Engineering quantity extraction and checking method and system based on vector feature analysis of engineering drawings

Abstract

The application discloses an engineering quantity extraction and reconciliation method and system based on engineering drawing vector feature analysis, relates to the field of digital engineering cost, and comprises the following steps: obtaining engineering drawing files and contract list data, extracting original vector elements and text annotations, and constructing a weighted vector relationship graph; combining the semantic feature decomposition of the geometric properties of the graph nodes and the neighborhood text annotations, and realizing scale adaptive fusion according to the local node distribution density and topological correlation; constructing a hierarchical vocabulary atlas based on the contract list data, performing hierarchical graph matching between the vector relationship graph and the vocabulary atlas, and determining the standard items corresponding to the component entities; and finally, completing the engineering quantity summary, reconciliation analysis and reliability evaluation according to the measurement rules, and outputting a reconciliation report. Therefore, the relevance, accuracy and review support capability of the engineering quantity extraction and business reconciliation processing are improved.

Description

Technical Field

[0001] This invention belongs to the field of digital engineering cost technology, and in particular relates to a method and system for extracting and reconciling engineering quantities based on vector feature analysis of engineering drawings. Background Technology

[0002] With the continuous expansion of the scale of construction and infrastructure projects, the evolution of engineering cost management towards digitalization and intelligence has become an important trend in the industry. Although digital technologies such as Building Information Modeling (BIM) have been gradually adopted in some scenarios, two-dimensional engineering CAD drawings and corresponding electronic files still occupy an important position in design expression, approval process, and construction implementation. These two-dimensional drawings not only carry intuitive information such as component outlines and dimensions, but also contain various expressive contents related to component attributes, material requirements, and project implementation. Therefore, quantity processing based on two-dimensional engineering drawings remains one of the fundamental links in engineering cost management.

[0003] Currently, common practices for quantity surveying fall into two categories: one involves using digital measurement tools, where cost estimators select graphic elements, delineate areas, and perform manual calculations within the drawing interface; the other involves converting 2D drawings into image format and automatically recognizing graphic objects and annotations to output the corresponding quantity results. While these methods improve calculation efficiency to some extent, they still rely heavily on local graphic recognition, text extraction, or manual interaction, and are insufficient in utilizing the geometric relationships, topological connections, and semantic correspondence between text and graphics in the original vector elements of the drawings.

[0004] Furthermore, in engineering cost estimation, quantity processing typically involves more than just generating the results; it also involves the connection and verification between the quantity data and contract lists, budget data, and related management information. Current processing methods often focus more on generating the quantity results, while offering relatively limited consideration for subsequent data correlation, discrepancy verification, and review support. In scenarios with dense drawing annotations, complex component relationships, or large project data volumes, there is still room for improvement in the consistency, relevance, and review efficiency of quantity processing results. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method and system for extracting and reconciling engineering quantities based on vector feature analysis of engineering drawings.

[0006] The technical solution of the present invention is as follows:

[0007] A method for quantity extraction and reconciliation of engineering work based on vector feature analysis of engineering drawings, the method comprising:

[0008] Obtain the engineering drawing file to be processed and the corresponding contract list data, and extract the original vector element set and the text annotation set corresponding to the original vector element set from the engineering drawing file;

[0009] Each native vector element in the set of native vector elements is used as a graph node, and a weighted vector relationship graph is constructed based on the spatial geometric relationship and connection status between each native vector element.

[0010] For each graph node in the weighted vector relation graph, the geometric attribute features corresponding to the graph node are extracted, and combined with the corresponding text annotations within the neighborhood of the graph node, semantic feature decomposition processing is performed on the geometric attribute features and the corresponding text annotations to obtain the basic semantic feature vector corresponding to the graph node; according to the local node distribution density of the graph node and the topological association in the weighted vector relation graph, scale adaptive fusion processing is performed on the basic semantic feature vector to generate a multi-scale fusion feature vector mapped to a unified semantic space;

[0011] A hierarchical vocabulary graph is constructed based on the contract list data. Based on the multi-scale fusion feature vector, the topological associations in the weighted vector relationship graph, and the vocabulary hierarchy relationships in the hierarchical vocabulary graph, hierarchical graph matching is performed between the weighted vector relationship graph and the hierarchical vocabulary graph to determine the standard entries corresponding to the component entities composed of the graph nodes.

[0012] For each of the aforementioned standard items, the geometric attribute features corresponding to each drawing node in the component entity mapped to the standard item are summarized and calculated according to the measurement rules corresponding to the standard item to obtain the target extracted engineering quantity. The target extracted engineering quantity is then reconciled with the contract list data. A reliability assessment is performed based on the matching confidence level of each of the aforementioned standard items and the corresponding reconciliation deviation. Based on the assessment results, a reconciliation report containing discrepancy information and risk warnings is output.

[0013] Furthermore, the construction of a weighted vector relationship graph based on the spatial geometric relationships and connection states between each original vector element includes:

[0014] For any two graph nodes in the original vector element set, extract the common edge length, contact length or overlap length, directional angle, and spatial centroid distance between the two graph nodes;

[0015] Using a preset geometric association evaluation function, the connection edge weight between the two graph nodes is calculated by linearly weighting and summing the normalized ratio of the common edge length, contact length or overlap length relative to the global maximum representation length, the directional consistency gain term obtained by converting the directional angle, and the distance attenuation factor based on the spatial centroid distance.

[0016] When the weight of the connecting edge is greater than a preset connectivity threshold, a connecting edge with corresponding weight is established between the two graph nodes to construct an initial weighted vector relationship graph.

