Method, computing device, non-transitory computer-readable recording medium, and computer program product for tracing and verifying for construction inspection form

Abstract

The present application provides a method for tracing and verifying for a construction inspection form, which is performed by a computing device after loading and executing a computer program product. The method includes the following steps: receiving a construction inspection form, which is associated with a construction regulatory text and has a plurality of construction inspection items and a plurality of construction inspection criteria corresponding to the plurality of construction inspection items, respectively; inputting the construction inspection form into a embedding model; by the embedding model, performing word embedding vectorization on the contents of the construction inspection form, and then generating and outputting a construction inspection vectorization result corresponding to the construction inspection form; comparing the construction inspection vectorization result with a plurality of project drawing name vectorization results stored in a construction illustration vector database for similarity, and then searching a plurality of project candidate drawings corresponding to a traceability verification item in the plurality of construction inspection items from the construction illustration vector database; inputting the plurality of project candidate drawings into a visual language model, and inputting at least one of the traceability verification item and the traceability verification criterion corresponding to the traceability verification item into the visual language model; by the visual language model, analyzing the contents of the plurality of project candidate drawings and filtering an project final drawing from the plurality of project candidate drawings based on at least one of the traceability verification item and the traceability verification criterion; by the visual language model, scoring and ranking the plurality of project final cropping sub-drawings corresponding to the project final drawing based on at least one of the traceability verification item and the traceability verification criterion, and then extracting a best cropping sub-drawing from the plurality of project final cropping sub-drawings; and comparing the contents of the best cropping sub-drawing with the contents of the traceability verification criterion for semantic consistency, and then generating a verified comparison result; wherein the plurality of project drawing name vectorization results are results generated after performing word embedding vectorization on the content of a plurality of project drawing names by the embedding model; and wherein the plurality of project final cropping sub-drawings are generated by cropping the project final drawing. Thereby, the method can be applied to traceability verification of construction inspection forms, and solve the problem that inspectors are unable to effectively perform traceability verification. In addition, a computing device, a non-transitory computer-readable recording medium, and a computer program product for tracing and verifying for a construction inspection form are also provided.

Description

[Technical Field]

[0001] This application relates to a method for verification, a computer device, a non-transitory computer-readable recording medium, and a computer program product, and more particularly to a method for tracing and verifying construction inspection forms, a computer device, a non-transitory computer-readable recording medium, and a computer program product. [Previous Technology]

[0002] First, for construction projects, in order to ensure the construction quality of each project, it is often necessary to conduct engineering inspections. More specifically, inspectors will conduct engineering inspections based on the contents of the inspection form to confirm that the construction project has indeed been carried out in accordance with the contents stipulated in the construction specifications during the construction process, thereby ensuring the construction quality of the project.

[0003] Furthermore, in practice, before using inspection forms to conduct engineering inspections, it is often necessary to first check the contents of the inspection forms against the contents recorded in the construction drawings (i.e., the documents with textual and / or numerical descriptions accompanying the engineering drawings) to confirm whether the inspection items and inspection criteria in the inspection forms are consistent with the contents recorded in the construction drawings. This ensures that inspectors can correctly conduct engineering inspections based on inspection forms whose contents are consistent with the construction drawings, and avoids erroneous inspection results caused by inspection forms whose contents are inconsistent with the construction drawings. [Summary of the Invention]

[0004] Generally speaking, the inspection of the inspection forms usually requires manpower to carefully examine the contents of hundreds of pages of construction drawings and documents. However, this method is not only time-consuming and laborious, but also difficult to implement in practice, which often leads to the failure to implement the inspection of the inspection forms in practice.

[0005] According to the statistical results of common deficiencies in the quality management system for public works construction inspection mentioned in "Exploring the Implementation Effectiveness of the Three-Level Quality Management System for Public Works Construction from an Engineering Practice Perspective" (the statistics are from January 1, 2016 to June 30, 2020), the most common deficiency is "the quality control self-inspection checklist is not implemented, or the inspection standards are not quantified, allowable error values ​​are not defined, or the inspection values ​​are not accurately recorded," with a deficiency rate as high as 84.01%. Meanwhile, the fifth and eighth ranked items in this statistical result are "the lack of established quality management standards for each material / equipment and construction, or non-compliance" (with a deficiency rate as high as 41.91%) and "the lack of established quality management standards for each sub-item of the project" (with a deficiency rate as high as 31.71%), both of which are also related to the lack of established or non-compliance of inspection management standards. Therefore, the aforementioned statistical results reflect the current practical difficulty in implementing the inspection of checklists.

[0006] Furthermore, with the evolution of artificial intelligence, various industries and technological fields have introduced technologies such as generative artificial intelligence (GAI) in the hope of using the high-speed computing capabilities of computer devices to solve practical problems caused by manpower requirements. However, taking large language models as an example, although current large language models possess powerful natural language processing capabilities, enabling them to analyze text content in various technological fields and understand the contextual semantics and intent of the text content to a certain extent, thereby generating corresponding results, current large language models still have significant limitations. For example, current large language models can only analyze pure text content and cannot analyze construction drawings, thus making it impossible to trace the content of construction inspection forms back to the various engineering drawings in the construction drawings for verification.

[0007] Therefore, how to solve the above-mentioned problems encountered by traditional inspection methods and effectively and accurately determine whether the contents of the construction inspection form conform to the contents of the construction drawing documents has become an urgent problem that this technical field hopes to solve.

[0008] To address the above problems, this application provides a method for tracing and verifying construction inspection forms. This method is executed after a computer program is loaded and executed via a computer device. The method includes the following steps: receiving a construction inspection form, wherein the construction inspection form is associated with a construction drawing document, and the construction inspection form has a plurality of construction inspection items and a plurality of construction inspection standards corresponding to the plurality of construction inspection items; inputting the construction inspection form into a vectorized model; and, through the vectorized model, targeting... The content of the construction inspection form is vectorized using word embedding, generating and outputting a construction inspection vectorization result corresponding to the form. This vectorization result is then compared for similarity with the vectorization results of multiple engineering drawing names stored in a construction drawing vector database. For a traceability inspection item within the multiple construction inspection items, multiple candidate engineering drawings corresponding to that traceability inspection item are searched from the database. These candidate drawings are then input into a visual language model, and... At least one of the traceability verification and inspection items and a corresponding traceability verification and inspection standard is input into the visual language model; through the visual language model, the content of the plurality of candidate engineering drawings is analyzed, and a final engineering drawing is selected from the plurality of candidate engineering drawings based on at least one of the traceability verification and inspection items and the traceability verification and inspection standard; through the visual language model, based on at least one of the traceability verification and inspection items and the traceability verification and inspection standard, the plurality of engineering final drawings corresponding to the final engineering drawing are analyzed. The process involves selecting and ranking the cut-out diagrams, extracting the best cut-out diagram from the plurality of final cut-out diagrams, comparing the semantic consistency of the content of the best cut-out diagram with the content of the traceability verification standard, and generating a verification comparison result. The vectorization result of the plurality of engineering drawing names is generated by performing word embedding vectorization processing on the names of the plurality of engineering drawings in the construction drawing specification document using the vectorization model. Furthermore, the plurality of final cut-out diagrams are generated by cropping the final engineering drawing.

[0009] In some embodiments, the vectorization results of the plurality of engineering drawing names are generated through the following steps: receiving the construction drawing document, wherein the construction drawing document has the plurality of engineering drawings; inputting the construction drawing document into the vectorization model; and through the vectorization model, performing word embedding vectorization processing on the names of the plurality of engineering drawings respectively, and generating and outputting the vectorization results of the plurality of engineering drawing names corresponding to the names of the plurality of engineering drawings respectively.

[0010] In some embodiments, the method further includes the following steps: inputting a plurality of engineering drawings in the construction drawing file into an object detection and clipping model; performing object detection on the plurality of engineering drawings through the object detection and clipping model, and generating a plurality of object detection results corresponding to the plurality of engineering drawings respectively; and clipping the plurality of engineering drawings into a plurality of engineering clipping sub-drawings corresponding to the plurality of engineering drawings based on the plurality of object detection results through the object detection and clipping model, and outputting the plurality of engineering clipping sub-drawings; wherein the object detection and clipping model is a trained artificial intelligence engine.

