An artificial intelligence-based construction engineering industry knowledge graph construction method

Abstract

The application discloses a kind of based on artificial intelligence to building engineering industry knowledge graph construction method, obtains the unstructured file of building engineering industry, and according to unstructured file according to file type it is classified as image file or text file;Based on image file, by the mode of image processing, the logical relationship between entity and entity in each image file is obtained;Based on text file, by the mode of semantic recognition, the logical relationship between entity and entity in each text file is obtained;Based on the logical relationship between entity in image file and the logical relationship between entity in text file, constructs knowledge graph.The present application is based on the large amount of unstructured file of existing building engineering industry, using the mode of semantic recognition and image processing, carries out extraction to unstructured file, obtains unified logical relationship, and constructs knowledge graph based on these logical relationships, so that the constructed knowledge graph is perfectly fitted in building engineering industry.

Description

A method for constructing knowledge graphs in the construction engineering industry based on artificial intelligence Technical Field

[0001] This invention relates to the field of building engineering technology, and in particular to a method for constructing a knowledge graph for the building engineering industry based on artificial intelligence. Background Technology

[0002] With the increasing informatization of the construction industry, enterprises have accumulated a large amount of highly unstructured and specialized knowledge data, such as process manuals, construction plans, and acceptance reports. Traditional methods rely on manual processing, resulting in serious knowledge silos, low retrieval efficiency, lagging knowledge updates, and low levels of intelligence. In recent years, although NLP and knowledge graph technologies have made some progress, general models often suffer from semantic bias and incorrect output knowledge information when dealing with the professional logic of construction engineering due to inadequate processing of structured data, and cannot fully adapt to the construction engineering industry. Summary of the Invention

[0003] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence. Based on a large number of unstructured documents in the existing construction engineering industry, semantic recognition and image processing are used to extract the unstructured documents, obtain unified logical relationships, and construct a knowledge graph based on these logical relationships, so that the constructed knowledge graph is perfectly aligned with the construction engineering industry.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, comprising: Step 1: acquiring unstructured files of the construction engineering industry, and classifying them into image files or text files according to their file type; Step 2: based on the image files, obtaining the entities and logical relationships between entities in each image file through image processing; Step 3: based on the text files, obtaining the entities and logical relationships between entities in each text file through semantic recognition; Step 4: constructing a knowledge graph based on the logical relationships between entities in the image files and the logical relationships between entities in the text files.

[0005] Furthermore, in step 1, after obtaining unstructured files from the construction industry and before classifying the unstructured files into image files or text files according to file type, the process includes: obtaining the content of the unstructured files; and decomposing the unstructured files into multiple image files and text files.

[0006] Furthermore, between step 1 and step 2, the method further includes: converting the image files to a format such that all image files have the same file extension; and converting the text files to a format such that all text files have the same file extension.

[0007] Furthermore, in step 2, the step of obtaining entities and logical relationships between entities in each image file through image processing includes: processing the image file into black and white lines so that each pixel in the image file is only black and white, wherein the black pixels form the lines in the image file; determining the entities in the image file based on the shape of the lines in the image file; and determining the logical relationships between entities based on the distance between entities in the image file.

[0008] Furthermore, determining the entities in the image file based on the shape of the lines in the image file includes: establishing a two-dimensional coordinate system; representing the image file in the two-dimensional coordinate system after the black and white lines are converted into lines; cutting each line at its inflection point to obtain multiple line segments; representing the coordinates of each pixel of each line segment using a set, and normalizing the set of each line segment; obtaining the corresponding entities based on the sets of each line segment using an entity database; the entity database is used to store entities and one or more sets corresponding to the entities.

[0009] Furthermore, determining the logical relationship between entities based on the distance between entities in the image file includes: obtaining the image size of the entities in the image file and the image distance between entities; obtaining the actual size of the entities in the image file based on the crawler program; determining the ratio between the image size and the actual size of the entities; determining the actual distance between entities based on the image distance between entities and the ratio; and extracting the positions of entities in the image file based on the actual distance between entities to determine the logical relationship between entities.

