters and verify the logical consistency of multi-source data; integrate and output rule matching results, completion information and logical verification conclusions.

8. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: In step S3, the intelligent agent collaboratively revises the document content using three technologies, generates standardized documents based on power industry rules, strengthens business review rules, and simultaneously handles document data conflicts and format deviations.

9. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: The standardized service interfaces mentioned in step S4 include image text extraction service, plant and station information intelligent matching service, generated content manual revision service, on-site survey document generation service, intelligent invoicing service, intelligent invoice review service, daily station and shift meeting document generation service, and document template management service.

10. The method for generating security management documents based on image recognition model training technology as described in claim 1, characterized in that: The display and output described in step S5 supports interactive display on PC and mobile devices. The output file format is Word, which includes power industry safety management documents such as on-site survey documents, work order documents, and daily shift meeting documents.

Images