[0017] For any graph node in the initial weighted vector graph, when it is determined that there are adjacent graph nodes, the weighted topological centrality algorithm is used to calculate the centrality score of the graph node in the global topology by multiplying and aggregating the relative proportion of the edge weights between the graph node and each of its adjacent graph nodes in the sum of local adjacency weights with a decay factor based on the shortest path length of the inverse edge weight; when it is determined that there are no adjacent graph nodes, its centrality score is recorded as zero.

[0018] Based on the centrality scores of each graph node and the spatial proximity relationships between graph nodes, graph nodes with centrality scores lower than a preset centrality threshold are identified as fragmented graph nodes. Geometric aggregation is then performed on the fragmented graph nodes that meet the preset merging conditions. The corresponding graph nodes and connecting edges are updated based on the aggregation results to generate a weighted vector relationship graph.

[0019] Further, the semantic feature decomposition process performed on the geometric attribute features and corresponding text annotations to obtain the basic semantic feature vector corresponding to the graph node includes:

[0020] Using a pre-trained language model for the engineering domain, word segmentation and entity recognition are performed on the corresponding text annotations located in the neighborhood of the graph node, extracting the set of text words associated with the graph node, and mapping each word in the set of text words to a corresponding text embedding vector;

[0021] The length, cross-sectional scale, orientation angle, and layer identifier features of the graph nodes are extracted as the geometric attribute features, and the corresponding geometric attribute vectors are constructed.

[0022] By using a linear mapping transformation with learning parameters, the geometric attribute vector is projected onto the same feature dimension space as the text embedding vector to obtain the geometric mapping vector;

[0023] When the text word set is not empty, the text embedding vector corresponding to each word in the text word set is extracted and the mean feature vector is calculated. Based on the preset semantic fusion coefficient, the geometric mapping vector and the mean feature vector are weighted and fused to generate the basic semantic feature vector corresponding to the graph node. When the text word set is empty, the geometric mapping vector is directly determined as the basic semantic feature vector corresponding to the graph node.

[0024] Further, the step of performing scale-adaptive fusion processing on the basic semantic feature vector based on the local node distribution density of the graph nodes and the topological associations in the weighted vector relationship graph to generate a multi-scale fused feature vector mapped to a unified semantic space includes:

[0025] In the weighted vector relationship diagram, with the target graph node as the center, the number of graph nodes within a preset search radius is counted, and the local node distribution density of the target graph node is calculated based on the ratio of the statistical result to the area of ​​the circular search region corresponding to the preset search radius.

[0026] The length feature or cross-sectional scale feature in the geometric attribute features of the target graph node is extracted as its feature size, and an adaptive fusion weight is calculated based on the local node distribution density and the feature size; wherein, the adaptive fusion weight is jointly determined by the sensitivity control function based on the local node distribution density and the relative size ratio of the feature size in the global reference dimension;

[0027] Based on the topological associations in the weighted vector relation graph, determine the set of adjacent graph nodes of the target graph node;

[0028] When the set of adjacent graph nodes is not empty, the adaptive fusion weight is used to perform weighted feature aggregation on the basic semantic feature vector of the target graph node and the mean of the basic semantic feature vectors of each graph node in the set of adjacent graph nodes, to generate the multi-scale fusion feature vector corresponding to the target graph node; when the set of adjacent graph nodes is empty, the basic semantic feature vector of the target graph node is determined as the corresponding multi-scale fusion feature vector.

[0029] Further, the construction of a hierarchical lexical graph based on the contract list data, and the hierarchical relationship of the lexical graph, involves performing hierarchical graph matching between the weighted vector relationship graph and the hierarchical lexical graph based on the multi-scale fusion feature vector, the topological associations in the weighted vector relationship graph, and the lexical hierarchy relationships in the hierarchical lexical graph, to determine the standard entries corresponding to the component entities composed of the graph nodes, including:

[0030] Based on the contract list data and the preset engineering quantity calculation specifications, the category attributes and measurement features of each standard item contained in the contract list data are extracted, the hierarchical vocabulary map with a tree-like hierarchical structure is constructed, and each standard semantic vector is generated and mapped to the unified semantic space.

[0031] Calculate the cosine similarity between the multi-scale fusion feature vector of each graph node and the standard semantic vector of each graph node to generate a matching score matrix composed of each graph node and each standard entry.

[0032] Using the set of graph nodes in the weighted vector relation graph as the first vertex set and the set of standard entries in the hierarchical vocabulary graph as the second vertex set, a bipartite graph matching model is constructed based on the matching score matrix.

[0033] A greedy search algorithm with unique mapping constraints for graph nodes is executed. Matching candidate edges are selected sequentially from the bipartite graph matching model in descending order of matching scores. After determining the mapping relationship between graph nodes and standard entries, other candidate edges connected to the graph node are removed from the search space to determine the candidate standard entries corresponding to each graph node.

[0034] The hierarchical consistency check is performed on the candidate standard entries corresponding to each graph node using the hierarchical lexical hierarchy relationship in the hierarchical lexical graph. After the check passes, the topological associations in the weighted vector relationship graph are combined to aggregate multiple graph nodes that are spatially adjacent and correspond to the same candidate standard entry into a component entity, thereby determining the standard entry corresponding to the component entity composed of the graph nodes.

[0035] Further, for each of the aforementioned standard items, according to the measurement rules corresponding to the standard item, the geometric attribute features corresponding to each drawing node in the component entity mapped to the standard item are summarized and calculated to obtain the target extracted engineering quantity, including:

[0036] When the measurement rule corresponding to the standard item is identified as length measurement, the length features of each graph node in the component entity are extracted and summed. Then, the sum of the common edge length, contact length or overlap length corresponding to the undirected connecting edges between connected graph nodes inside the component entity is subtracted. The length extraction quantity is obtained by topological deduplication calculation.