[0010] Furthermore, in step 3, the step of obtaining entities and logical relationships between entities in each text file through semantic recognition based on the text file includes: identifying entities in the text file through a reasoning model based on the text file; and obtaining logical relationships between entities using semantic recognition based on the text file.

[0011] Furthermore, after step 3 and before step 4, the process includes: verifying and detecting conflicts between entities based on a heuristic rule base in the architecture domain; and deleting the logical relationships between entities that fail the verification and conflict detection.

[0012] Furthermore, after step 4, the method further includes: acquiring newly added knowledge nodes and feedback data; obtaining the node connection relationships and weights of the knowledge graph based on a time-aware graph neural network; updating the current knowledge graph based on the obtained node connection relationships and weights; the time-aware graph neural network is used to output the node connection relationships and weights of the knowledge graph based on the newly added knowledge nodes and feedback data.

[0013] Compared with existing technologies, this invention has the following advantages: This invention provides a method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence. By combining semantic recognition and image processing technologies, it effectively solves problems such as knowledge silos, low retrieval efficiency, and lagging knowledge updates caused by highly unstructured and specialized knowledge data in the construction engineering industry. For different types of documents such as process manuals, construction plans, and acceptance reports, the system first classifies them into image files and text files, and further ensures the standardization and compatibility of subsequent processing through format unification preprocessing. For image files, black and white line processing and two-dimensional coordinate system mapping are used. Through line segment normalization and matching with the entity database, engineering entities in drawings and schematic diagrams are accurately identified. Based on the ratio between image size and actual size, the actual distance and positional relationship between entities are calculated, thereby extracting spatial and logical connections that conform to engineering reality. For text files, reasoning models and semantic recognition technologies are used to extract the entities and their relationships hidden in the text. Verification and conflict detection are performed using a heuristic rule base in the construction field to ensure that the extracted logical relationships conform to industry standards and professional common sense. Ultimately, the system integrates entities and relationships extracted from images and text to construct a well-structured and accurately represented knowledge graph. This graph not only achieves effective integration and unified representation of architectural engineering knowledge, but also supports dynamic updates and weight optimization through a time-aware graph neural network. It can continuously adapt to the evolution and feedback of industry knowledge, thereby significantly improving the accuracy of knowledge retrieval and the reliability of decision support. Overall, it enhances the intelligent level of knowledge management in the architectural engineering field, perfectly aligning with the professional characteristics and practical application needs of the architectural engineering industry.

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0015] Figure 1 is a schematic diagram of the overall process of a knowledge graph construction method for the construction engineering industry based on artificial intelligence provided by the present invention. Detailed Implementation

[0016] As shown in Figure 1, the present invention provides a method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, comprising: Step 1: acquiring unstructured files from the construction engineering industry and classifying them into image files or text files according to their file type; Step 2: based on the image files, obtaining the entities and logical relationships between entities in each image file through image processing; Step 3: based on the text files, obtaining the entities and logical relationships between entities in each text file through semantic recognition; Step 4: constructing a knowledge graph based on the logical relationships between entities in the image files and the logical relationships between entities in the text files.

[0017] This invention aims to address the core problems identified in the background section regarding the highly unstructured nature of knowledge in the construction engineering industry, semantic biases in general processing techniques, and difficulty in aligning with industry-specific logic. To this end, this invention proposes a dedicated knowledge graph construction method that integrates multimodal information processing and domain knowledge verification. In general, this invention first performs intelligent classification and preprocessing on massive amounts of unstructured files (such as CAD drawings, construction photos, process texts, and acceptance reports). Then, through customized image processing and enhanced semantic recognition processes, it accurately extracts entities and their logical relationships within the construction engineering domain from image and text files. Subsequently, it introduces heuristic rules from the construction domain to clean and verify the extracted relationships. Finally, it integrates all information to construct a knowledge graph and utilizes a time-aware mechanism to achieve dynamic evolution and optimization of the graph, thereby forming an intelligent knowledge base capable of deeply understanding the professional context of construction engineering and accurately reflecting the industry's knowledge system.