[0011] In some embodiments, the step of selecting the final engineering drawing from the plurality of engineering candidate drawings based on at least one of the traceability verification item and the traceability verification standard includes the following sub-steps: analyzing the content of at least one of the traceability verification item and the traceability verification standard, and extracting a location condition information and a detail condition information based on the content of at least one of the traceability verification item and the traceability verification standard; determining whether each of the plurality of engineering candidate drawings satisfies the location condition information and the detail condition information; and determining the engineering candidate drawing that satisfies the location condition information and the detail condition information as the final engineering drawing.

[0012] In some embodiments, the method further includes the following steps: storing the traceability inspection item, the traceability inspection standard, and the best cutting sub-graph as a historical inspection item, a historical inspection standard, and a historical best sub-graph respectively in the construction drawing vector database.

[0013] In some embodiments, the method further includes the following steps: inputting the best cropped sub-image and a plurality of historical best sub-images stored in the construction drawing vector database into a feature extractor; performing feature extraction on each of the best cropped sub-image and the plurality of historical best sub-images through the feature extractor, and generating and outputting a plurality of feature extraction results corresponding to each of the best cropped sub-image and the plurality of historical best sub-images; and retrieving the best cropped sub-image and the plurality of historical best sub-images based on the plurality of feature extraction results. A sub-image is classified into two groups, forming a majority group and a minority group; it is determined whether the best cropped sub-image is the same as the plurality of historical best sub-images; when the best cropped sub-image is different from the plurality of historical best sub-images, it is determined whether the best cropped sub-image belongs to the minority group; and when the best cropped sub-image belongs to the minority group, the plurality of project final cropped sub-images are compared with the historical best sub-images belonging to the majority group, a feature-similar cropped sub-image is extracted from the plurality of project final cropped sub-images, and the feature-similar cropped sub-image is taken as the best cropped sub-image.

[0014] In some embodiments, the method further includes the following steps: when the check comparison result is a content mismatch information, correcting the content of the traceability check standard according to the content of the optimal cropped sub-image, and generating a correction check form.

[0015] Furthermore, this application also provides a computer device for tracing and verifying construction inspection forms, comprising: a storage module configured to store a computer program product; and a processing module configured to be coupled to the storage module; wherein, after loading and executing the computer program product, the processing module is capable of executing any of the methods described in this application for tracing and verifying construction inspection forms.

[0016] Furthermore, this application also provides a non-transitory computer-readable recording medium for tracing and verifying construction inspection forms. After a computer device loads a computer program product stored in the non-transitory computer-readable recording medium and executes the computer program product, the computer device can execute any of the methods described in this application for tracing and verifying construction inspection forms.

[0017] Furthermore, this application also provides a computer program product for tracing and verifying construction inspection forms. After a computer device loads and executes the computer program product, the computer device can execute any of the methods described in this application for tracing and verifying construction inspection forms.

[0018] Accordingly, the technical means provided by this application can produce advantageous effects that were previously unattainable. Specifically, one advantageous effect achieved by this application is the ability to utilize artificial intelligence to trace and verify construction inspection forms, thereby effectively and accurately determining whether the content of the construction inspection forms conforms to the content of the construction drawings. This enables relevant personnel to implement the verification work of the inspection forms, and allows inspectors to correctly conduct project inspections based on construction inspection forms that conform to the content of the construction drawings, thereby ensuring the construction quality of each construction project. In other words, this application not only provides more effective assistance for project management and execution, but also improves the efficiency and accuracy of project supervision.

Implementation Method

[0019] This application will be described in detail through the embodiments and accompanying drawings described below, so as to help those skilled in the art to which this application pertains to understand the purpose, features and effects of this application.

[0020] It should be noted that the various steps described in this application may be performed sequentially, in reverse order, or by appropriately changing or skipping the order during control processing. It should be noted that "the first step may be performed after the second step" described in this application may mean "the first step is performed directly after the second step is performed" and / or "other steps (such as the third step) are performed after the second step is performed, and then the first step is performed."

[0021] Furthermore, in the description of this application, it should be noted that terms such as "first," "second," and "third" are used to distinguish between elements, rather than to limit the elements themselves or indicate a specific order of elements. It should also be noted that in the description below, the same elements or steps may be represented by the same number.

[0022] Furthermore, the term “coupled” as described in this application may mean “directly connected” and / or “indirectly connected”. Specifically, “the first element is configured to be coupled to the second element” may mean “the first element is configured to be directly connected to the second element” and / or “the first element is configured to be indirectly connected to the second element”.

[0023] For the sake of brevity, although the steps in the method for tracing and verifying construction inspection forms described in this application are executed through a single computer device, in some embodiments, the steps may also be executed through multiple computer devices. That is, the steps in the method for tracing and verifying construction inspection forms described in this application may also be implemented through the collaborative operation of multiple computer devices (e.g., a computer device and a remote server).

[0024] Please refer to Figure 1, which is a block diagram illustrating the computer device 200 for tracing and verifying construction inspection forms of this application.

[0025] In some embodiments, the computer device 200 may be a finished product known to those skilled in the art to which this application pertains, specifically such as various models or specifications of desktop computers, notebook computers, laptop computers, tablet computers, or other electronic devices with equivalent configurations, but not limited thereto. The computer device 200 includes a processing module 220 and a storage module 230; furthermore, in some embodiments, the computer device 200 also includes a receiving module 210, an output module 240, and / or a display module 250. Taking FIG1 as an example, the computer device 200 may include a receiving module 210, a processing module 220, a storage module 230, an output module 240, and a display module 250. Hereinafter, each module will be described in more detail.

[0026] The receiving module 210 is configured to receive various data, images, and / or instructions from a remote server (not shown). Taking this application as an example, the receiving module 210 can receive construction drawings and / or construction inspection forms from an electronic device (not shown) such as a scanner via a physical signal line, so as to perform subsequent processing on the received construction drawings and / or construction inspection forms. In some embodiments, the receiving module 210 can be a finished product known to those skilled in the art to which this application pertains, such as various types or specifications of input / output interfaces, but is not limited thereto. In some embodiments, the file format of the construction drawings and / or construction inspection forms received by the computer device 200 can be, for example, a portable document format (PDF), but is not limited thereto. Furthermore, the computer device 200 can also receive construction drawings and / or construction inspection forms from a cloud database (not shown) or a cloud server (not shown) via virtual transmission.

[0027] The processing module 220 is configured to be coupled to the storage module 230 and is configured to perform the steps of any of the methods described in this application for tracing and verifying construction inspection forms. More specifically, after loading and executing a computer program product, the processing module 220 is able to perform the steps of any of the methods described in this application for tracing and verifying construction inspection forms, thereby implementing any of the methods described in this application for tracing and verifying construction inspection forms. In some embodiments, the processing module 220 may be a finished product known to those skilled in the art, such as various models or specifications of central processing units or graphics processors, but is not limited thereto.

[0028] Furthermore, the processing module 220 can be configured to be coupled to the receiving module 210, the output module 240 and the display module 250 so as to receive specific data, images and / or instructions via the receiving module 210, output specific data, images and / or instructions via the output module 240 and display specific data and / or images via the display module 250.

[0029] The storage module 230 is configured to store a computer program product, such that after the processing module 220 loads and executes the stored computer program product, the processing module 220 can execute the various steps of any of the methods described in this application for tracing and verifying construction inspection forms. The computer program product described in this application may include a series of code and / or instruction sets, particularly including specific code and / or instruction sets corresponding to the respective steps of any of the methods described in this application for tracing and verifying construction inspection forms.

[0030] In some embodiments, the storage module 230 may include one or more non-volatile memory and one or more volatile memory. In some embodiments, the volatile memory may be a finished product known to those skilled in the art to which this application pertains, such as various types of dynamic random access memory or static random access memory, but is not limited thereto. In some embodiments, the non-volatile memory may be a finished product known to those skilled in the art to which this application pertains, such as various types of read-only memory or flash memory, but is not limited thereto.