[0037] When the measurement rule corresponding to the standard entry is identified as area measurement, a closed topological loop is searched in the graph node set corresponding to the component entity based on the connectivity of the weighted vector relationship graph, and the searched topological loop is divided into outward contour boundaries and inward hole boundaries according to the directional attributes and inclusion relationships of the topological loop.

[0038] Calculate the base area enclosed by the outward contour boundary, and subtract the sum of the invalid areas enclosed by each of the inward hole boundaries to obtain the area extraction quantity.

[0039] The engineering quantity calculation results that match the measurement rules corresponding to the standard entries are determined as the target extracted engineering quantities; the engineering quantity calculation results include length extracted engineering quantities or area extracted engineering quantities.

[0040] Furthermore, the process of performing a reliability assessment based on the matching confidence level of each standard item and the corresponding reconciliation deviation, and outputting a reconciliation report containing discrepancy information and risk warnings based on the assessment results, includes:

[0041] Each of the aforementioned standard items is sequentially used as a target standard item. The set of graph nodes that are finally mapped to the target standard item after the hierarchical graph matching is determined. The matching score between each graph node in the set of graph nodes and the target standard item is obtained. The matching confidence level corresponding to the target standard item is calculated based on the aggregation result of each matching score. The predetermined quantity of work, the budget benchmark price and the contract execution price corresponding to the target standard item are parsed from the contract list data.

[0042] The absolute difference between the target extraction quantity and the predetermined quantity of work corresponding to the target standard item is calculated as the quantity deviation, and the absolute difference between the budget benchmark price and the contract execution price is calculated as the price deviation, so as to constitute the reconciliation deviation of the target standard item.

[0043] The comprehensive credibility score corresponding to the target standard item is calculated by a preset reliability evaluation function. The comprehensive credibility score is based on the matching confidence level and is obtained by exponential decay penalty according to the degree of deviation of the quantity deviation from the preset quantity tolerance correction parameter and the degree of deviation of the price deviation from the preset price tolerance correction parameter.

[0044] The project's overall funding deviation is obtained by weighting and summing the deviations in the quantities of each target standard item against the corresponding budget benchmark price. The total budget benchmark value in the funding dimension is then obtained by globally summing the products of the predetermined quantities of each target standard item against the corresponding budget benchmark price. The ratio of the project's overall funding deviation to the total budget benchmark value is used as the overall reconciliation deviation rate of the entire project.

[0045] The target standard items whose comprehensive credibility score is lower than the preset security threshold are marked as high-risk items, and each comprehensive credibility score, each high-risk item, and the global reconciliation deviation rate are determined as the evaluation results;

[0046] By combining the global reconciliation deviation rate and each of the high-risk items, a reconciliation report is generated that includes the discrepancy information and risk warnings; wherein the global reconciliation deviation rate, the quantity deviation, and the price deviation are used as the discrepancy information, and the high-risk items are used as the risk warnings.

[0047] A system for extracting and reconciling quantities of engineering work based on vector feature analysis of engineering drawings, the system comprising:

[0048] The drawing list extraction unit is used to obtain the engineering drawing file to be processed and the corresponding contract list data, and to extract the original vector element set and the text annotation set corresponding to the original vector element set from the engineering drawing file.

[0049] The vector relationship construction unit is used to take each native vector element in the native vector element set as a graph node and construct a weighted vector relationship graph based on the spatial geometric relationship and connection status between each native vector element.

[0050] The semantic feature fusion unit is used to extract the geometric attribute features corresponding to each graph node in the weighted vector relationship graph, and combine them with the corresponding text annotations located in the neighborhood of the graph node to perform semantic feature decomposition processing on the geometric attribute features and the corresponding text annotations to obtain the basic semantic feature vector corresponding to the graph node; according to the local node distribution density of the graph node and the topological association in the weighted vector relationship graph, the unit performs scale adaptive fusion processing on the basic semantic feature vector to generate a multi-scale fusion feature vector mapped to a unified semantic space;

[0051] The standard entry matching unit is used to construct a hierarchical vocabulary graph based on the contract list data, and perform hierarchical graph matching between the weighted vector relationship graph and the hierarchical vocabulary graph according to the multi-scale fusion feature vector, the topological association in the weighted vector relationship graph and the vocabulary hierarchy relationship in the hierarchical vocabulary graph, so as to determine the standard entry corresponding to the component entity composed of the graph nodes.

[0052] The quantity reconciliation assessment unit is used to summarize and calculate the geometric attribute features corresponding to each drawing node in the component entity mapped to the standard item according to the measurement rules corresponding to the standard item, to obtain the target extracted quantity of work, and to perform reconciliation analysis between the target extracted quantity of work and the contract list data; to perform a reliability assessment based on the matching confidence of each standard item and the corresponding reconciliation deviation, and to output a reconciliation report containing difference information and risk warnings based on the assessment results.

[0053] An electronic device, comprising:

[0054] At least one processor;

[0055] The system also includes a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the steps in the above-described method for extracting and reconciling quantities based on vector feature analysis of engineering drawings.

[0056] A storage medium storing a computer program that, when executed by a processor, implements the steps in the above-described method for extracting and reconciling quantities of engineering quantities based on vector feature analysis of engineering drawings.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] A unified relational and semantic representation mechanism is established around the native vector elements in engineering drawings. By organizing the spatial geometric relationships and connection states between native vector objects into a weighted relational structure, and performing joint semantic decomposition of the geometric attributes corresponding to graph nodes with neighborhood text annotations, and then combining local node distribution density and topological relationships to achieve multi-scale adaptive fusion, the originally scattered line and surface features and annotation information in the drawings can be transformed into feature representations with unified semantic orientation. This not only helps maintain the original accuracy of the geometric representation of the drawings, but also enhances the overall grasp of component boundaries, component combination relationships and their semantic reference relationships, thereby improving the stability and consistency of component entity recognition and semantic determination in complex drawing scenarios.