[0018] In step 1, unstructured files from the construction industry are obtained and classified into image files or text files according to their file type.

[0019] Step 1 involves data acquisition and preliminary classification. The system collects various unstructured documents from the enterprise, including corporate policies, process standard manuals, construction plans, standards and specifications, contract texts, project meeting minutes, approval processes, and project acceptance reports. These data are widely distributed across the enterprise's ERP, OA, and other business systems, and come in formats such as text, tables, images, and PDFs, exhibiting highly unstructured and specialized characteristics. The system first uses file extensions and header information to initially classify files into image files (primarily visual information, such as design drawings, site photos, and diagrams) and text files (primarily character sequences, such as construction plan texts, material specifications, and safety briefing records). This classification forms the basis for subsequent differentiated processing, as it requires distinctly different feature extraction and understanding techniques for image and text data. This separation of processing paths creates conditions for applying optimal image processing and natural language processing techniques separately.

[0020] Furthermore, in step 1, after obtaining unstructured files from the construction industry and before classifying the unstructured files into image files or text files according to file type, the process includes: obtaining the content of the unstructured files; and decomposing the unstructured files into multiple image files and text files.

[0021] In practice, many unstructured documents are composite documents; for example, a PDF file may contain both text pages and drawing pages. Therefore, a file decomposition sub-step is added before the initial classification in step 1. The system parses the original file content and uses layout analysis technology to output visible graphic blocks as independent image files and text blocks as independent text files. The principle behind this approach is to refine the data granularity, ensuring that each image or text data unit is a simple modality, facilitating subsequent processing. Its effect is to improve the usability of the original data and avoid information loss caused by composite file formats.

[0022] Between step 1 and step 2, the method further includes: converting the image files to the same format so that all image files have the same file extension; and converting the text files to the same format so that all text files have the same file extension.

[0023] Before the data enters the core processing flow (steps 2 and 3), a format standardization preprocessing is performed. All image files (e.g., .jpg, .png, .bmp) are converted to a unified format well-supported by the processing library (e.g., .png). Similarly, the content of all text files is extracted and cleaned, and saved as a unified plain text format (e.g., .txt) or structured text format (e.g., .json). The principle is to eliminate processing interference caused by differences in file encoding, compression formats, and accompanying metadata, providing consistent and clean input for subsequent algorithms. The effect is to enhance the robustness and processing efficiency of the system and simplify the interface design between the image processing module and the text processing module.

[0024] In step 2, based on the image files, the entities in each image file and the logical relationships between the entities are obtained through image processing.

[0025] Step 2 involves entity and relationship extraction from image files. For the classified image files, this step executes a dedicated image processing workflow. Its core principle is that entities in architectural engineering images (especially technical drawings) are typically composed of specific lines, symbols, and annotations, and the spatial positions and connections between entities contain important engineering logic (such as "beam supporting a slab" or "pipes passing through a wall"). This invention uses a series of image analysis techniques to transform pixel-level visual information into symbol-level entity and relationship representations. In specific implementation, the refined advanced methods provided by this invention can be used. The effect of this step is to overcome the limitations of traditional text mining in processing image knowledge, digitizing the rich spatial and structural relationships contained in the drawings, which is key to constructing a knowledge graph that fits the actual engineering project.

[0026] The step of obtaining entities and logical relationships between entities in each image file through image processing includes: converting the image file into black and white lines so that each pixel in the image file is only black and white, with black pixels forming lines in the image file; determining entities in the image file based on the shape of the lines in the image file; and determining logical relationships between entities based on the distance between entities in the image file.

