A steel structure management system

Abstract

The application provides a steel structure management system, which integrates a contract information collection submodule, a third-party OCR calling submodule, a local open source large model processing submodule, an order information mapping entry submodule and an information checking and review submodule in an order management module, forms a full-process automatic processing system from contract collection, character recognition, semantic analysis, structured extraction, automatic entry to checking and review, and realizes accurate differentiation of printed and handwritten information, positioning of core elements and standardized conversion of formats through local deployment of an open source large model and special fine tuning for a steel structure contract scene, thereby completely solving the problems of low efficiency of manual entry, difficulty in handwritten character recognition and data privacy leakage. The application greatly reduces manual intervention, improves the accuracy of order information entry and the security of data, and promotes the intelligent and efficient transformation of order management of steel structure enterprises.

Description

Technical Field

[0001] This invention relates to the field of steel structure technology, and in particular to a steel structure management system. Background Technology

[0002] Currently, order information entry in management systems within the steel structure industry is generally done manually. Staff must painstakingly extract core order elements from steel structure business contracts (paper, electronic PDF / image, etc.), including contract number, partner information, steel structure product specifications (model, size, material, quantity), delivery time, delivery location, contract amount, and payment method, and then enter these elements one by one into the order module of the steel structure management system.

[0003] Some companies use simple text recognition tools or third-party OCR APIs to extract contract text, but these can only achieve basic character conversion and cannot perform semantic analysis and structured processing on the extracted text. At the same time, existing solutions mostly use large cloud models for text processing, which poses a risk of data privacy leakage. Furthermore, there is no professional optimization for the steel structure industry, and the extracted text is messy and still needs to be manually checked and sorted before it can be entered into the system. Handwritten contract information is even more difficult to process effectively due to recognition accuracy issues.

[0004] This type of existing technology has been widely used in the entire process management of steel structure production, processing, warehousing and sales. It is the conventional way for steel structure companies to manage orders. The core relies on manual operation to complete the conversion of contract information into system orders. At the technical level, it only achieves basic character recognition and has not completed the intelligentization of the entire process from "recognition" to "structured input". Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a steel structure management system that overcomes or at least partially solves the above problems.

[0006] This invention provides a steel structure management system, comprising: The contract information collection submodule is used to collect steel structure contract documents, generate a unique contract collection number for each steel structure contract document and associate it with the order number; The third-party OCR calling submodule is used to call the third-party OCR API to recognize and convert the steel structure contract document into editable plain text information, and mark the contract area corresponding to each plain text information and mark the recognition type of each character; The local open-source large model processing submodule is used to perform semantic parsing, structured extraction, and format standardization conversion on the plain text information on a local server to generate a standardized structured information dataset. The order information mapping and entry submodule is used to automatically enter the structured information dataset into the corresponding order entry items according to the preset standardized mapping relationship, and generate preliminary order information; The information verification and review submodule is used to perform automated verification on the preliminary order information and simultaneously display the original contract and preliminary order information on the order editing page for manual review and correction. After the verification is passed, it is confirmed as formal order information and the order creation is completed.

[0007] Optionally, the third-party OCR calling submodule is also used for: The steel structure contract document is denoised, watermark removed, and the underlying text information covered by the watermark is repaired by a combination algorithm of frequency domain filtering and deep learning image restoration. The pattern unrelated to the order text information in the steel structure contract document is masked by an instance segmentation and recognition region locking algorithm; Invalid content that does not conform to the coding standard is filtered out from the text initially identified by OCR. Then, the semantic coherence of the filtered text is checked by a 3-gram language model trained on a corpus specific to the steel structure industry, and invalid information below the semantic coherence threshold is removed. The algorithm uses regular expression matching and structured reorganization to remove whitespace characters from text that has passed semantic coherence verification, and then performs paragraph reorganization and line breaking on the text after removing whitespace characters.

[0008] Optionally, the local industry fine-tuning large model of the local open-source large model processing submodule is based on the open-source general large model base and is deployed after being trained on a local server for a steel structure contract scenario-specific fine-tuning; the fine-tuning training adopts the low-rank adaptation LoRA lightweight fine-tuning scheme, and the training dataset includes at least the enterprise's historical anonymized contracts, industry standard contracts, professional knowledge texts, and scenario-enhanced samples.

[0009] Optionally, the local open-source large model processing submodule is used for: The identified contract text is standardized and preprocessed, and then the standardized preprocessed contract text is segmented into sentences and words by using the general typesetting rules of steel structure contracts and the semantic verification of the industry fine-tuning large model. By combining the character type labels and coordinate area information transmitted by the third-party OCR calling submodule, printed and handwritten information are differentiated and prioritized. Based on the pre-set entity tagging system of core elements of steel structure contracts, a fusion algorithm of industry-fine-tuned large model named entity recognition and sliding window semantic matching is adopted to accurately locate various entities and the text areas corresponding to entity information. The method of semantic reasoning and cross-validation of business rules is used to perform semantic disambiguation and validation on entity information.