[0059] The multi-scale semantic expression formed on the drawing side is further mapped to the hierarchical business semantic system on the contract list side, and the mapping result is directly used in subsequent measurement and verification processing. Using a hierarchical lexical graph and its hierarchical constraints, component entities in the drawings can be assigned to standard entries within a unified business semantic framework. Based on this, relevant geometric attributes are summarized and calculated according to the measurement rules of the corresponding standard entries, and the resulting extracted quantities are reconciled with the contract list data. Simultaneously, a reliability assessment result is formed by combining matching confidence and reconciliation deviation. This ensures that the extracted quantities not only have numerical significance but also possess clear entry attribution, business correspondence, and verification support value, thereby improving the relevance, verifiability, and efficiency of discrepancy location.

[0060] This technical solution achieves cross-layer semantic integration from the native vector relationships in engineering drawings to standard cost items, and organically connects component identification, quantity calculation, and bill of quantities reconciliation during this process. Based on this continuous technical link from underlying primitive relationships, component semantic expression, standard item mapping to result verification output, the quantity processing results can simultaneously possess high semantic certainty, measurement consistency, and business usability, thus providing more stable and systematic technical support for quantity extraction and reconciliation analysis in engineering drawing scenarios. Attached Figure Description

[0061] Figure 1 The flowchart shows a method for extracting and reconciling quantities of engineering work based on vector feature analysis of engineering drawings.

[0062] Figure 2 Flowchart for constructing a weighted vector relationship graph;

[0063] Figure 3 This is a flowchart illustrating the process of determining the standard entries corresponding to component entities by performing hierarchical graph matching.

[0064] Figure 4 This is a schematic diagram illustrating the system operation mechanism of a method for extracting and reconciling quantities of engineering quantities based on vector feature analysis of engineering drawings.

[0065] Figure 5 A schematic diagram illustrating the comparative simulation results of the error evolution in quantity extraction using different methods under varying drawing density.

[0066] Figure 6 A schematic diagram illustrating the comparative simulation results of different methods processing time evolution under different drawing complexities;

[0067] Figure 7 A structural block diagram of a quantity extraction and reconciliation system based on vector feature analysis of engineering drawings, according to an embodiment of this application, is shown. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0069] It should be noted that in current related technologies, some solutions for the automatic recognition and parsing of engineering drawings have attempted to introduce methods such as object detection, semantic segmentation, and text recognition to extract component objects, dimension annotations, and explanatory information from the drawings. These solutions typically convert 2D drawings into raster images before performing image-level processing. Their technical approach focuses more on the recognition and classification of visual content, while making relatively limited use of the connection relationships, adjacency relationships, and geometric constraints carried by the original vector elements of the drawings. Since engineering quantity processing has high requirements for the accurate representation of underlying geometric relationships, if the original vector structural information is not fully utilized, the results often still have room for improvement in terms of geometric representation stability and object attribution accuracy in scenarios such as complex contours, dense annotations, or multiple adjacent objects.

[0070] On the other hand, in processing drawing text and annotation information, current technologies mostly employ text detection or character recognition to extract dimensions, elevations, and attribute information. However, text annotations in engineering drawings often exhibit diverse directions, dense distribution, and overlap with geometric lines, making simple text extraction methods susceptible to the complexity of the drawing. In particular, some processing solutions typically treat graphic recognition and text recognition separately, paying insufficient attention to the spatial correspondence, referential relationships, and combinatorial semantics between text content and adjacent geometric entities. Consequently, it is difficult to further organize scattered graphic and annotation information into component entity representations with complete business semantics.

[0071] Furthermore, while existing digital measurement or intelligent quantity calculation methods can assist in quantity acquisition to some extent, many processing steps still rely heavily on manual selection, marking, or subsequent verification. It is also difficult to maintain consistency in data representation and result organization across different operational paths. Simultaneously, there is often a lack of direct structured correlation between quantity processing results and business information such as contract lists and budget data. This results in extraction results remaining largely at the numerical output level, with shortcomings in subsequent item matching, discrepancy verification, and review support. Especially in application scenarios with diverse engineering drawing types, complex component relationships, or large project data scales, how to simultaneously address geometric relationship expression, semantic association between graphics and text, and business data integration during the drawing parsing stage remains a problem that requires further improvement in current technologies.

[0072] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the above description of the relevant technologies is only used to illustrate the technical background and the technical problem to be solved in this application, and does not constitute a limitation on the scope of protection of this application.

[0073] Figure 1A flowchart illustrating an example of a method for extracting and reconciling quantities based on vector feature analysis of engineering drawings, according to an embodiment of this application, is shown.

[0074] Regarding the execution subject of the method in the embodiments of this application, it can be any controller or processor with computing or processing capabilities, such as the controller of an intelligent engineering cost processing platform. It implements the method of the embodiments of this application by running programs or instructions stored in a storage medium. In some examples, it can be integrated and configured in an electronic device or terminal through software, hardware, or a combination of both, and the type of terminal or electronic device can be diverse.

[0075] like Figure 1 As shown, in step S110, the engineering drawing file to be processed and the corresponding contract list data are obtained, and the set of native vector elements and the set of text annotations corresponding to the set of native vector elements are extracted from the engineering drawing file.