[0027] Furthermore, determining the entities in the image file based on the shape of the lines in the image file includes: establishing a two-dimensional coordinate system; representing the image file in the two-dimensional coordinate system after the black and white lines are converted into lines; cutting each line at its inflection point to obtain multiple line segments; representing the coordinates of each pixel of each line segment using a set, and normalizing the set of each line segment; obtaining the corresponding entities based on the set of each line segment using an entity database; the entity database is used to store entities and one or more sets corresponding to the entities.

[0028] Meanwhile, determining the logical relationship between entities based on the distance between entities in the image file includes: obtaining the image size of the entities in the image file and the image distance between entities; obtaining the actual size of the entities in the image file based on the crawler program; determining the ratio between the image size and the actual size of the entities; determining the actual distance between entities based on the image distance between entities and the ratio; and extracting the positions of entities in the image file based on the actual distance between entities to determine the logical relationship between entities.

[0029] First, the image is processed into black and white lines. Algorithms such as grayscale conversion, binarization, and line thinning are applied to the input image to obtain an image containing only black lines and a white background. The principle is that the core information of engineering drawings is carried by lines; removing color and grayscale interference greatly simplifies subsequent shape analysis. The effect is to highlight the structural skeleton of the image.

[0030] Next, shape-based entity recognition is performed. The processed image is placed in a two-dimensional coordinate system, with black pixels forming lines. Lines are cut at their inflection points to obtain a set of basic line segments. For each line segment, all its pixel coordinates are normalized to a standard scale (e.g., the [0,1] interval), forming a normalized coordinate set representing its shape characteristics. The system has a pre-built entity database storing various architectural engineering entities (such as "rectangular column," "I-beam," and "gate valve symbol") and their corresponding one or more standard line segment feature sets. By calculating the similarity (e.g., cosine similarity, Euclidean distance) between the extracted line segment feature set and the standard features in the database, the entity represented by the line segment or combination of line segments can be identified. The principle is to use geometric shapes as stable and invariant features of entities. The effect is to achieve high-precision, interpretable recognition of drawing symbols and component outlines, overcoming the low recognition rate of general image recognition models on professional symbols.

[0031] Then, spatial relationship derivation based on distance is performed. After identifying entities, their relative positions and image distances are measured. To obtain practical logical relationships with engineering significance (such as "5 meters away", "adjacent", "above"), image distances need to be converted into actual distances. The system obtains the standard actual dimensions of the identified entities from authoritative material libraries, standard specification libraries, or by parsing annotations in the images through a web crawler. A scale is obtained by calculating "entity image size / entity actual size". This scale is then used to convert the image distances between entities into actual distances. Finally, based on the actual distances and relative orientations, the logical relationships between entities are determined and output, such as "adjacent to", "distance 5m", "support by", etc. The principle is to introduce prior domain knowledge (actual entity dimensions) to calibrate the image measurement results, realizing a mapping from pixel space to physical space. The effect is that the extracted spatial relationships have real physical meaning and engineering value, enabling the knowledge graph to truly reflect the constraints of the physical world.

[0032] In step 3, based on the text files, the entities in each text file and the logical relationships between the entities are obtained through semantic recognition.

[0033] Step 3 involves entity and relation extraction from the text files. For the categorized text files, this step performs an enhanced semantic recognition process. The principle is that construction engineering texts contain numerous technical terms (entities) and complex relationships between them, such as processes, sequences, and causal relationships. General NLP models lack domain knowledge and are prone to ambiguity. This invention uses a pre-trained industry language model or a model combined with a domain dictionary as the reasoning model to more accurately identify technical entities in the text (such as "C30 concrete" and "prestressed tensioning machine"). Subsequently, a relation extraction model, combined with syntactic analysis, is used to identify semantic relationships between entities (such as "applies to," "precedes," and "requires"). The effect of this step is to achieve deep semantic understanding of unstructured text, accurately capturing abstract knowledge such as process flows and specification requirements described in the document.