[0031] The output module 240 is configured to output various data, images, and / or instructions to a remote server (not shown), a terminal device (not shown), or a cloud database (not shown). Taking this application as an example, the output module 240 can output the contents of a construction inspection form, the inspection comparison results corresponding to the traceability inspection items and standards in the construction inspection form, the engineering final selection diagram, and / or the optimal cutting sub-diagram, etc., but is not limited thereto. In some embodiments, the output module 240 can be a finished product known to those skilled in the art to which this application pertains, such as various models or specifications of input / output interfaces, but is not limited thereto.

[0032] The display module 250 is configured to display specific data and / or images, etc. Taking this application as an example, the display module 250 can display to the user the contents of the construction inspection form, the inspection comparison results corresponding to the traceability inspection items and traceability inspection standards in the construction inspection form, the engineering final selection diagram and / or the best cutting sub-diagram, etc., but is not limited thereto. In some embodiments, the display module 250 can be a finished product known to those skilled in the art to which this application pertains, such as various models or specifications of displays or display panels, etc., but is not limited thereto.

[0033] Furthermore, taking Figure 1 as an example, the computer device 200 also includes a construction drawing vector database 410 to store the vectorized results of the engineering drawing names, but is not limited thereto. The aforementioned vectorized results of the engineering drawing names refer to the vectorized data generated after word embedding vectorization processing of the names of multiple engineering drawings in the construction drawing document through the vectorization model described later. In addition, the name of the construction project, the name of the construction drawing document, the drawing directory information of the construction drawing document, the names of each engineering drawing, the images of each engineering drawing, and / or the engineering cut-out sub-drawings described later can also be stored in the construction drawing vector database 410. In some embodiments, the construction drawing vector database 410 can be a finished product known to those skilled in the art to which this application pertains, such as Neo4j Graph Database, but is not limited thereto. In addition, the computer device 200 can also read various data stored in the construction drawing vector database 410.

[0034] Furthermore, taking Figure 1 as an example, the computer device 200 also includes a vectorization model 310, a visualization language model 320, and an object detection and clipping model 330, for processing construction drawings and / or construction inspection forms. The following will provide a more detailed explanation of each model.

[0035] The vectorization model 310 is configured to convert the string content in a file into corresponding vectorized results (i.e., high-dimensional vector representations) so that the computer device 200 can understand the string content in the file through the vectorized results. More specifically, taking this application as an example, after the construction drawing documents and construction inspection forms are respectively input into the vectorization model 310, the vectorization model 310 can perform word embedding vectorization processing on the content of the drawing names of each engineering drawing in the construction documents and the content of the construction inspection forms, thereby generating and outputting the corresponding vectorized results, namely, the vectorized results of the engineering drawing names and the vectorized results of the construction inspection forms. With the operating characteristics of the vectorization model 310, the computer device 200 can capture the semantic similarity and correlation between words in the file, thereby improving the accuracy and efficiency of the computer device 200 in understanding the string content in the file. In some embodiments, the vectorization model 310 may be a finished product known to those skilled in the art to which this application pertains, such as OpenAI's text-embedding-ada-002 or open-source Google-bert / bert-base-chinese or shibing624 / text2vec-base-chinese on Hugging Face, but is not limited thereto.

[0036] The visualization language model 320 is configured to analyze the content of a plurality of candidate engineering drawings (described below), and to select a final engineering drawing from the plurality of candidate engineering drawings based on at least one of traceability inspection items and traceability inspection standards. It is also configured to score and rank a plurality of final engineering cut-off sub-drawings corresponding to the final engineering drawing based on at least one of traceability inspection items and traceability inspection standards, and to extract the best cut-off sub-drawing from the plurality of final engineering cut-off sub-drawings. More specifically, taking this application as an example, after the traceability verification standards and at least one of the traceability verification standards, as well as the plurality of candidate engineering drawings, are input into the visualization language model 320, the visualization language model 320 can analyze the content of the traceability verification standards and at least one of the traceability verification standards, as well as the content of the plurality of candidate engineering drawings, and then select the final engineering drawing from the plurality of candidate engineering drawings based on the content of the traceability verification items and at least one of the traceability verification standards; subsequently, the visualization language model 320 can score and rank the plurality of final engineering cut-off sub-drawings corresponding to the final engineering drawing based on the content of the traceability verification items and at least one of the traceability verification standards, and extract the best cut-off sub-drawing from the plurality of final engineering cut-off sub-drawings. Leveraging the characteristics of the visual language model 320, this application can first identify the most relevant engineering drawing from numerous engineering drawings in the construction drawing specification document based on the content of at least one of the traceability verification items and traceability verification standards, and then identify the most relevant detail drawing on that engineering drawing from the most relevant engineering drawing. In some embodiments, the visual language model 320 may be a finished product known to those skilled in the art to which this application pertains, such as OpenAI's gpt-4-vision-preview, but is not limited thereto.

[0037] The object detection and cropping model 330 is configured to detect objects on an image and crop the objects on the image based on the object detection results. The object detection and cropping model 330 is a trained artificial intelligence engine, also known as a trained image recognition model. More specifically, taking this application as an example, after multiple engineering drawings in the construction drawing document are input into the object detection and cropping model 330, the object detection and cropping model 330 can analyze each engineering drawing, detect the objects on each engineering drawing, and generate object detection results corresponding to each engineering drawing. Subsequently, the object detection and cropping model 330 can crop each engineering drawing according to the object detection results of each engineering drawing, thereby generating multiple cropped sub-images corresponding to each engineering drawing. Furthermore, in some embodiments, those skilled in the art to which this application pertains can use a plurality of annotated training drawings as training data sets to train the object detection and clipping model 330, and then use a plurality of test drawings as test data sets to examine the effectiveness of the object detection and clipping model 330. This enables the object detection and clipping model 330, after training and examination, to automatically detect objects on the drawings and automatically clip the drawings into a plurality of clipping sub-drawings based on the detection results. Additionally, in some embodiments, the object detection and clipping model 330 can be a finished product known to those skilled in the art to which this application pertains, such as YOLOv8, but is not limited thereto.

[0038] With the above configuration, the computer device 200 can execute the steps of any of the methods described in this application for tracing and verifying construction inspection forms, so as to implement any of the methods described in this application for tracing and verifying construction inspection forms, and then perform tracing and verification on the construction inspection forms, thereby effectively and accurately determining whether the content of the construction inspection forms conforms to the content of the construction drawing documents.

[0039] Please refer to Figure 2, which is a flowchart illustrating the construction of a construction drawing vector database for traceability verification according to this application. As shown in Figure 2, this flowchart may include steps S210, S220, S230 and S240, and each step can be executed through the processing module 220 of the computer device 200 shown in Figure 1.

[0040] In some embodiments, step S220 may be executed after step S210, step S230 may be executed after step S220, and step S240 may be executed after step S230.

[0041] In step S210, construction drawing documents are received. More specifically, the computer device 200 can receive construction drawing documents from an electronic device such as a scanner or a cloud database, wherein the construction drawing documents have a plurality of engineering drawings associated with the construction project, and these engineering drawings may be, for example, schematic diagrams, configuration diagrams, construction reference diagrams, cross-sectional views, structural diagrams, etc., but are not limited thereto. In addition, one or more detailed views, sectional views, enlarged partial views, diagrams, etc., are also drawn on each engineering drawing, but are not limited thereto. Furthermore, related textual descriptions and / or numerical specifications are also attached to these engineering drawings. In some embodiments, the file format of the construction drawing documents received by the computer device 200 may be, for example, a portable file format, but is not limited thereto.

[0042] In step S220, the construction drawing document is input into the vectorization model. More specifically, the computer device 200 can input the received construction drawing document into the vectorization model (e.g., OpenAI's text-embedding-ada-002 or open-source Google-bert / bert-base-chinese or shibing624 / text2vec-base-chinese on Hugging Face, but not limited to these), so that the vectorization model can perform word embedding vectorization processing on the input construction drawing document, i.e., step S230 described later.