[0076] In some implementations, the engineering drawing files to be processed can be engineering design drawings from different disciplines, such as architectural floor plans, structural reinforcement drawings, and electromechanical installation drawings, or electronic files exported from them that retain vector characteristics. The system can directly read the graphic data in the drawings through the drawing parsing module, extracting basic geometric objects with engineering expression significance, such as straight lines, polylines, arcs, curves, closed contours, and block references, to form a set of native vector elements. At the same time, the system can also synchronously parse text boxes, leader lines, dimensions, and explanatory text in the drawings, and extract their position, orientation, and spatial correspondence with surrounding graphic elements to form a set of text annotations corresponding to the set of native vector elements. For contract list data, the system can obtain project codes, project names, project feature descriptions, units of measurement, budget control information, and contract execution information by parsing list files, structured tables, or business databases.

[0077] In this technical scenario, the objects represented and the annotation formats in drawings from different disciplines often vary significantly. For example, in structural reinforcement drawings, the drawings may simultaneously contain steel bar segments, component cross-sectional outlines, and specification annotations such as "Φ16 steel bar"; in architectural floor plans, room boundaries, door and window outlines, axes, and explanatory text such as "wall thickness 200" may appear simultaneously; and in electromechanical installation drawings, densely arranged pipeline segments, elbow outlines, equipment symbols, and corresponding numbers and specification annotations may appear. By directly extracting from native vector and native text objects, rather than first converting the drawings into pure pixel images and then performing recognition, the geometric accuracy, primitive boundary relationships, and textual spatial relationships in the drawings can be preserved more completely, providing an input foundation for subsequent primitive organization and joint analysis of text and image semantics based on topology.

[0078] In step S120, each native vector element in the native vector element set is used as a graph node, and a weighted vector relationship graph is constructed based on the spatial geometric relationship and connection status between each native vector element.

[0079] In some implementations, the system abstracts the original vector elements extracted in step S110 into graph nodes and analyzes the spatial geometric relationships and connection states between graph nodes in a unified coordinate system. The spatial geometric relationships can include contact, adjacency, overlap, intersection, continuous extension, and proximity relationships between graph elements; the connection state reflects whether the graph elements collectively form the same local boundary, the same contour unit, or the same structural unit in the drawing representation. Based on these relationships, the system can establish connecting edges between graph nodes with effective associations and assign corresponding weights according to the degree of association, thereby constructing a weighted vector relationship graph.

[0080] In actual drawings, a physical component is usually not represented by a single graphic element, but rather by a combination of multiple discrete graphic elements according to certain rules. For example, in architectural floor plans, a wall is often represented by two parallel lines and several corner segments; in structural drawings, the outline of a beam or column may be defined by multiple boundary lines and local annotations; in MEP drawings, a continuous pipeline may be represented by several line segments, turning segments, and connecting symbols. By constructing a weighted vector relationship diagram, the originally discrete set of vector elements can be organized into a relational network with topological connectivity, so that subsequent processing is no longer limited to the isolated identification of individual graphic elements, but can identify which graphic elements are more likely to belong to the same component or the same unit of measurement in the drawing representation.

[0081] In step S130, for each graph node in the weighted vector relation graph, the geometric attribute features corresponding to the graph node are extracted, and combined with the corresponding text annotations within the neighborhood of the graph node, semantic feature decomposition processing is performed on the geometric attribute features and the corresponding text annotations to obtain the basic semantic feature vector corresponding to the graph node; according to the local node distribution density of the graph node and the topological association in the weighted vector relation graph, scale adaptive fusion processing is performed on the basic semantic feature vector to generate a multi-scale fusion feature vector mapped to a unified semantic space.

[0082] In some implementations, the system first extracts corresponding geometric attribute features for each graph node. These features may include element length, local scale, orientation, layer attributes, and other information reflecting the element's shape and expressive attributes. Simultaneously, centered on the target graph node, the system retrieves spatially related text annotations within its neighborhood and performs word segmentation, entity recognition, and semantic encoding on these annotations. For example, for a linear element in a structural reinforcement drawing, the neighborhood text may contain specification information such as "Φ16 rebar"; for a boundary element in an architectural plan, the neighborhood text may contain construction information such as "wall thickness 200". By jointly decomposing and representing the geometric attribute features and textual semantic information, the system generates a basic semantic feature vector corresponding to the graph node, ensuring that the representation of the graph node reflects not only its geometric appearance but also its semantic meaning within the drawing context.

[0083] Furthermore, the system performs scale-adaptive fusion processing on the basic semantic feature vectors based on the local node distribution density of the region where the graph node is located and the topological relationships in the weighted vector relationship graph. For graph nodes located in regions with simple structures and relatively sparse connections, more of their own features can be preserved; for graph nodes located in regions with dense connections and complex local configurations, the contextual information of adjacent graph nodes can be appropriately introduced for joint fusion. For example, in structural graphs, locally dense regions may correspond to the concentrated expression of microscopic objects such as steel bars and holes; in architectural graphs, they may correspond to the intersection of local boundaries of larger-scale regions such as room outlines and floor slab boundaries. By adopting fusion processing based on local density and topological relationships, the final multi-scale fused feature vector can simultaneously reflect the primitive's own attributes, neighborhood semantics, and local structural environment, thereby improving the semantic recognition stability in complex drawing scenarios.

[0084] In step S140, a hierarchical vocabulary graph is constructed based on the contract list data. Based on the multi-scale fusion feature vector, the topological associations in the weighted vector relationship graph, and the vocabulary hierarchy relationships in the hierarchical vocabulary graph, hierarchical graph matching is performed between the weighted vector relationship graph and the hierarchical vocabulary graph to determine the standard entries corresponding to the component entities composed of graph nodes.