[0010] Optionally, the local open-source large model processing submodule is used for: Spatial clustering is performed based on the character recognition type labels and coordinate region information transmitted by the third-party OCR calling submodule to generate printed text sets and handwritten text sets respectively. The association between handwritten text blocks and printed clauses is determined by calculating coordinate overlap and semantic similarity. The association includes newly added supplementary agreements, modifications to existing clauses, and irrelevant remarks. Based on the rule that handwritten amendments to steel structure contracts take precedence over printed fixed clauses, the handwritten content determined to be an amendment to the original clauses is automatically replaced with the corresponding printed clause's entity information, and the handwritten content determined to be a new supplementary agreement is added to the text set of the corresponding clause.

[0011] Optionally, the local open-source large model processing submodule is used for: A database of easily confused information features is built based on the business rules of steel structure contracts, and entity information with confusion risk is initially screened out through rule matching; The initially screened entity information and its contextual text are input into the local industry fine-tuning model, and the true semantic reference and business dimension of the entity information are inferred through contextual semantic reasoning. The disambiguated entity information is cross-validated with the business rules of steel structure order management. The cross-validation includes consistency verification of the entity information and the unified credit code, mathematical logic verification of the amount and tax rate, chronological logic verification of the time dimension, and matching verification of the component specifications and industry standards. If the verification passes, the semantic and text regions of the entity information are locked; if the verification fails, the semantic parsing process is retried.

[0012] Optionally, the order information mapping and entry submodule is further used for: Each automatically entered entity information entry is labeled with an identification type and the source of the original contract text, whereby the identification type includes printed or handwritten text.

[0013] Optionally, the information verification and review submodule is further used for: The initial order information is subjected to format compliance verification, logical relationship verification, and consistency verification by reverse comparison with the original contract. Any verification errors are marked with prominent annotations and error type labels. After the user corrects the error information, the corrected order information is re-verified for format compliance, logical relationship verification, and consistency verification by reverse comparison with the original contract until the verification passes.

[0014] Optionally, the system further includes a permission management module and an audit log module. The permission management module is used to provide refined permission control based on job responsibilities for various operations within the order management module; The audit log module is used to record the operational behavior within the order management module in detail.

[0015] Optionally, the order management module is further used for: After the formal order information is created, it will be synchronized to other business modules of the system. These other business modules include at least a data dashboard, design management, cost management, material management, planning management, production management, quality management, production personnel, equipment management, basic management, system management, and auxiliary functions.

[0016] This invention has the following advantages: This invention integrates a contract information collection submodule, a third-party OCR call submodule, a local open-source large model processing submodule, an order information mapping and input submodule, and an information verification and review submodule within the order management module, forming a fully automated processing system from contract collection, character recognition, semantic parsing, structured extraction, automatic input to verification and review. By deploying an open-source large model locally and fine-tuning it specifically for steel structure contract scenarios, it achieves accurate differentiation between printed and handwritten information, core element positioning, and standardized format conversion, completely solving the problems of low efficiency in manual input, difficulty in handwritten recognition, and data privacy leakage. This invention significantly reduces manual intervention, improves the accuracy of order information input and data security, and promotes the transformation of steel structure enterprises' order management towards intelligence and efficiency. Attached Figure Description

[0017] Figure 1 This is a flowchart of the local open-source large model processing submodule within the order management module of the steel structure management system provided in this embodiment of the invention; Figure 2 This is a flowchart of the operation of the order management module of the steel structure management system provided in this embodiment of the invention. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] This invention adds a core function to the order management module of a steel structure management system. The core innovation lies in integrating an OCR function with invalid information filtering and a locally deployed industry-fine-tuned large model into the order management module. It makes specific optimizations for the core information of steel structure contracts, and builds a full-process functional system for automated collection, recognition, structured extraction, and standardized input of contract information. At the same time, it supports handwritten information recognition, convenient manual review, and refined access control, realizing fully automated processing of steel structure business contracts and order information in the order management module, greatly reducing manual intervention and ensuring the privacy and security of enterprise data.

[0020] In this invention, there are no special hardware requirements. Conventional computers, scanners, cameras (for scanning / photographing paper contracts), and local servers (for deploying large models) can all support the operation of the function. The core is to complete the design, collaborative work, and accurate information processing logic of each functional sub-module within the order management module.

[0021] 1. Composition of functional sub-modules within the order management module, and the functions and relationships of each module. This invention adds a contract information collection submodule, a third-party OCR call submodule, a local open-source large model processing submodule, an order information mapping and input submodule, and an information verification and review submodule to the order management module of the steel structure management system. It also enables data linkage with the system's original permission management module, audit log module, and other business modules (data dashboard, design management, cost management, material management, planning management, production management, quality management, production personnel, equipment management, basic management, system management, and auxiliary functions).

[0022] The newly added sub-modules within the order management module collaborate hierarchically, with data flow shifting unidirectionally as follows: Contract Information Collection Sub-module → Third-Party OCR Call Sub-module → Local Open Source Large Model Processing Sub-module → Order Information Mapping and Input Sub-module → Information Verification and Review Sub-module. The system permission management module provides permission control for all operational steps within the order management module. The information verification and review sub-module simultaneously provides feedback on the information recognition / input content to the human end and feedback on the verification results to the abnormal steps, forming a closed-loop process of "automated processing + convenient human review." After the order management module completes the generation of formal order information, it synchronizes data with other business modules in the system.