[0043] In step S230, word embedding vectorization processing is performed on the titles of the plurality of engineering drawings, and the vectorization results of the plurality of engineering drawing titles are generated and output. More specifically, after the construction drawing specification document is input into the vectorization model, the vectorization model can perform word embedding vectorization processing on the content of the title of each engineering drawing in the construction drawing specification document, thereby generating and outputting the corresponding results, that is, the vectorization results of the plurality of engineering drawing titles, wherein each of the vectorization results of the plurality of engineering drawing titles corresponds to the respective engineering drawing.

[0044] By leveraging the operational characteristics of the vectorization model (converting each word into a high-dimensional vector representation), the computer device 200 is able to capture the semantic similarity and correlation between the words in the names of each engineering drawing, thereby improving the accuracy and efficiency of the computer device 200 in understanding the string content of the names of each engineering drawing in the construction drawing specification document. In other words, the computer device 200 is better able to understand the content of the names of each engineering drawing in the construction drawing specification document.

[0045] In step S240, the vectorization results of the plurality of engineering drawing names are stored in a construction drawing vector database. More specifically, the computer device 200 may store the vectorization results of each engineering drawing name in a construction drawing vector database (such as Neo4j Graph Database, but not limited thereto).

[0046] Through the operations described above, the computer device 200 can establish a construction drawing vector database, which stores the vectorized results of the engineering drawing names associated with the construction drawing documents, for subsequent traceability verification of construction inspection forms associated with these construction drawing documents. Simultaneously, these steps enable the establishment of independent construction drawing vector databases; that is, each database can store the vectorized results of the engineering drawing names corresponding to each construction drawing document. This allows for the use of independent databases to address various traceability verification needs for different construction drawing documents, ensuring that all relevant drawing content is stored without being affected by the language model's word limit, thus effectively preventing the language model from responding to content exceeding the database's scope. In other words, this helps to treat each construction drawing vector database as an expert familiar with the construction drawing documents, enabling the computer device 200 to utilize each database separately for traceability verification of associated construction inspection forms based on actual traceability verification needs.

[0047] Please refer to Figure 3, which is a flowchart illustrating the process of cutting multiple engineering drawings according to this application. As shown in Figure 3, this flowchart may include steps S310, S320, and S330, and each step may be executed by the processing module 220 of the computer device 200 shown in Figure 1.

[0048] In some embodiments, step S310 may be executed after step S210, step S320 may be executed after step S310, and step S330 may be executed after step S320.

[0049] In step S310, a plurality of engineering drawings are input into the object detection and clipping model. More specifically, the computer device 200 can input a plurality of engineering drawings from the received construction drawing file into the object detection and clipping model (e.g., YOLOv8, but not limited thereto), so that the object detection and clipping model can perform object detection and clipping processing on each of the input engineering drawings, i.e., steps S320 and S330 described later.

[0050] In step S320, object detection is performed on a plurality of engineering drawings, and a plurality of object detection results are generated. More specifically, after a plurality of engineering drawings in the construction drawing file are input into the object detection and clipping model, the object detection and clipping model can perform object detection on each engineering drawing separately, thereby detecting the objects on each engineering drawing, and generating and outputting corresponding results, that is, a plurality of object detection results corresponding to each engineering drawing.

[0051] Taking the first engineering drawing as an example, the object detection and clipping model can analyze the content of the first engineering drawing and then detect various objects on the first engineering drawing, such as the first object, the second object, the third object, etc. The various objects on the first engineering drawing can be, for example, detailed drawings, cross-sectional drawings, enlarged partial drawings, charts, etc., but are not limited to these.

[0052] In step S330, based on the plurality of object detection results, the plurality of engineering drawings are cropped into a plurality of engineering cropping sub-drawings, and the plurality of engineering cropping sub-drawings are output. More specifically, the object detection cropping model can crop each engineering drawing separately based on the plurality of object detection results corresponding to each engineering drawing, thereby cropping a plurality of engineering cropping sub-drawings corresponding to each engineering drawing, and outputting the engineering cropping sub-drawings of each engineering drawing.

[0053] Taking the first engineering drawing as an example, after detecting each object on the first engineering drawing, the object detection and clipping model can continue to clip the first engineering drawing based on the object detection results of the first engineering drawing, and then clip out the first engineering clipping sub-drawing, the second engineering clipping sub-drawing, the third engineering clipping sub-drawing, etc., of the first engineering drawing. These engineering clipping sub-drawings can be, for example, detailed drawings, cross-sectional drawings, enlarged partial drawings, charts, etc., but are not limited to these.

[0054] Since the object detection and clipping model has been trained using a plurality of labeled training engineering drawings as training data set and tested using a plurality of test engineering drawings as test data set, the trained object detection and clipping model can detect each object on the engineering drawing and clip it. Therefore, the computer device 200 can use the object detection and clipping model to automatically, effectively and accurately clip each engineering drawing in the construction drawing document into engineering clipping sub-drawings corresponding to each engineering drawing, thereby reducing the human resources and costs required for this work.

[0055] Furthermore, since each engineering drawing contains various engineering cut-out sub-drawings, and each engineering cut-out sub-drawing contains its own images, text descriptions and / or numerical specifications, cutting each engineering drawing into its own engineering cut-out sub-drawings will help with the subsequent traceability and inspection of the construction inspection form. That is, during the traceability and inspection process, the most relevant engineering cut-out sub-drawings can be found, and then it can be determined whether the contents of the construction inspection form are consistent with the contents of the engineering cut-out sub-drawings.

[0056] Please refer to Figure 4, which is a flowchart illustrating the method for tracing and verifying construction inspection forms according to this application. The method for tracing and verifying construction inspection forms can be executed through the processing module 220 of the computer device 200 shown in Figure 1, and the method may include steps S410, S420, S430, S440, S450, S460, S470 and S480.

[0057] In some embodiments, step S420 may be executed after step S410, step S430 may be executed after step S420, step S440 may be executed after step S430, step S450 may be executed after step S440, step S460 may be executed after step S450, step S470 may be executed after step S460, and step S480 may be executed after step S470.

[0058] It should be noted that the method for tracing and verifying construction inspection forms, as shown in FIG4, may include the steps shown in FIG2, to first establish a construction drawing vector database for tracing and verification by executing the steps shown in FIG2, and then to perform tracing and verification of the construction inspection forms by executing the steps shown in FIG4. In other words, in some embodiments, step S410 may be executed after step S240 shown in FIG2.

[0059] Furthermore, after establishing the construction drawing vector database for traceability verification by executing the steps shown in FIG2, the computer device 200 can repeatedly execute the steps shown in FIG4 to perform traceability verification on multiple construction inspection forms. That is, first, steps S410 to S480 of the first round are executed to perform traceability verification on the first construction inspection form, and then steps S410 to S480 of the second round are executed to perform traceability verification on the second construction inspection form, and so on. In other words, when the vectorized results of multiple engineering drawing names generated after word embedding vectorization of the content of the drawing names of each engineering drawing in the construction drawing document are stored in the construction drawing vector database, the computer device 200 can perform traceability verification on the construction inspection forms associated with this construction drawing document by executing the steps shown in FIG4.

[0060] Alternatively, the method for tracing and verifying construction inspection forms as shown in FIG4 may also include the steps shown in FIG3, wherein each engineering drawing in the construction drawing document is first cut out by executing the steps shown in FIG3, and then the construction inspection forms are traced and verified by executing the steps shown in FIG4. In other words, in some embodiments, step S410 may be executed after step S330 as shown in FIG3.

[0061] In step S410, a construction inspection form is received. More specifically, the computer device 200 may receive the construction inspection form from an electronic device such as a scanner or a cloud database, wherein the construction inspection form has a plurality of construction inspection items and a plurality of construction inspection standards, and each construction inspection standard corresponds to a respective construction inspection item, that is, a first construction inspection standard corresponds to a first construction item, and a second construction standard corresponds to a second construction item, and so on. In some embodiments, the file format of the construction inspection form received by the computer device 200 may be, for example, a portable file format, but is not limited thereto.