[0085] In some implementations, the system performs structured parsing of contract list data, extracting item names, category relationships, measurement attributes, and descriptive business information, and constructs a hierarchical lexical graph based on this. The hierarchical lexical graph can form a tree-like or multi-level organizational structure according to professional categories, component categories, sub-specifications, and measurement attributes, used to express the hierarchical and semantic relationships between standard items on the business side. For example, the lexical graph can include standard items such as "beam," "column," and "wall," and can be further refined to business objects at different professional or specification levels. Simultaneously, the system models the graph structure on the drawing side, composed of graph nodes and their multi-scale fused feature vectors, in correspondence with the hierarchical lexical graph on the business side.

[0086] In the specific matching process, the system comprehensively considers multi-scale fused feature vectors, topological associations between graph nodes, and lexical hierarchy relationships in the hierarchical lexical graph, performing hierarchical graph matching between the weighted vector relationship graph and the hierarchical lexical graph. For multiple graph nodes that are spatially adjacent, semantically similar, and can be classified as the same entry at the lexical level, the system can aggregate them into component entities and further determine the standard entry corresponding to the component entity. For example, in an architectural floor plan, multiple adjacent graphic elements that constitute the same wall outline can be aggregated and mapped to the "wall" entry; in a structural drawing, graphic elements that constitute the same beam boundary or column boundary can be aggregated and mapped to the "beam" or "column" entry. Through this step, a seamless mapping can be achieved from discrete graphic elements on the drawing side to component entities, and then to standard entries on the business side, thus providing a clear object attribution basis for subsequent quantity extraction and reconciliation analysis based on standard entries.

[0087] In step S150, for each standard item, the geometric attribute features corresponding to each drawing node in the component entity mapped to the standard item are summarized and calculated according to the measurement rules corresponding to the standard item to obtain the target extracted engineering quantity. The target extracted engineering quantity is then reconciled with the contract list data. A reliability assessment is performed based on the matching confidence of each standard item and the corresponding reconciliation deviation. Based on the assessment results, a reconciliation report containing discrepancy information and risk warnings is output.

[0088] In some implementations, for component entities whose item mapping has been completed, the system summarizes and calculates the geometric attribute features corresponding to each drawing node within the component entity according to the measurement rules of the corresponding standard item to obtain the target extracted quantity. During this process, the system can combine the boundary relationships, connection relationships, and repetitive expression relationships between drawing nodes within the component entity to perform merging, deduplication, or boundary deduction processing on relevant geometric quantities, ensuring that the obtained quantity results are consistent with the measurement caliber of the corresponding item. For different types of component entities, the measurement caliber used can correspond to the attributes of their standard item; for example, some items focus more on linear extension, while others focus more on closed area quantities, thus enabling the automatic extraction results to effectively align with the business-side list caliber.

[0089] After extracting the target quantities, the system reconciles them with the corresponding information in the contract list data. It then performs a reliability assessment based on the matching confidence level output from the hierarchical diagram matching stage and the reconciliation discrepancies obtained in the current step. Furthermore, the system can mark items with high uncertainty, significant quantity-price discrepancies, missing annotations, or dimensional anomalies as risk items, and generate a reconciliation report containing discrepancy information and risk warnings based on the overall analysis results. The reconciliation report can be output in the form of a tabular file, structured data table, or database records, and can include supplementary information such as drawing location, text matching source, and item attribution for subsequent manual review and auditing. Thus, it not only completes the extraction of quantities but also establishes an automated verification relationship between the drawing analysis results and the contract list, forming a closed-loop processing flow from drawing data acquisition, structured analysis, item attribution, to reconciliation output.

[0090] Figure 2 A flowchart illustrating an example of constructing a weighted vector relationship graph according to an embodiment of this application is shown.

[0091] like Figure 2 As shown, in step S210, for any two graph nodes in the original vector element set, the common side length, contact length or overlap length, directional angle and spatial centroid distance between the two graph nodes are extracted.

[0092] In some implementations, the native vector elements in engineering drawings often exhibit a complex spatial distribution. Different elements may have direct contact, partial overlap, or local shared edges, or they may simply be spatial neighbors. To accurately characterize whether elements have a basis for further edge construction, the system preferably extracts association information from multiple geometric dimensions. Specifically, the common edge length, contact length, or overlap length reflects the degree of fit between two elements at their local boundaries; the directional angle characterizes the consistency of two elements in their extension direction; and the spatial centroid distance reflects the proximity of two elements in their overall position. By jointly extracting the above geometric relationships, a basic input can be provided for the subsequent calculation of connection edge weights.

[0093] In actual drawings, the scenarios corresponding to the above-mentioned geometric relationships have strong engineering significance. For example, in architectural floor plans, two locally parallel line segments with a contact boundary may jointly constitute a wall boundary; in structural drawings, several adjacent edge lines may jointly define the outline of a beam, column, or slab; in MEP drawings, continuous linear elements usually have strong directional consistency and spatial continuity. Therefore, by simultaneously extracting the degree of fit, directional relationship, and spatial distance, the system can no longer limit its judgment of the relationship between elements to a single geometric condition, but rather form a more complete spatial relationship description.

[0094] In step S220, the connection edge weight between two graph nodes is calculated by using a preset geometric association evaluation function, which is achieved by linearly weighting and summing the normalized ratio of the common edge length, contact length or overlap length relative to the global maximum representation length, the directional consistency gain term obtained by converting the directional angle, and the distance attenuation factor based on the spatial centroid distance.

[0095] For example, the weight of the connection edge between two graph nodes It can be obtained by calculation using the following formula:

[0096] ;

[0097] In the formula, Represents graph nodes Graph Nodes The weights of the connecting edges between them; This indicates the common edge length, contact length, or overlap length between two graph nodes; Indicates the global maximum connection representation length; Indicates the angle between directions; Pi; Indicates the distance between the centroids of space; This is a preset reference distance for the spatial centroid; , , The preset weighting coefficients are used, and they satisfy the following conditions: .