[0034] Specifically, the step of obtaining entities and logical relationships between entities in each text file through semantic recognition based on the text file includes: identifying entities in the text file through an inference model based on the text file; and obtaining logical relationships between entities using semantic recognition based on the text file.

[0035] For text file processing, a two-stage model is employed. In the first stage, pre-trained models such as BERT and ERNIE, fine-tuned on a construction engineering corpus, are used as inference models to perform named entity recognition, accurately locating and classifying specialized entities within the text. In the second stage, a relation extraction model based on a pre-trained model, or a pattern matching method combining dependency parsing and rules, is used to semantically identify logical relationships such as "material_used_for_process", "risk_leads_to_accident", and "regulation_prohibited_behavior" from sentences or paragraphs where entities co-occur. The principle is to leverage industry-knowledge-enhanced deep learning models to understand the deep semantics of specialized texts. The effect is a significant reduction in semantic biases observed in general models within specialized domains, and an improvement in the accuracy of relation extraction.

[0036] After step 3 and before step 4, the process includes: verifying and detecting conflicts between entities based on a heuristic rule base in the architecture domain; and deleting the logical relationships between entities that fail the verification and conflict detection.

[0037] After step 3 and before the fusion construction in step 4, a quality control step is introduced. The system maintains a heuristic rule base for the construction field, containing industry common sense and standards (e.g., "concrete pouring" must follow "formwork installation"; the strength grade of "reinforcing steel" must match the design). The system uses this rule base to validate all logical relationships extracted from text and images. For example, if the relationship "concrete pouring precedes formwork installation" is detected, which conflicts with the process order in the rule base, the relationship will be marked as conflicting. All relationships that fail validation or are detected as conflicting will be deleted or marked as pending review and will not be included in the final knowledge graph. The principle is to use domain expert knowledge as an external constraint to logically cleanse the machine learning results. The effect is to greatly improve the professional correctness and reliability of the constructed knowledge graph, ensuring that the knowledge output of the graph conforms to engineering logic, which is one of the core guarantees of "perfectly fitting the construction engineering industry".

[0038] In step 4, a knowledge graph is constructed based on the logical relationships between entities in the image file and the logical relationships between entities in the text file.

[0039] Step 4 is knowledge graph construction. The entity-relationship pairs extracted from images in Step 2 are merged with those extracted from text in Step 3. The system first aligns the same entity from different sources (for example, the "main beam" in text and a specific beam component identifier in a drawing should be the same entity). Then, using entities as nodes and logical relationships as edges, a graph-structured data model, i.e., the knowledge graph, is constructed. The principle is that graph structures are a natural form for expressing complex relationships. Its effect is to completely break down the "knowledge silos" between different documents and modalities, forming a unified, interconnected, and computable knowledge network, providing the underlying data structure for intelligent retrieval, question answering, and decision support.

[0040] Following step 4, the process further includes: acquiring newly added knowledge nodes and feedback data; obtaining the node connection relationships and weights of the knowledge graph based on a time-aware graph neural network; updating the current knowledge graph based on the obtained node connection relationships and weights; the time-aware graph neural network is used to output the node connection relationships and weights of the knowledge graph based on the newly added knowledge nodes and feedback data.

[0041] Knowledge graph construction is not a one-time process. After building the initial graph in step 4, the system enters a dynamic maintenance phase. When new knowledge nodes (entities or relationships) are added, or when user feedback on existing knowledge is received (such as "This relationship is not commonly used" or "This answer is helpful"), the system initiates an update process. A time-aware graph neural network is crucial in this stage, taking the graph structure, node attributes, and timestamps of relationship occurrences or feedback as input. Its principle is to capture the patterns of knowledge and user feedback evolution over time. Through training, this network can output predictions of updated connections between nodes and adjustments to relationship weights (e.g., relationships frequently validated positively increase in weight, while relationships not used for a long time or receiving negative feedback decrease in weight). The system incrementally updates the knowledge graph based on this output. The effect is that the knowledge graph possesses the ability to self-learn and evolve, adapting to the release of new technologies and standards, and optimizing the priority of knowledge presentation based on actual usage feedback, maintaining the timeliness and practicality of the knowledge graph.