[0062] Furthermore, since the content of the construction inspection form originates from the construction drawing document, the construction inspection form that the computer device 200 needs to perform traceability verification is associated with the construction drawing document. That is, the computer device 200 can receive the first construction inspection form associated with the first construction drawing document and then trace the content of the first construction inspection form back to the first construction drawing document for verification.

[0063] Furthermore, during the traceability verification process, the computer device 200 can perform traceability verification on each construction inspection item and construction inspection standard in the construction inspection form. For example, the computer device 200 can first use the first construction inspection item and the first construction inspection standard as the traceability verification inspection item and traceability verification standard described later, and then use the second construction inspection item and the second construction inspection standard as the traceability verification inspection item and traceability verification standard described later, and so on.

[0064] In step S420, the construction inspection form is input into the vectorization model. More specifically, the computer device 200 can input the received construction inspection form into the vectorization model (e.g., OpenAI's text-embedding-ada-002 or open-source Google-bert / bert-base-chinese or shibing624 / text2vec-base-chinese on Hugging Face, but not limited to these), so that the vectorization model can perform word embedding vectorization processing on the input construction inspection form, i.e., step S430 described later.

[0065] In step S430, word embedding vectorization processing is performed on the content of the construction inspection form, and the construction inspection vectorization result is generated and output. More specifically, after the construction inspection form is input into the vectorization model, the vectorization model can perform word embedding vectorization processing on the content of the construction inspection form, thereby generating and outputting the corresponding result, that is, the construction inspection vectorization result.

[0066] In some embodiments, the vectorization model can perform word embedding vectorization processing on the content of multiple construction inspection items and multiple construction inspection standards in the construction inspection form, and then generate and output corresponding results, namely, multiple inspection item vectorization results and multiple inspection standard vectorization results, wherein each of the multiple inspection item vectorization results corresponds to the content of each construction inspection item, and each of the multiple inspection standard vectorization results corresponds to the content of each construction inspection standard.

[0067] By leveraging the operational characteristics of the vectorization model (converting each word into a high-dimensional vector representation), the computer device 200 is able to capture the semantic similarity and correlation between words in the input construction inspection form, thereby improving the accuracy and efficiency of the computer device 200 in understanding the string content in the construction inspection form. That is, the computer device 200 is better able to understand the content of each construction inspection item and each construction inspection standard in the construction inspection form.

[0068] Furthermore, since the titles of the various engineering drawings in the construction drawing specifications and the construction inspection forms are processed by word embedding vectorization through the same vectorization model, the computer device 200 can have similar or identical vectorization results for the similarities or similarities between the content of the titles of the various engineering drawings in the construction drawing specifications and the content of the construction inspection forms. That is, the computer device 200 can have similar or identical understanding of similar or identical content, which helps to improve the accuracy of tracing the construction inspection forms back to the construction drawing specifications for verification.

[0069] In step S440, the similarity comparison is performed between the construction inspection vectorization result and the vectorization results of multiple engineering drawing names stored in the construction drawing vector database, and multiple candidate engineering drawings are searched from the construction drawing vector database. More specifically, the computer device 200 can perform a similarity comparison between the construction inspection vectorization result and the vectorization results of multiple engineering drawing names stored in the construction drawing vector database, that is, perform a comparison by correlation search, and then calculate the ones with the shortest distance in the vector space (Top K), such as the top 10 with the shortest distance (Top 10), but not limited to this; subsequently, the computer device 200 can know the vectorization results of the top 10 engineering drawing names with the shortest distance based on the similarity comparison result, and then search for multiple engineering drawings corresponding to them from the construction drawing vector database, and use the searched multiple engineering drawings as candidate engineering drawings. In some embodiments, the aforementioned similarity comparison may be conducted in a manner known to those skilled in the art to which this application pertains, such as Euclidean distance (also known as L2 distance), but is not limited thereto.

[0070] Taking the traceability inspection item as an example, the computer device 200 can compare the similarity between the vectorized result of the inspection item and / or the vectorized result of the inspection standard and the vectorized result of the multiple engineering drawing names stored in the construction drawing vector database, so as to calculate the vectorized result of the multiple engineering drawing names that are closer in the vector space for the traceability inspection item in the construction inspection form, and then search for the corresponding multiple engineering drawings from the construction drawing vector database, and use these engineering drawings as multiple engineering candidate drawings corresponding to the traceability inspection item.

[0071] In step S450, a plurality of candidate engineering images are input into a visual language model, and at least one of the traceability verification items and traceability verification criteria is input into the visual language model. More specifically, the computer device 200 can input at least one of the traceability verification items and traceability verification criteria, as well as a plurality of candidate engineering images corresponding to the traceability verification items, into a visual language model (e.g., OpenAI's gpt-4-vision-preview, but not limited thereto), so that the visual language model can perform image analysis processing on the input plurality of candidate engineering images based on at least one of the traceability verification items and traceability verification criteria, i.e., steps S460 and S470 described later.

[0072] In step S460, the content of a plurality of candidate engineering drawings is analyzed, and the final engineering drawing is selected from the plurality of candidate engineering drawings based on at least one of the traceability inspection items and traceability inspection standards. More specifically, after at least one of the traceability inspection items and traceability inspection standards, and the plurality of candidate engineering drawings corresponding to the traceability inspection items, are input into the visual language model, the visual language model can analyze at least one of the traceability inspection items and traceability inspection standards, and then extract key information (such as location condition information or detailed condition information, but not limited to this) based on the content of at least one of the traceability inspection items and traceability inspection standards; subsequently, the visual language model can analyze the content of each candidate engineering drawing, and use the aforementioned key information as a screening condition to determine whether the content of each engineering drawing meets the aforementioned screening condition, and then select the engineering drawing that meets the screening condition as the final engineering drawing. Alternatively, in some embodiments, step S460 can also be implemented by executing the various sub-steps shown in FIG5, but is not limited thereto. Taking the Top 10 engineering candidate images as an example, the visualization language model can first judge the first engineering candidate image, then judge the second engineering candidate image, and so on.

[0073] In step S470, a plurality of engineering final selection cut-off sub-graphs are scored and ranked according to at least one of the traceability inspection items and traceability inspection standards, and the best cut-off sub-graph is extracted from the plurality of engineering final selection cut-off sub-graphs. More specifically, the visual language model can analyze at least one of the traceability inspection items and traceability inspection standards, and then extract scoring condition information based on the content of at least one of the traceability inspection items and traceability inspection standards; subsequently, the visual language model can analyze the content of the plurality of engineering final selection cut-off sub-graphs on the engineering final selection graph, and score and rank each engineering final selection cut-off sub-graph using the aforementioned scoring condition information, and then extract the engineering final selection cut-off sub-graph with the highest score from the plurality of engineering final selection cut-off sub-graphs based on the scoring results, and take the engineering final selection cut-off sub-graph with the highest score as the best cut-off sub-graph.

[0074] For example, the computer device 200 can use prompting words to teach a visual language model to break down the traceability inspection items and standards into multiple independent sub-conditions. This allows the visual language model to score whether each sub-condition is satisfied by a plurality of engineering selection cutout subgraphs (Top K), and to average the scores of each sub-condition to represent the score of each engineering selection cutout subgraph. Based on this score, the plurality of engineering selection cutout subgraphs are reordered to find the engineering selection cutout subgraph with the highest score, which is then selected as the optimal cutout subgraph. This approach helps improve the accuracy of the content in the traceability inspection construction form.