[0023] The specific functions of each submodule are as follows: (1) Contract Information Collection Submodule As the contract data entry point for the order management module, it is embedded in the order creation page of the order management module. It supports the uploading and collection of steel structure contracts in multiple formats, and has file format verification and clarity detection functions. It provides targeted collection optimization prompts for contracts containing handwritten core information to ensure the validity of the basic data. After collection, a unique contract collection number is generated for each contract, which is linked and bound to the order number in the order management module to achieve full-process data traceability. At the same time, the recognition priority of core information is preset to ensure the priority extraction of key information.

[0024] (2) Third-party OCR calling submodule Embedded within the order management module, it interfaces with OCR APIs provided by external vendors, receives contract files transmitted from the contract information collection submodule, accurately recognizes printed and handwritten characters, and converts contract content in image / non-editable text format into editable plain text information. It has a built-in invalid information filtering function that automatically removes watermarks, garbled characters, irrelevant patterns, blank characters, etc. from the contract, while marking the specific contract area and character recognition type corresponding to the recognized text. It performs specialized recognition optimization for the character features of core information, and transmits the data to the local open-source large model processing submodule after associating it with the contract collection number.

[0025] The specific implementation method of the invalid information filtering function is as follows: This function employs a two-layer filtering architecture of "image layer preprocessing + text layer rule verification." It matches corresponding algorithm models and processing flows to different types of invalid information in steel structure contracts, accurately removing invalid information without compromising the core valid text. The specific implementation is as follows: 1) Watermark Removal: A combined algorithm of frequency domain filtering and deep learning image inpainting The processing consists of three steps: ① Image preprocessing: The uploaded color image of the contract is converted into a single-channel grayscale image. Bilateral filtering is used to smooth image noise while preserving character edge features. ② Semi-transparent watermark removal: A two-dimensional discrete Fourier transform is performed on the grayscale image to convert it from the spatial domain to the frequency domain. An adaptive band-stop filter is used to filter out the periodic frequency peaks corresponding to the watermark. Then, an inverse Fourier transform is used to restore the image to the spatial domain, eliminating the interference of the semi-transparent watermark on character recognition. ③ Opaque watermark / strong overlay watermark processing: A YOLOv8 instance segmentation model optimized for steel structure contract scenarios is used. A feature dataset of the company logo, official seal, and overlay watermark is pre-trained to accurately locate the pixel coordinates of the watermark area. Then, a Context Encoder generative adversarial network is used to perform pixel-level repair of the watermark area, restoring the underlying text information covered by the watermark and avoiding the loss of effective text.

[0026] 2) Garbled character removal: A dual verification algorithm combining encoding verification and semantic coherence. The processing consists of two steps: ① Encoding format verification: The character set initially identified by OCR is verified for UTF-8 / GBK encoding standards, and invalid content such as garbled characters outside the encoding standards, invisible ASCII control characters, and irregular symbols are automatically filtered out; ② Semantic coherence verification: A 3-gram language model is trained based on a corpus specific to the steel structure industry to perform probability scoring on the semantic coherence of the recognized text. Garbled segments with a single sentence semantic score lower than a preset threshold are judged as invalid information and automatically removed, while retaining the paragraph logic of the valid text in the context.

[0027] 3) Irrelevant pattern masking: Instance segmentation + recognition region locking algorithm Processing procedure: An optimized YOLOv8 instance segmentation model is used to train features on irrelevant patterns commonly found in steel structure contracts, such as company logos, official seals, decorative patterns, signature panel illustrations, and QR codes / barcodes. The model accurately identifies and outputs the coordinates of the smallest bounding rectangle of these irrelevant patterns. In the OCR recognition stage, the recognition operation of these coordinate areas is automatically blocked to avoid interference from irrelevant patterns on text recognition. At the same time, invalid recognition content corresponding to the patterns is directly removed, and only the recognition results of the contract text area are retained.

[0028] 4) Whitespace character handling: Regular expression matching + structured recombination algorithm The processing consists of two steps: ① Batch whitespace character removal: Using preset regular expressions, the system accurately matches and identifies whitespace characters such as full-width / half-width spaces, tabs, consecutive line breaks, and consecutive blank lines between paragraphs in the text, and performs a full removal operation; ② Text restructuring: Based on the standard typesetting rules of steel structure contracts (commercial clauses divided into paragraphs, product lists divided into lines, and monetary information divided into columns), the text after removing whitespace characters is restructured into paragraphs and lines, preserving the semantic logic and typesetting structure of the core content, and avoiding semantic errors caused by the removal of whitespace characters.

[0029] (3) Local open-source large model processing submodule As the core text processing submodule within the order management module, it is based on an open-source large model deployed on the enterprise's local server, completing text processing entirely locally without any external data transmission, fundamentally ensuring the privacy and security of the enterprise's sensitive contract data.