[0042] In summary, through a series of collaborative technical means such as multimodal processing, domain knowledge fusion, logic verification, and dynamic evolution detailed in the above specific embodiments, this invention constructs a professional knowledge graph that is truly tailored to the characteristics of the construction engineering industry, accurate, vivid, and intelligently applicable, effectively solving the various challenges raised in the background technology.

[0043] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, characterized in that, The process includes: Step 1: Obtaining unstructured documents from the construction industry and classifying them into image files or text files based on their file type; Step 2: Using image processing, obtaining the entities and logical relationships between them in each image file; Step 3: Using semantic recognition, obtaining the entities and logical relationships between them in each text file. Step 4: Construct a knowledge graph based on the logical relationships between entities in image files and between entities in text files.

2. The method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence as described in claim 1, characterized in that, In step 1, after obtaining unstructured files from the construction industry and before classifying the unstructured files into image files or text files according to file type, the steps include: obtaining the content of the unstructured files; and decomposing the unstructured files into multiple image files and text files.

3. The method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence as described in claim 1, characterized in that, Between step 1 and step 2, the method further includes: converting the image files to the same format so that all image files have the same file extension; and converting the text files to the same format so that all text files have the same file extension.

4. A method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, as described in claim 1, is characterized in that... In step 2, the step of obtaining entities and logical relationships between entities in each image file through image processing includes: processing the image file into black and white lines so that each pixel in the image file is only black and white, where the black pixels form the lines in the image file; determining the entities in the image file based on the shape of the lines in the image file; and determining the logical relationships between entities based on the distance between entities in the image file.

5. A method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, as described in claim 4, is characterized in that... The step of determining entities in an image file based on the shape of lines in the image file includes: establishing a two-dimensional coordinate system; representing the image file in the two-dimensional coordinate system after processing the black and white lines, wherein the black pixels in the image file are lines in the two-dimensional coordinate system; cutting each line at its inflection point to obtain multiple line segments; representing the coordinates of each pixel of each line segment using a set, and normalizing the set of each line segment; and obtaining the corresponding entity based on the set of each line segment using an entity database; wherein the entity database is used to store entities and one or more sets corresponding to the entities.

6. A method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, as described in claim 5, is characterized in that... The step of determining the logical relationship between entities based on the distance between entities in the image file includes: obtaining the image size of the entities in the image file and the image distance between the entities; obtaining the actual size of the entities in the image file based on the crawler program; determining the ratio between the image size and the actual size of the entities; determining the actual distance between the entities based on the image distance between the entities and the ratio; and extracting the positions between the entities in the image file based on the actual distance between the entities to determine the logical relationship between the entities.

7. A method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, as described in claim 1, is characterized in that... In step 3, the step of obtaining entities and logical relationships between entities in each text file through semantic recognition based on the text file includes: identifying entities in the text file through an inference model based on the text file; and obtaining logical relationships between entities using semantic recognition based on the text file.

8. A method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, as described in claim 1, is characterized in that... After step 3 and before step 4, the process includes: verifying and detecting conflicts between entities based on a heuristic rule base in the architecture domain; and deleting the logical relationships between entities that fail the verification and conflict detection.

9. A method for constructing a knowledge graph for the construction engineering industry based on artificial intelligence, as described in claim 1, characterized in that, Following step 4, the process further includes: acquiring newly added knowledge nodes and feedback data; obtaining the node connection relationships and weights of the knowledge graph based on a time-aware graph neural network; updating the current knowledge graph based on the obtained node connection relationships and weights; the time-aware graph neural network is used to output the node connection relationships and weights of the knowledge graph based on the newly added knowledge nodes and feedback data.

Images