[0075] In step S480, the content of the optimal cropped sub-image is compared with the content of the traceability verification standard for semantic consistency, and a verification comparison result is generated. More specifically, the computer device 200 can compare the content of the corresponding optimal cropped sub-images with the content of the traceability verification standard for semantic consistency, and generate a verification comparison result. The verification comparison result can be content matching information or content non-matching information, and the content matching information and content non-matching information can be represented by flag bits, for example, a high logic level ("1") indicates that the content of the traceability verification standard conforms to the content of the construction drawing document, and a low logic level ("0") indicates that the content of the traceability verification standard does not conform to the content of the construction drawing document, but is not limited thereto. In some embodiments, when the semantics of the traceability verification standard fully conforms to the requirements of the construction drawing document, the computer device 200 can generate a verification comparison result of content matching information, but is not limited thereto. By comparing semantic consistency, the computer device 200 can identify whether there are semantic and / or numerical discrepancies between the content of the optimal cropped subgraph and the content of the source code verification standard. In some embodiments, the aforementioned semantic consistency comparison can be performed in a manner known to those skilled in the art, such as teaching large language models to perform semantic consistency comparison using Chain of Thought (CoT), Few-Shot Learning, Prompt Engineering, or Agentic Reasoning, but is not limited to these methods.

[0076] Through the operation of the above steps, the computer device 200 can perform traceability verification on the contents of the construction inspection form, thereby effectively and accurately determining whether the contents of the construction inspection form conform to the contents of the associated construction drawing documents. This enables relevant personnel to use artificial intelligence to carry out the inspection work of the construction inspection form, and allows inspectors to correctly conduct project inspections based on the construction inspection form that conforms to the contents of the construction drawing documents, thereby ensuring the construction quality of each construction project.

[0077] Furthermore, in some embodiments, the method shown in FIG4 further includes the following step: storing the traceability inspection items, traceability inspection standards, and optimal cutting sub-graphs as historical inspection items, historical inspection standards, and historical optimal sub-graphs respectively in a construction drawing vector database. This step can be performed after, for example, step S480 as shown in FIG4. This step helps to examine the accuracy and usability of the method for traceability inspection of construction inspection forms described in this application using the characteristics of the database, and allows the computer device 200 to use the data in the database to determine whether there are any anomalies in the results generated after each traceability inspection.

[0078] Furthermore, in some embodiments, the method shown in FIG4 further includes the following steps: when the check comparison result is content discrepancy information, the content of the traceability check standard is corrected according to the content of the optimal cropped sub-image, and a correction check form is generated. This step can be performed after, for example, step S480 as shown in FIG4. More specifically, when the check comparison result is content discrepancy information, this indicates that the content of the traceability check standard in the construction check form does not conform to the content in the construction drawing document, and the computer device 200 can correct the non-conforming content in the traceability check standard, that is, cover the non-conforming content in the traceability check standard with the content of the optimal cropped sub-image, thereby generating a new construction check form, and using the new construction check form as the correction check form. This step utilizes accurate information from the construction drawings to correct any discrepancies in the construction inspection forms. This facilitates the provision of corrected construction inspection forms to inspectors after traceability verification, thereby improving user convenience and enabling inspectors to conduct accurate project inspections based on construction inspection forms that align with the construction drawings. Ultimately, this ensures the construction quality of each project.

[0079] Please refer to Figure 5, which is a detailed flowchart illustrating the process of selecting a final engineering drawing from a plurality of candidate engineering drawings based on at least one of traceability inspection items and traceability inspection standards. As shown in Figure 5, this flowchart may include sub-steps S510, S520, and S530, and each sub-step can be executed by the processing module 220 of the computer device 200 shown in Figure 1. Furthermore, step S460 shown in Figure 4 can be completed by executing the sub-steps shown in Figure 5.

[0080] In some embodiments, sub-step S510 may be executed after step S450, sub-step S520 may be executed after sub-step S510, and sub-step S530 may be executed after sub-step S520.

[0081] In sub-step S510, location condition information and detailed condition information are extracted. More specifically, after at least one of the traceability inspection items and traceability inspection standards, and a plurality of candidate engineering drawings corresponding to the traceability inspection items, are input into the visualization language model, the visualization language model can analyze at least one of the traceability inspection items and traceability inspection standards, and then extract location condition information and detailed condition information based on the content of at least one of the traceability inspection items and traceability inspection standards. The detailed condition information can be, for example, various engineering inspection values ​​such as rebar size, rebar spacing, rebar quantity, weld length, pile elevation, number of spacers, spacer spacing, excavation depth, and casing size, but is not limited thereto. In some embodiments, as is known to those skilled in the art to which this application pertains, location condition information and detailed condition information can be extracted by teaching large language models reasoning, for example, using thought chains, providing examples with a small number of samples, prompting engineering or role settings, but is not limited thereto.

[0082] In sub-step S520, it is determined whether each of the plurality of candidate engineering drawings satisfies the positional condition information and the detail condition information. More specifically, after extracting the positional condition information and the detail condition information, the visualization language model can determine whether the content of each candidate engineering drawing satisfies the positional condition information and the detail condition information one by one, and generate the judgment result for each candidate engineering drawing respectively. In some embodiments, the visualization language model can first determine whether the candidate engineering drawing satisfies the positional condition information, and then determine whether the candidate engineering drawing that satisfies the positional condition information also satisfies the detail condition information. Unlike simultaneously determining whether the positional condition information and the detail condition information are satisfied, this approach helps to reduce the amount of computation required by the visualization language model and reduce the judgment time required by the visualization language model.

[0083] In sub-step S530, the candidate engineering drawing that satisfies both the location condition information and the detail condition information is determined as the final engineering drawing. More specifically, the visualization language model can find the candidate engineering drawing that satisfies both the location condition information and the detail condition information from among many candidate engineering drawings based on the judgment results of each candidate engineering drawing, and determine this candidate engineering drawing as the final engineering drawing.

[0084] Since location condition information can reflect whether a component appears on the engineering candidate drawing, and detail condition information can reflect whether the relevant specifications of a component appear on the engineering candidate drawing, using location condition information and detail condition information to screen engineering candidate drawings will help to select the most relevant engineering candidate drawings from many engineering candidate drawings, thereby increasing the accuracy of traceability verification.

[0085] Taking the location condition information "main bridge section P110" as an example, the visual language model can use this location judgment information to determine whether the content of each engineering candidate drawing contains "main bridge section P110". When the engineering candidate drawing contains "main bridge section P110", the visual language model will determine that the engineering candidate drawing meets the location condition information, and the visual language model will then determine whether the engineering candidate drawing also meets the detailed condition information; conversely, when the engineering candidate drawing does not contain "main bridge section P110", the visual language model will determine that the engineering candidate drawing does not meet the location condition information, and the visual language model will then determine whether the content of "main bridge section P110" appears on another engineering candidate drawing.

[0086] Furthermore, taking the detailed condition information "pile diameter and length: L=55m; D=2m" as an example, the visual language model can determine whether the content "pile diameter and length: L=55m; D=2m" appears on the candidate engineering drawings that meet the location condition information. When the candidate engineering drawing contains "pile diameter and length: L=55m; D=2m", the visual language model will determine that this candidate engineering drawing meets both the location condition information and the detailed condition information, and the visual language model will identify this candidate engineering drawing as the most relevant candidate engineering drawing, i.e., the final engineering drawing; conversely, when the candidate engineering drawing does not contain "pile diameter and length: L=55m; D=2m", the visual language model will determine that this candidate engineering drawing does not meet the detailed condition information, and the visual language model will continue to determine whether another candidate engineering drawing meets both the location condition information and the detailed condition information.

[0087] Please refer to Figure 6, which is a partial flowchart illustrating the method for tracing and verifying construction inspection forms according to this application. As shown in Figure 6, this flowchart may include steps S610, S620, S630, S640, S645, S650, S655 and S660, and each step can be executed through the processing module 220 of the computer device 200 shown in Figure 1.

[0088] In some embodiments, step S610 may be executed after step S470, step S620 may be executed after step S610, step S630 may be executed after step S620, step S640 may be executed after step S630, steps S645 and S650 may be executed after step S640, and steps S655 and S660 may be executed after step S650.