[0030] This model adopts a full-process training system of "base selection - dataset construction - multi-stage lightweight fine-tuning - performance evaluation - local deployment optimization". Based on the open-source general large model base, it completes the special optimization for steel structure contract scenarios. The entire process is completed on the enterprise's local server, with no risk of data leakage. The specific training process is as follows: Phase 1: Selection of Basic Pre-trained Models The preferred choice is the open-source version of Tongyi Qianwen 2 (Qwen-7B / 14B) as the basic pre-training platform. Alternatives include the open-source models Llama3 8B / 70B and ChatGLM3-6B. Other open-source models can also be selected. The core selection criteria are: ① Fully open-source and commercially viable, with no intellectual property risks, and suitable for enterprise private deployment needs; ② Excellent Chinese semantic understanding capabilities and strong adaptability to contextual understanding of long text contracts; ③ Supports lightweight LoRA fine-tuning, requiring low training computing power, and can be trained on enterprise-grade GPU servers; ④ Supports 32K context windows, capable of fully processing single steel structure contracts of dozens of pages; ⑤ Natively supports RAG retrieval enhancement architecture, seamlessly integrating with the steel structure industry's dedicated knowledge base.

[0031] Phase 2: Construction and Preprocessing of Specialized Training Datasets To address the steel structure contract business scenario, a multi-dimensional, high-quality specialized training dataset was constructed, comprehensively covering various contract formats and scenarios in actual enterprise operations. The specific process is as follows: The dataset is composed of four main categories: First, a core enterprise business dataset, containing over 3,000 historical anonymized steel structure processing, installation, and sales contracts from the construction industry, covering different project types, contract formats, and handwritten modification scenarios; second, an industry standard dataset, containing over 1,000 standard contract templates and specification texts issued by the Ministry of Housing and Urban-Rural Development and the Steel Structure Industry Association; third, a professional knowledge dataset, containing textual materials such as steel structure design / construction specifications, material standards, professional terminology dictionaries, and industry business clause libraries; and fourth, a scenario enhancement dataset, which generates over 2,000 enhanced samples for high-frequency core scenarios such as handwritten modifications, monetary and tax matters, component specifications, and delivery terms, through back translation, random replacement, and scenario augmentation, improving the model's adaptability to complex scenarios.

[0032] Data preprocessing: First, the entire dataset is anonymized, and sensitive information such as corporate entities, bank accounts, amounts, and contact persons in the contracts is de-identified to ensure data security. Next, data cleaning is performed, filtering out invalid text, garbled characters, and incomplete samples, and standardizing the format, punctuation, and encoding of all text. Then, a specific annotation standard for steel structure contracts is developed, and annotation is completed jointly by steel structure business experts and professional data annotators. Annotation dimensions include core entity labels, entity relationships, semantic types, handwritten modification relationships, and format standards, forming a high-quality annotated dataset. Finally, the dataset is split into training, validation, and test sets in an 8:1:1 ratio for model training, performance validation, and generalization ability testing, respectively.

[0033] Phase 3: Multi-stage lightweight fine-tuning training We adopt a lightweight fine-tuning scheme using LoRA (Low-Rank Adaptation), which trains only the newly added low-rank adaptation matrix without modifying the weights of the model's backbone network. This significantly reduces the computational requirements for training while fully preserving the general capabilities of the basic model. Using steel structure industry standards, standard contracts, and a professional terminology database as training data, we employ fully supervised fine-tuning (SFT) to train the model's understanding of steel structure industry professional terminology, text structure, and semantic logic. This addresses the core issues of insufficient industry knowledge and errors in professional terminology recognition in general large models. With minimizing cross-entropy loss as the training objective, we iteratively optimize the model's accuracy in understanding industry semantics.

[0034] Phase 4: Model Performance Evaluation and Iterative Optimization Establish a multi-dimensional evaluation index system specifically for steel structure contract scenarios, and conduct full-dimensional effect verification of the fine-tuned model. The core evaluation indexes include: core entity recognition accuracy ≥99%, core information extraction recall ≥99.5%, semantic parsing accuracy ≥98%, format standardization compliance rate 100%, and single contract reasoning response time ≤10s. Through batch testing on the test set and actual business scenario testing, supplement training samples for complex scenarios that do not meet the standards, and iteratively fine-tune the model until all dimensions of the index meet the business requirements.

[0035] Phase 5: Local Lightweight Deployment Optimization After completing model training and performance verification, deployment optimization was performed on the enterprise's local server environment: GPTQ / AWQ quantization algorithms were used to perform INT4 / INT8 low-bit quantization on the model, significantly reducing the model's GPU memory usage while keeping accuracy loss under control; through operator fusion, KV cache optimization, and batch processing scheduling optimization, the model inference speed was further improved; finally, the optimized model was encapsulated into a standardized intranet API interface, which was seamlessly integrated with the order management module, third-party OCR call submodule, and information verification and review submodule, realizing closed-loop data flow throughout the entire process within the intranet.