[0098] The above formula embodies the joint evaluation logic of multi-dimensional geometric information. Specifically, the normalized ratios of common edge length, contact length, or overlap length are used to enhance primitive pairs with high fit; the directional consistency gain term is used to ensure that primitives with more consistent orientations make a higher contribution; and the distance attenuation factor is used to suppress primitive pairs with large spatial spans. By linearly weighting and summing these three parts, the edge weights can comprehensively reflect the actual geometric correlation between two primitives, thus providing a quantitative basis for determining which connections to retain and which to weaken in the graph structure.

[0099] In step S230, when the weight of the connecting edge is greater than the preset connectivity threshold, a connecting edge with corresponding weight is established between two graph nodes to construct an initial weighted vector relationship graph.

[0100] In some implementations, the system does not perform fully connected graph construction on all graph nodes. Instead, it filters nodes based on a comparison between the edge weights and a preset connectivity threshold. Only when the edge weight between two graph nodes is greater than the preset connectivity threshold is a weighted edge established between them; for graph node pairs whose weights do not reach the threshold, no edge is established. In this way, weak relationships in engineering drawings caused by accidental proximity, local overlap, or representational noise can be eliminated, allowing the initial weighted vector graph to retain more structurally meaningful connections.

[0101] Thus, the discrete set of primitives in the drawing is organized into an initial weighted vector relationship diagram. This diagram not only preserves whether connections exist between primitives, but also further preserves the strength differences between different connections. After this processing, local boundaries, continuous contours, and combined components in the drawing can be presented in the diagram structure with a clearer connection pattern.

[0102] In step S240, for any graph node in the initial weighted vector relation graph, when it is determined that there are adjacent graph nodes, the weighted topological centrality algorithm is used to multiply and aggregate the relative proportion parameter of the edge weight between the graph node and each of its adjacent graph nodes in the sum of local adjacency weights with the attenuation factor composed of the shortest path length based on the reciprocal of the edge weight to obtain the centrality score of the graph node in the global topology; when it is determined that there are no adjacent graph nodes, its centrality score is recorded as zero.

[0103] For example, the weighted topological centrality algorithm can be expressed as follows:

[0104] ;

[0105] In the formula, Represents graph nodes Centrality score; Represents graph nodes To graph nodes The shortest path length is calculated based on the reciprocal of the edge weights; Represents graph nodes For graph nodes The local connectivity contribution weights, where, ; Represents graph nodes The set of adjacent graph nodes. In determining graph nodes... If there are no adjacent graph nodes, score their centrality. Record it as 0.

[0106] As a supplement, the shortest path length By connecting graph nodes With graph nodes Each feasible path is obtained by performing a path cost search, where the path cost is determined by the sum of the inverses of the weights of each edge on the path.

[0107] The centrality score described above considers both the local connectivity strength distribution between the current graph node and its neighboring nodes, as well as the node's reachability within the overall graph structure. Graph nodes that are tightly connected to surrounding elements and occupy a central position in the global graph structure typically receive higher centrality scores; conversely, nodes that are scattered, have weak local connections, or are detached from the main structure usually receive lower centrality scores. This process distinguishes between core boundary elements and secondary fragment elements in the drawing, thus providing a basis for subsequent element correction.

[0108] In step S250, based on the centrality scores of each graph node and the spatial proximity relationships between graph nodes, graph nodes with centrality scores lower than a preset centrality threshold are identified as fragmented graph nodes. Geometric aggregation is performed on fragmented graph nodes that meet preset merging conditions. The corresponding graph nodes and connecting edges are updated based on the aggregation results to generate a weighted vector relationship graph.

[0109] In some implementations, the system identifies graph nodes with a centrality score below a preset centrality threshold based on the centrality scores of each graph node and the spatial proximity between graph nodes, and designates them as candidate fragment graph nodes. For candidate fragment graph nodes that meet preset merging conditions, the system can perform geometric aggregation processing on them by combining their geometric proximity, connection strength, and local boundary continuity with neighboring backbone graph nodes, and update the graph node attributes and connection edge distribution based on the aggregation results; for graph elements that do not meet the aggregation conditions, their independent state can be retained or further determined in subsequent processing.

[0110] In engineering drawings, due to factors such as drawing habits, fragmented blocks, local decorative lines, or boundary breaks, short, isolated line segments, locally repetitive line segments, or scattered elements that have lost their significance in the main structure often appear. By correcting elements based on centrality scoring and spatial proximity, fragmented elements in the initial weighted vector relationship diagram can be merged or weakened, thus making the final weighted vector relationship diagram more stably reflect the main boundary structures and component connections in the drawing. The resulting diagram structure is clearer in expressing the continuity of component outlines, local connectivity, and overall topological organization.

[0111] This application's embodiments complete the process from extracting geometric relationships between native vector elements, to weighted connectivity graph construction, and then to primitive correction based on topological centrality. This enables discrete primitives in engineering drawings to be organized into a weighted vector relationship graph with relatively clear topological relationships. This processing result helps to accurately preserve the main structural boundary relationships in the drawings and reduces the interference of local fragmented primitives on the graph's structural representation.

[0112] Regarding the implementation details of parsing the basic semantic feature vectors corresponding to the graph nodes in step S130, in some examples of embodiments of this application, firstly, a pre-trained engineering domain language model is used to analyze the basic semantic feature vectors located at the graph nodes. Word segmentation and entity recognition are performed on the corresponding text annotations within the neighborhood range, and graph nodes are extracted. Related text word set and the text word set Each word in the text is mapped to a corresponding text embedding vector. .