[0089] In step S610, the best cropped sub-image and a plurality of historical best sub-images are input into the feature extractor. More specifically, the computer device 200 can input the best cropped sub-image extracted during this source tracing check and a plurality of historical best sub-images stored in the construction drawing vector database into the feature extractor, so that the feature extractor can perform feature extraction on the input best cropped sub-image and each historical best sub-image, i.e., step S620 described later. In some embodiments, the feature extractor can be a finished product known to those skilled in the art to which this application pertains, such as ResNet50, but is not limited thereto.

[0090] In step S620, feature extraction is performed on the best cropped sub-image and the plurality of historical best sub-images respectively, and feature extraction results are generated and output. More specifically, after the best cropped sub-image and each historical best sub-image are input into the feature extractor, the feature extractor can perform feature extraction on the best cropped sub-image and each historical best sub-image respectively, and generate and output a plurality of feature extraction results corresponding to the best cropped sub-image and each historical best sub-image respectively.

[0091] In step S630, the best cropped sub-image and the multiple historical best sub-images are classified into two groups based on the results of multiple feature extractions. More specifically, the computer device 200 can classify the best cropped sub-image and the multiple historical best sub-images according to the results of each feature extraction (K-means), for example, classifying the best cropped sub-image and the multiple historical best sub-images into two groups (2-means), forming a majority group and a minority group.

[0092] In step S640, it is determined whether the best cropped sub-image is the same as a plurality of historical best sub-images. More specifically, the computer device 200 can perform image comparison between the best cropped sub-image and each historical best sub-image to determine whether the best cropped sub-image is the same as each historical best sub-image. In some embodiments, the aforementioned image comparison can be performed in a manner known to those skilled in the art, such as obtaining the respective feature vectors through image recognition models such as ResNet50 or ResNet101, and then using the L2 distance between the feature vectors of each sub-image to measure the similarity between images to achieve the aforementioned image comparison, but is not limited to this.

[0093] It should be noted that the method for tracing and verifying construction inspection forms described in this application can be used for batch processing of multiple construction inspection forms, that is, tracing and verifying each construction inspection form separately. In this case, the historical best sub-graph refers to the best trimmed sub-graph generated during the previous tracing and verification. Since the best trimmed sub-graph and each historical best sub-graph are the results of similar or identical tracing and verification items and standards for different construction inspection forms, the best trimmed sub-graph generated during this tracing and verification should be the same as each historical best sub-graph.

[0094] When the best cropped sub-image is the same as each of the historical best sub-images (i.e., the judgment result of step S640 is "yes"), the computer device 200 can continue to execute step S645; and when the best cropped sub-image is different from each of the historical best sub-images (i.e., the judgment result of step S640 is "no"), the computer device 200 can continue to execute step S650.

[0095] In step S645, the best cropped sub-image is determined to be the most relevant image. More specifically, when the computer device 200 determines that the best cropped sub-image generated during this source tracing check is the same as each of the historical best sub-images stored in the construction drawing vector database, this indicates that the randomness of the visualization language model has not affected this source tracing check, that is, the visualization language model has correctly extracted the best cropped sub-image. In this case, the computer device 200 can perform subsequent processing on the checked best cropped sub-image, i.e., step S480 as shown in FIG4.

[0096] In step S650, it is determined whether the best cropped sub-image belongs to a minority group. More specifically, when the computer device 200 determines that the best cropped sub-image generated during this source tracing check is different from the various historical best sub-images stored in the construction drawing vector database, this indicates that the randomness of the visualization language model may affect this source tracing check, that is, the visualization language model may extract an abnormal best cropped sub-image. In this case, the computer device 200 will further examine whether the best cropped sub-image generated during this source tracing check belongs to a minority group, that is, whether the feature extraction result of the best cropped sub-image generated during this source tracing check is different from the feature extraction results of many historical best sub-images.

[0097] When the best cropped sub-image does not belong to the minority group (i.e., the judgment result of step S650 is "no"), the computer device 200 may continue to execute step S655; while when the best cropped sub-image belongs to the minority group (i.e., the judgment result of step S650 is "yes"), the computer device 200 may continue to execute step S660.

[0098] In step S655, the best cropped sub-image is determined to be the most relevant image. More specifically, when the computer device 200 determines that the best cropped sub-image generated during this source tracing check belongs to the majority group rather than the minority group, this indicates that the feature extraction result of the best cropped sub-image generated during this source tracing check is the same as or highly similar to the feature extraction results of many historical best sub-images, that is, the visualization language model correctly extracts the best cropped sub-image. In this case, the computer device 200 can perform subsequent processing on the best cropped sub-image after double review, that is, step S480 as shown in FIG4.

[0099] In step S660, the similarity of a plurality of project selection cropping sub-images is compared with that of historical best sub-images belonging to the majority group, and feature-similar cropping sub-images are extracted from the plurality of project selection cropping sub-images. More specifically, when the computer device 200 determines that the best cropping sub-image generated during this source tracing check belongs to the minority group, this indicates that the feature extraction result of the best cropping sub-image generated during this source tracing check is different from the feature extraction results of many historical best sub-images, that is, the visualization language model extracts an abnormal best cropping sub-image. In this case, the computer device 200 will re-extract a new best cropping sub-image from many project selection cropping sub-images, that is, compare the similarity of each project selection cropping sub-image with each historical best sub-image belonging to the majority group, and then extract feature-similar cropping sub-images from many project selection cropping sub-images, and use this feature-similar cropping sub-image as the new best cropping sub-image. In some embodiments, the aforementioned similarity comparison may be conducted in a manner known to those skilled in the art to which this application pertains, such as L2 distance, but is not limited thereto.

[0100] By performing the above steps, the computer device 200 can self-check whether there are any abnormalities in the best cropped sub-image extracted by the visualization language model, thereby avoiding the randomness of the visualization language model from extracting the wrong best cropped sub-image, thereby improving the accuracy and usability of the method for tracing and verifying construction inspection forms described in this application.

[0101] In some embodiments, those skilled in the art to which this application pertains may use integrated development environments (IDEs) such as Jupyter Notebook, Visual Studio Code, PyCharm, and Sublime Text; programming languages ​​such as Python and PyTorch; and libraries and toolkits such as chroma-hnswlib, chromadb, chromadbx, fitz, gradio, langchain, langchain_community, neo4j, openai, Pillow, roboflow, sentence-transformers, tokenizers, tqdm, and transformers to implement any of the methods described in this application for tracing and verifying construction inspection forms.

[0102] Please refer to Figures 7, 8A, 8B, and Table (I). Figure 7 is a schematic diagram illustrating the results of searching and filtering the final engineering drawing 700 from multiple engineering drawings in the construction drawing documentation according to this application. Figures 8A and 8B are schematic diagrams illustrating the results of extracting the first optimal cutting sub-drawing 810 and the second optimal cutting sub-drawing 820 from multiple final engineering cutting sub-drawings in the final engineering drawing 700 shown in Figure 7. Table (I) illustrates the results of verifying the construction inspection form by tracing its origin back to the construction specification text using the method of this application for tracing back to the source of the construction inspection form. Table (I) Construction inspection items Construction Inspection Standards Best cropping sub-image Verification and comparison results Drilling operations - pile bottom elevation and pile length L~L+D / 2 L=55m; D=2m [Design Drawings of Complete Piling Foundations for Tamkang Bridge - P3 Detailed Drawings of Cast-in-Place Foundation Piles for Main Bridge Section - Reinforcement Table for Foundation Piles] Foundation Number: P110; Pile Diameter: D 200; Pile Length: PL 5500 Content matches information Reinforcement assembly - lap length According to the construction drawings, ≥150cm [Design Drawings of Complete Piling Foundations for Tamkang Bridge - P3 Main Bridge Section Foundation Cast-in-Place Piling Details - Main Reinforcement Splice Details] Main reinforcement lap length Ls - 180cm Content does not match information

[0103] In some embodiments, the steps in the method for tracing and verifying construction inspection forms described in this application may be further combined, replaced, repeatedly performed and / or modified to produce new embodiments without departing from the scope disclosed in this application.