[0036] This submodule receives plain text information transmitted from a third-party OCR calling submodule and performs three core functions. For detailed processing flow, please refer to [link / reference]. Figure 1 : 1) Precise semantic analysis: Professional sentence and word segmentation of contract text, accurate location of text areas corresponding to various core information, effective differentiation between printed and handwritten information content, and complete resolution of industry pain points such as confusion and inaccurate positioning of core information; This stage employs a three-tiered processing architecture: "lightweight fine-tuning of an open-source large model + NLP basic task pipeline + strong validation of a steel structure contract rule engine." Based on locally deployed open-source large models (such as Tongyi Qianwen open-source version, Llama 3, Qwen, etc.), it performs specialized fine-tuning for steel structure contract scenarios. Combined with basic natural language processing algorithms and a contract business rule engine, it achieves accurate semantic parsing of contract text across all dimensions. The complete processing consists of four core steps, the specific implementation of which is as follows: Step 1: Contract text preprocessing and professional sentence and word segmentation. This step takes the structured plain text output by OCR as input and adopts a dual-track processing mode of "initial processing by rule engine + semantic verification of large model with industry fine-tuning" to complete the sentence segmentation and word segmentation operations that conform to the semantic logic of steel structure contracts. The specific implementation process is as follows: ① Text standardization preprocessing: First, the OCR-recognized text is standardized by unifying full-width and half-width characters, standardizing Chinese and English punctuation, and standardizing line breaks and paragraph breaks to eliminate the format chaos caused by OCR recognition. At the same time, meaningless scattered characters remaining in the preprocessing stage are filtered out to generate a standardized continuous text stream. ② Rule-level initial sentence segmentation: Based on the general typesetting rules of steel structure contracts, a sentence segmentation feature rule library is constructed. The basic sentence segmentation identifiers are period, semicolon, line break, chapter title number, and bullet point. At the same time, exception rules are set for scenarios in steel structure contracts such as "component specification parameters with decimal points / semicolons" and "amount values ​​with thousands separators" to avoid sentence breaks, complete the initial sentence segmentation of the text, and generate basic sentence units. ③ Large Model Semantic Sentence Segment Validation: Input the preliminary sentence segmentation results into the locally fine-tuned open-source large model, and guide the large model to perform semantic coherence validation on the sentence segmentation results through the preset Prompt command, correct the errors of semantic breakage and sentence merging, and finally generate a standard sentence set that conforms to the semantic logic of the contract. ④ Steel Structure Specialty Word Segmentation: Based on a dedicated corpus for the steel structure industry, the Jieba word segmentation algorithm is optimized. High-frequency terms such as steel structure material standards (Q235B, Q355B, etc.), component types (tube trusses, H-beams, box columns, etc.), processing technology, and professional terms in business terms are added to a custom dictionary and set to the highest segmentation priority. First, basic word segmentation is completed using the optimized word segmentation algorithm. Then, the word segmentation results are validated for professional terms using a local large model to correct errors in professional terminology splitting and homophone / similar character recognition. Finally, word segmentation results that conform to the semantic specifications of the steel structure industry are generated, laying the foundation for subsequent semantic analysis.

[0037] Step 2: Differentiating and prioritizing printed and handwritten information This step combines the character type labels and coordinate region information transmitted by the OCR submodule, and uses the method of "label clustering and hierarchical structure + large model semantic association verification" to accurately distinguish between printed and handwritten information and determine modification relationships. The specific implementation process is as follows: ① Character type label clustering: Based on the "printed / handwritten" type label, page number and coordinate information of each character output by OCR, spatial clustering is performed on the text content. Text content of the same coordinate area and the same type is aggregated to generate printed text set and handwritten text set respectively. At the same time, the area of ​​printed text corresponding to each handwritten text block is marked. ② Spatial and semantic dual association matching: For each handwritten text block, firstly, the corresponding related clause area in the printed text is matched by calculating the coordinate overlap and horizontal / vertical distance; then, the handwritten text and the corresponding printed clause are input into the local large model, and the association relationship between the two is determined by semantic similarity calculation (based on the Sentence-BERT model to generate sentence vectors and calculate cosine similarity) and semantic reasoning, including three categories: "new supplementary agreement", "modification of existing clauses" and "irrelevant remarks". ③ Content Priority Determination and Text Update: Based on the industry-standard rule in steel structure contracts that "handwritten amendments take precedence over printed fixed clauses," handwritten content deemed as "amendments to existing clauses" is automatically replaced with the core information of the corresponding printed clauses; handwritten content deemed as "new supplementary agreements" is added to the text set of the corresponding clauses; and content deemed as "irrelevant remarks" is marked separately and not included in the core information extraction scope. Ultimately, standardized semantic text with distinct types and clear priorities is generated, completely resolving the issues of confusion between printed and handwritten content and unclear amendment relationships.

[0038] Step 3: Precisely locate the text area corresponding to the core information This step, based on the core element system of steel structure contracts, employs a fusion algorithm of "fine-tuned large-model named entity recognition (NER) + sliding window semantic matching" to achieve accurate positioning of text regions corresponding to various core information types. The specific implementation process is as follows: ① Construct a core element entity tag system for steel structure contracts: Based on the core input fields of steel structure order management, 12 major categories of core entity tags are preset, including contract subject information, contract subject information, amount and tax information, delivery and performance information, project management information, component specification information, quality requirement information, payment method information, liability for breach of contract information, rights and obligations of both parties information, dispute resolution information, and other agreed information. Each major category of tags is further subdivided into sub-entity tags to form a complete entity tag system; ② Large-scale model NER fine-tuning: Based on thousands of annotated steel structure contract corpora, the local open-source large model is fine-tuned using lightweight LoRA to enable it to accurately identify the core entities of steel structure contracts; the standardized text that has been segmented into sentences and words and distinguished by type is input into the fine-tuned large model, and all core entities in the text are identified through the NER task, and the entity label, text content, sentence and start and end character positions of each entity are output; ③ Precise text region locking and coordinate association: Based on the start and end character positions of the entity recognition results, the sentence text region corresponding to the core information is locked, and the character coordinate information output by the OCR submodule is associated to generate the contract page number and text block coordinate range corresponding to the core information, thus completing the strong binding of "core entity - semantic text region - contract original spatial location"; ④ Secondary verification using a sliding window: For high-frequency core information such as amount, component specifications, and delivery time, a sliding window with a fixed step size is used to perform a secondary traversal of the entire text. Combined with keyword regular expression matching and semantic similarity calculation, the NER recognition results of the large model are completed and verified to avoid missing core information and ensure 100% accurate positioning of the text regions corresponding to all core information.

[0039] Step 4: Semantic disambiguation and obfuscated information verification closed loop This step addresses easily confused core information in steel structure contracts by employing a method of "contextual semantic reasoning + cross-validation of business rules" to achieve semantic disambiguation and completely resolve the industry pain point of core information confusion. The specific implementation process is as follows: ① Initial screening of ambiguous information: Based on the business rules of steel structure contracts, a database of easily confused information features is constructed, including high-frequency confusion scenarios such as "confusing the main information of Party A / Party B", "confusing the amount including tax / excluding tax / tax amount", "confusing the provisional amount / fixed amount", "confusing the delivery time / signing time / warranty period", and "confusing the specifications and parameters of different components". Entity information with confusion risks is initially screened out through rule matching; ② Large Model Context Semantic Disambiguation: Input the initially screened ambiguous information, its context text, and the steel structure contract business rules into the local open-source large model. Through context semantic reasoning, clarify the true semantic reference and business dimension of the ambiguous information, and complete semantic disambiguation. For example, for multiple sets of monetary values, through the logical relationship of "amount excluding tax × tax rate = tax amount" and context semantics, accurately distinguish easily confused values ​​such as amount excluding tax, amount including tax, tax amount, and prepayment amount. ③ Business rule cross-validation closed loop: The disambiguated core information is cross-validated with the business rules of steel structure order management, including consistency verification of main information and unified credit code, mathematical logic verification of amount and tax rate, logical verification of time sequence, and matching verification of component specifications with industry standards. If the verification passes, the semantic and text areas of the core information are locked; if the verification fails, the semantic parsing process is re-triggered until the verification passes, forming a complete closed loop of semantic disambiguation and obfuscation verification, fundamentally solving the problems of core information obfuscation and inaccurate positioning.

[0040] 2) Structured Information Extraction: Based on the requirements of steel structure order management, all core information such as contract number, information of both parties, bank account information, contract signing date, tax rate, various contract amounts and their contents are automatically extracted from the contract text, realizing the dimensional breakdown and extraction of information; 3) Format standardization conversion: All extracted core information is converted into a standard format that matches the order entry requirements of the order management module. For example, date information is uniformly converted into a fixed number format, tax rate is uniformly converted into a percentage format, amount information is uniformly retained to a specified number of decimal places, and handwritten text and amount are converted into printed text format. Finally, a standardized structured information dataset is generated, which is then synchronously transmitted to the order information mapping and entry submodule and the information verification and review submodule after being associated with the contract collection number.

[0041] (4) Order Information Mapping and Input Submodule As the core data entry submodule within the order management module, it pre-establishes a standardized mapping relationship between the structured information of steel structure contracts and the order entry items in the order management module. It receives the standardized structured information dataset transmitted from the local open-source large model processing submodule, automatically enters various core information into the corresponding order entry items in the order management module, and generates preliminary system order information. It marks the identification type (printed / handwritten) and the source of the original contract for each entered information, binds it with the order number to realize full-process traceability of information, and transmits it to the information verification and review submodule after associating it with the contract collection number.

[0042] (5) Information verification and review submodule This is a core sub-module within the order management module that ensures accurate information entry. It features both automated, precise verification and convenient manual review. It incorporates proprietary verification logic designed specifically for the core information in steel structure contracts and is embedded in the order editing page of the order management module. 1) Automated Validation: Three-fold validation rules are executed. First, format validation: the format and number of characters of information such as date, tax rate, amount, unified credit code, and bank account number are validated for compliance to match the format requirements of the order management module. Second, logical validation: the mathematical logical relationship between various amounts and tax rates is precisely validated to ensure the accuracy of data calculation. Third, consistency validation: the information entered into the order management module is reverse-checked with the original contract text recognized by OCR to ensure that no information is missed or incorrectly entered.

[0043] 2) Convenient manual review: The original contract text after OCR recognition and the automatically entered order information are displayed simultaneously on the order editing page of the order management module. The display area is clearly divided according to the information dimension. Users can manually copy the recognized content to modify the entered information. Abnormal information found by automatic verification is highlighted in red and the abnormality type is marked, which makes it easy for users to make targeted manual corrections. After correction, automatic verification can be run again until the verification passes.

[0044] Once the verification passes, the preliminary order information is confirmed as the official order information, and the order is created within the order management module. If the verification fails, the user must manually correct the information within the order management module and then re-verify it, thus forming a closed-loop process.