[0104] In some embodiments, the steps of the method for tracing and verifying construction inspection forms described in this application may be stored in a non-transitory computer-readable recording medium, such as a hard disk, optical disk, magnetic disk, USB flash drive, or a database accessible via a network, but not limited thereto. After the non-transitory computer-readable recording medium loads a computer program product into memory through a computer device and executes the computer program product, the computer device is able to implement any of the methods of the method for tracing and verifying construction inspection forms described in this application.

[0105] In some embodiments, the computer program product for tracing and verifying construction inspection forms described in this application may include a series of code and / or instruction sets, particularly specific code and / or instruction sets corresponding to each step of any of the methods for tracing and verifying construction inspection forms described in this application, so that after the computer device loads and executes the computer program product, the computer device can implement any of the methods for tracing and verifying construction inspection forms described in this application.

[0106] This application has been further described through the above embodiments and accompanying drawings. However, those skilled in the art to which this application pertains can still make many modifications and changes without departing from the scope and spirit set forth in the claims of this application. Therefore, the scope of protection of this application should still be determined by the claims of the patent applications and should not be limited by the content disclosed in the specification. [Simplified Explanation of the Diagram]

[0107] Figure 1 is a block diagram illustrating the computer device for traceability verification of construction inspection forms according to this application. Figure 2 is a flowchart illustrating the construction of a construction drawing vector database for traceability verification according to this application. Figure 3 is a flowchart illustrating the cropping of multiple engineering drawings according to this application. Figure 4 is a flowchart illustrating the method for traceability verification of construction inspection forms according to this application. Figure 5 is a detailed flowchart illustrating the selection of a final engineering drawing from multiple candidate engineering drawings based on at least one of traceability verification items and traceability verification standards according to this application. Figure 6 is a partial flowchart illustrating the method for traceability verification of construction inspection forms according to this application. Figure 7 is a schematic diagram illustrating the results of searching and cropping a final engineering drawing from multiple engineering drawings in the construction drawing file according to this application. Figures 8A and 8B are schematic diagrams illustrating the results of extracting a first optimal cropping sub-drawing and a second optimal cropping sub-drawing from multiple cropping sub-drawings of the final engineering drawing shown in Figure 7.

Claims

1. A method for tracing and verifying a construction inspection form, the method being executed after loading and executing a computer program product via a computer device, the method comprising the following steps: (a) receiving a construction inspection form, wherein the construction inspection form is associated with a construction drawing document, and the construction inspection form has a plurality of construction inspection items and a plurality of construction inspection standards corresponding to the plurality of construction inspection items; (b) inputting the construction inspection form into a vectorization model; (c) performing word embedding vectorization processing on the content of the construction inspection form through the vectorization model, and generating and outputting a construction inspection vectorization result corresponding to the construction inspection form; (d) comparing the similarity of the construction inspection vectorization result with the vectorization results of a plurality of engineering drawing names stored in a construction drawing vector database, and searching the construction drawing vector database for a traceability verification inspection item among the plurality of construction inspection items to find a plurality of candidate engineering drawings corresponding to the traceability verification inspection item; (e) (f) Input the plurality of candidate engineering images into a visual language model, and input at least one of the traceability verification item and a traceability verification standard corresponding to the traceability verification item into the visual language model; (g) Analyze the content of the plurality of candidate engineering images through the visual language model, and select a final engineering image from the plurality of candidate engineering images based on at least one of the traceability verification item and the traceability verification standard; (h) Through the visual language model, score and rank the plurality of final engineering image cutouts corresponding to the final engineering image based on at least one of the traceability verification item and the traceability verification standard, and extract an optimal cutout image from the plurality of final engineering image cutouts; and (h) Perform a semantic consistency comparison between the content of the optimal cutout image and the content of the traceability verification standard, and generate a verification comparison result; wherein, The vectorization result of the multiple engineering drawing names is the result of word embedding vectorization processing of the names of multiple engineering drawings in the construction drawing document through the vectorization model; and the multiple engineering final selection cropped sub-drawings are the drawing generated by cropping the engineering final selection drawing.

2. The method as described in request item 1, wherein, The vectorization results of the multiple engineering drawing names are generated through the following steps: receiving the construction drawing document, wherein the construction drawing document has the multiple engineering drawings; inputting the construction drawing document into the vectorization model; and through the vectorization model, performing word embedding vectorization processing on the drawing names of the multiple engineering drawings respectively, and generating and outputting the vectorization results of the multiple engineering drawing names corresponding to the drawing names of the multiple engineering drawings respectively.

3. The method as described in request item 1, wherein, The method also includes the following steps: inputting a plurality of engineering drawings from the construction drawing document into an object detection and clipping model; performing object detection on the plurality of engineering drawings through the object detection and clipping model, and generating a plurality of object detection results corresponding to the plurality of engineering drawings respectively; And through the object detection and clipping model, based on the multiple object detection results, the multiple engineering drawings are clipped into multiple engineering clipping sub-drawings corresponding to the multiple engineering drawings, and the multiple engineering clipping sub-drawings are output; wherein, the object detection and clipping model is a trained artificial intelligence engine.

4. The method as described in claim 1, wherein the step of selecting the final engineering drawing from the plurality of engineering candidate drawings based on at least one of the traceability verification item and the traceability verification standard comprises the following sub-steps: analyzing the content of at least one of the traceability verification item and the traceability verification standard, and extracting a location condition information and a detail condition information based on the content of at least one of the traceability verification item and the traceability verification standard; determining whether each of the plurality of engineering candidate drawings satisfies the location condition information and the detail condition information; and determining the engineering candidate drawing that satisfies the location condition information and the detail condition information as the final engineering drawing.

5. The method as described in request item 1, wherein, The method also includes the following steps: storing the traceability inspection item, the traceability inspection standard, and the optimal cutting sub-graph as a historical inspection item, a historical inspection standard, and a historical optimal sub-graph respectively in the construction drawing vector database.

6. The method as described in request item 1, wherein, The method further includes the following steps: inputting the optimal cropped sub-image and a plurality of historical optimal sub-images stored in the construction drawing vector database into a feature extractor; using the feature extractor, performing feature extraction for each of the optimal cropped sub-image and the plurality of historical optimal sub-images, and generating and outputting a plurality of feature extraction results corresponding to each of the optimal cropped sub-image and the plurality of historical optimal sub-images; classifying each of the optimal cropped sub-image and the plurality of historical optimal sub-images into two groups based on the plurality of feature extraction results, forming a majority group and a minority group; and determining whether the optimal cropped sub-image is the same as the plurality of historical optimal sub-images. When the optimal cropping subgraph is different from the plurality of historical optimal subgraphs, it is determined whether the optimal cropping subgraph belongs to the minority group; and when the optimal cropping subgraph belongs to the minority group, the plurality of project selection cropping subgraphs are compared with the historical optimal subgraphs belonging to the majority group, a feature-similar cropping subgraph is extracted from the plurality of project selection cropping subgraphs, and the feature-similar cropping subgraph is used as the optimal cropping subgraph.

7. The method as described in request item 1, wherein, The method also includes the following steps: When the verification result is a content discrepancy, the content of the traceability verification standard is corrected according to the content of the optimal cropped sub-image, and a correction verification form is generated.

8. A computer device for tracing and verifying construction inspection forms, comprising: a storage module configured to store a computer program product; and a processing module configured to be coupled to the storage module; wherein, After loading and executing the computer program product, the processing module is able to perform the method for tracing and verifying construction inspection forms as described in any of requests 1 to 7.

9. A non-transitory computer-readable recording medium for tracing and verifying construction inspection forms, wherein after a computer device loads and executes a computer program product stored in the non-transitory computer-readable recording medium, the computer device is capable of performing the method for tracing and verifying construction inspection forms as described in any one of claims 1 to 7.

10. A computer program product for tracing and verifying construction inspection forms, wherein after the computer program product is loaded and executed on a computer device, the computer device is able to perform the method for tracing and verifying construction inspection forms as described in any one of claims 1 to 7.

Images