[0045] (6) Existing module linkage function of the system Access Control Module: Provides fine-grained access control for all operations within the order management module. Based on the enterprise's organizational structure and job requirements, operation permissions are set according to the importance of information. Different operation functions within the order management module are opened to personnel in different positions, realizing "dedicated authority and dedicated responsibility for dedicated personnel". Audit Log Module: Records all operations within the order management module in detail, including the operator, operation time, operation content, and information comparison before and after modification; Other business modules: After the order management module completes the creation of a formal order, it will synchronize the core order data to all business modules such as data dashboard, design management, cost management, material management, planning management, production management, quality management, production personnel, equipment management, basic management, system management, and auxiliary functions, providing data support for each link. Each module opens information viewing and operation functions according to user permissions.

[0046] 2. Specific steps for operating the functions within the order management module For a detailed overview of the entire process within the order management module, from contract collection to order creation and data synchronization, please refer to [link / reference]. Figure 2 The specific operating steps are as follows: 1) Contract Information Collection: Staff members upload steel structure contract documents through the contract information collection submodule on the order creation page of the order management module. The submodule automatically detects the file format and clarity and transmits the file to a third-party OCR call submodule.

[0047] 2) Third-party OCR recognition: The third-party OCR calls the submodule to connect to the external API to complete the full character recognition of printed and handwritten text in the contract document, convert it into editable plain text, filter out invalid information, mark the text corresponding to the contract area, associate the collection number and transmit it to the local open source large model processing submodule.

[0048] 3) Local large model structuring: The local open-source large model processing submodule completes text processing on the local server, calls the steel structure industry-specific corpus, performs semantic parsing, core information structuring extraction and format standardization conversion on the plain text, generates a standardized structured information dataset, and synchronously transmits it to the order information mapping and entry submodule and the information verification and review submodule.

[0049] 4) Automatic order information entry: The order information mapping and entry submodule automatically enters various core information from the structured information dataset into the corresponding order entry items of the order management module according to the preset standardized mapping relationship, generates preliminary order information and marks the information source and identification type.

[0050] 5) Automated verification + manual review: The information verification and review sub-module synchronously displays the original contract and preliminary order information on the order editing page of the order management module, and supports manual copying / modification; it automatically performs triple verification of format, logic and consistency. If the verification passes, it is directly confirmed as formal order information in the order management module and the order creation is completed; if the verification fails, the abnormal information is marked in red, and the user has to manually correct it in the module and verify it again.

[0051] 3. Specific application operation process and results Application scenarios A building technology industrial company (Party B) and a construction engineering company (Party A) signed a printed steel structure processing contract. The printed version included core information such as the contract number, the names of both parties, the unified credit code, the tax rate, and the amount excluding tax. The handwritten version supplemented and modified key content such as the amount including tax, the tax amount, and Party A's bank account. Staff were required to accurately enter all the core information of the contract into the system's order management module and create a formal order. Strict protection of the privacy and security of sensitive data such as bank account and contract amount was required, and the order creation work had to be completed quickly.

[0052] Running process 1) On the order creation page of the system's order management module, staff members upload a clear scan of the contract through the contract information collection sub-module. The sub-module checks that the file format and clarity meet the standards, generates a unique collection number and binds it to the order number to be created, and then transfers the file to a third-party OCR call sub-module. 2) The third-party OCR call submodule connects to the external OCR API to complete the full recognition of printed and handwritten characters in the contract, filter out the company logo watermark in the contract to generate a plain text file, mark the contract area corresponding to each character, and associate the collection number to transmit to the local open source large model processing submodule. 3) The local open-source large model processing submodule completes text processing on the local server, accurately parsing all the core information in the contract, including the contract number, information of both parties, tax rate, and amount excluding tax in printed form, as well as the amount including tax, tax amount, and bank account number of Party A in handwritten form. After standardizing and converting all the information into a format, a structured information dataset is generated. 4) The order information mapping and entry submodule automatically enters all core information into the corresponding order entry items in the order management module according to the preset mapping relationship, generating preliminary order information; 5) The information verification and review submodule synchronously displays the original contract and preliminary order information on the order editing page of the order management module, and automatically performs triple verification: matching the input requirements of various information formats, matching the logical relationship between amount and tax rate, and ensuring that the input information is completely consistent with the original contract. If the verification passes, the information is confirmed as formal order information in the module, and the order creation is completed.

[0053] Execution results The entire contract information entry and order creation process takes approximately one minute. Staff only need to upload the contract file within the order management module, with no manual information extraction or entry operations. Core sensitive data is processed on the company's local server without any external transmission, completely eliminating the risk of data leakage. The official order information generated within the order management module is 100% consistent with the original contract (including handwritten modifications), with standardized information formats, accurate logical relationships, and no entry errors. Order information is quickly synchronized from the order management module to various business modules within the company, improving efficiency by over 97% compared to the traditional manual entry method. This avoids production errors, financial settlement problems, and business disputes caused by information errors from the source. All operations are subject to refined permission control and recorded in the audit log, achieving standardized and traceable management of order information within the order management module, fully adapting to the actual business needs of the company.

[0054] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0055] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0056] The steel structure management system provided by this invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application. The above embodiments are merely preferred embodiments given to fully illustrate this invention, and the protection scope of this invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on this invention are all within the protection scope of this invention.

Claims

1. A steel structure management system, characterized in that, The system includes an order management module, which includes: The contract information collection submodule is used to collect steel structure contract documents, generate a unique contract collection number for each steel structure contract document and associate it with the order number; The third-party OCR calling submodule is used to call the third-party OCR API to recognize and convert the steel structure contract document into editable plain text information, and mark the contract area corresponding to each plain text information and mark the recognition type of each character; The local open-source large model processing submodule is used to perform semantic parsing, structured extraction, and format standardization conversion on the plain text information on a local server to generate a standardized structured information dataset. The order information mapping and entry submodule is used to automatically enter the structured information dataset into the corresponding order entry items according to the preset standardized mapping relationship, and generate preliminary order information; The information verification and review submodule is used to perform automated verification on the preliminary order information and simultaneously display the original contract and preliminary order information on the order editing page for manual review and correction. After the verification is passed, it is confirmed as formal order information and the order creation is completed.

2. The steel structure management system according to claim 1, characterized in that, The third-party OCR calling submodule is also used for: The steel structure contract document is denoised, watermark removed, and the underlying text information covered by the watermark is repaired by a combination algorithm of frequency domain filtering and deep learning image restoration. The pattern unrelated to the order text information in the steel structure contract document is masked by an instance segmentation and recognition region locking algorithm; Invalid content that does not conform to the coding standard is filtered out from the text initially identified by OCR. Then, the semantic coherence of the filtered text is checked by a 3-gram language model trained on a corpus specific to the steel structure industry, and invalid information below the semantic coherence threshold is removed. The algorithm uses regular expression matching and structured reorganization to remove whitespace characters from text that has passed semantic coherence verification, and then performs paragraph reorganization and line breaking on the text after removing whitespace characters.

3. The steel structure management system according to claim 1, characterized in that, The local open-source large model processing submodule's local industry fine-tuning large model is based on an open-source general large model base and is deployed after being trained on a local server for a steel structure contract scenario. The fine-tuning training adopts a low-rank adaptation LoRA lightweight fine-tuning scheme, and the training dataset includes at least the enterprise's historical anonymized contracts, industry standard contracts, professional knowledge texts, and scenario-enhanced samples.

4. The steel structure management system according to claim 3, characterized in that, The local open-source large model processing submodule is used for: The identified contract text is standardized and preprocessed, and then the standardized preprocessed contract text is segmented into sentences and words by using the general typesetting rules of steel structure contracts and the semantic verification of the industry fine-tuning large model. By combining the character type labels and coordinate area information transmitted by the third-party OCR calling submodule, printed and handwritten information are differentiated and prioritized. Based on the pre-set entity tagging system of core elements of steel structure contracts, a fusion algorithm of industry-fine-tuned large model named entity recognition and sliding window semantic matching is adopted to accurately locate various entities and the text areas corresponding to entity information. The method of semantic reasoning and cross-validation of business rules is used to perform semantic disambiguation and validation on entity information.

5. The steel structure management system according to claim 4, characterized in that, The local open-source large model processing submodule is used for: Spatial clustering is performed based on the character recognition type labels and coordinate region information transmitted by the third-party OCR calling submodule to generate printed text sets and handwritten text sets respectively. The association between handwritten text blocks and printed clauses is determined by calculating coordinate overlap and semantic similarity. The association includes newly added supplementary agreements, modifications to existing clauses, and irrelevant remarks. Based on the rule that handwritten amendments to steel structure contracts take precedence over printed fixed clauses, the handwritten content determined to be an amendment to the original clauses is automatically replaced with the corresponding printed clause's entity information, and the handwritten content determined to be a new supplementary agreement is added to the text set of the corresponding clause.

6. The steel structure management system according to claim 4, characterized in that, The local open-source large model processing submodule is used for: A database of easily confused information features is built based on the business rules of steel structure contracts, and entity information with confusion risk is initially screened out through rule matching; The initially screened entity information and its contextual text are input into the local industry fine-tuning model, and the true semantic reference and business dimension of the entity information are inferred through contextual semantic reasoning. The disambiguated entity information is cross-validated with the business rules of steel structure order management. The cross-validation includes consistency verification of the entity information and the unified credit code, mathematical logic verification of the amount and tax rate, chronological logic verification of the time dimension, and matching verification of the component specifications and industry standards. If the verification passes, the semantic and text regions of the entity information are locked; if the verification fails, the semantic parsing process is retried.

7. The steel structure management system according to claim 1, characterized in that, The order information mapping and entry submodule is also used for: Each automatically entered entity information entry is labeled with an identification type and the source of the original contract text, whereby the identification type includes printed or handwritten text.

8. The steel structure management system according to claim 1, characterized in that, The information verification and review submodule is also used for: The initial order information is subjected to format compliance verification, logical relationship verification, and consistency verification by reverse comparison with the original contract. Any verification errors are marked with prominent annotations and error type labels. After the user corrects the error information, the corrected order information is re-verified for format compliance, logical relationship verification, and consistency verification by reverse comparison with the original contract until the verification passes.

9. The steel structure management system according to claim 1, characterized in that, The system also includes a permission management module and an audit log module. The permission management module is used to provide refined permission control based on job responsibilities for various operations within the order management module; The audit log module is used to record the operational behavior within the order management module in detail.

10. The steel structure management system according to claim 1, characterized in that, The order management module is also used for: After the formal order information is created, it will be synchronized to other business modules of the system. These other business modules include at least a data dashboard, design management, cost management, material management, planning management, production management, quality management, production personnel, equipment management, basic management, system management, and auxiliary functions.

Images