[0113] In practical implementation scenarios, text annotations in engineering drawings typically carry engineering semantic information directly related to surrounding graphics. For example, in structural drawings, text annotations might include material and reinforcement specifications such as "C30 concrete" and "Φ16@200"; in architectural drawings, text annotations might include wall thickness, construction methods, or room names; and in mechanical and electrical drawings, text annotations might also include equipment numbers, pipe diameter specifications, or descriptions of strong and weak current types. To establish spatial associations between text information and corresponding graphic elements, the system preferably uses the geometric centroid of the graphic node as a reference, combined with the center position of the bounding box of the text words, and retrieves adjacent text according to preset spatial distance constraints. As a supplementary explanation, the text word set... The logic for obtaining [the value] can be expressed by the following spatial distance constraint formula:

[0114] ;

[0115] In the formula, Represents graph nodes Geometric centroid and textual words The two-dimensional planar Euclidean distance between the centers of the bounding boxes; This indicates the preset threshold for the text association search radius.

[0116] Through the aforementioned neighborhood retrieval, text annotations related to the graph node space can be extracted. Considering the distinct domain attributes of technical terms in engineering drawings, the system employs a pre-trained engineering domain language model for word segmentation, entity recognition, and semantic encoding to improve the accuracy of identifying material categories, specifications, component attributes, and technical descriptions. After this processing, the discrete text is transformed into a unified-dimensional text embedding vector. This provides a numerical input for the joint representation of primitive geometric features and textual semantic features.

[0117] Then, extract the graph nodes. The length, cross-sectional scale, orientation angle, and layer identifier features are used as geometric attribute features, and corresponding geometric attribute vectors are constructed. .

[0118] In some implementations, the system can extract geometric attribute information closely related to the component representation from native vector primitives. Among these, length features reflect the linear extension scale of the primitive, cross-sectional scale features reflect the local width, thickness, or cross-sectional dimensions of the primitive, orientation angle features reflect the spatial orientation of the primitive in the two-dimensional plane, and layer identifier features inherit professional classification information pre-assigned during the drafting process. By numerically encoding and normalizing these geometric attributes, the system concatenates them to construct the corresponding geometric attribute vector. For example, it can be represented as:

[0119] ;

[0120] In the formula, , , , Numerical feature components corresponding to length, cross-sectional dimensions, orientation angle, and layer identifier, respectively; This represents the vector transpose operation.

[0121] Through this process, the originally scattered geometric appearance features are uniformly organized into a structured vector representation, enabling the morphological, spatial, and cartographic attributes of graph nodes to participate in the calculation under the same data form. This is beneficial for subsequently mapping geometric information and textual semantic information to the same feature space.

[0122] Then, through a linear mapping transformation with learning parameters, the geometric attribute vector is transformed... Projecting onto the same feature dimension space as the text embedding vector yields a geometric mapping vector. .

[0123] For example, geometric mapping vector It is obtained by calculation using the following formula:

[0124] ;

[0125] In the formula, For a trainable mapping weight matrix, This is a bias term.

[0126] In the above equation, a linear mapping transformation adjusts the geometric attribute vector to a feature dimension space consistent with that of the text embedding vector. Since geometric attribute vectors typically consist of a small number of physical attributes, while text embedding vectors often possess higher-dimensional semantic expressive power, directly fusing the two types of features can easily lead to scale inconsistencies and imbalanced representations. By introducing a trainable mapping weight matrix and bias terms, the system can continuously adjust the projection of geometric features into the semantic space based on training data, ensuring that geometric objects with similar business meanings have a more consistent representational basis after mapping.

[0127] Furthermore, in the text word set When the text is not empty, extract the text embedding vector corresponding to each word in the text word set and calculate the mean feature vector. Based on the preset semantic fusion coefficient, perform weighted fusion calculation on the geometric mapping vector and the mean feature vector to generate the basic semantic feature vector corresponding to the graph node.

[0128] For example, based on preset semantic fusion coefficients, the geometric mapping vector is... With each text embedding vector in the text word set Perform weighted fusion calculations to generate graph nodes. Corresponding basic semantic feature vector :

[0129] ;

[0130] In the formula, Represents graph nodes The corresponding basic semantic feature vector; Let represent the semantic fusion coefficients, and satisfy . ; Represents graph nodes A set of related text words; This represents the total number of words in the text word set; Represents the first word in the text word set. One word; Represents a geometric mapping vector; This represents the text embedding vector.

[0131] In some implementations, the text embedding vectors corresponding to each word in the text word set can first undergo mean pooling to form the overall semantic representation of the neighboring text, and then be weighted and fused with the geometric mapping vector according to the semantic fusion coefficient. After this processing, the resulting basic semantic feature vector retains both the morphological information of the graph node itself and the engineering semantic information expressed by the surrounding text annotations. For graph nodes with rich surrounding text, the basic semantic feature vector can reflect a more complete semantic context; for graph nodes with less surrounding text, it can retain more of the graph node's own geometric semantic representation.

[0132] In the text word set When empty, the geometric mapping vector will be... Determined as a graph node Corresponding basic semantic feature vector .

[0133] This fallback process ensures that even when no related text exists within the neighborhood of a target graph node, the system can still generate a basic semantic feature vector based on the node's own geometric mapping, thus maintaining consistency in feature output. In this way, regardless of the sufficiency of annotation information in the drawing, the process can provide a unified form of basic semantic representation for each graph node.

[0134] This embodiment achieves joint parsing of graph node geometric attributes and neighborhood text semantics, enabling discrete primitives to be represented as basic semantic feature vectors that simultaneously contain morphological and semantic information. This helps to reduce semantic ambiguity caused by a single geometric appearance and improves the representation integrity and distinguishability of graph nodes in complex engineering drawing scenarios.

[0135] Regarding the implementation details of generating multi-scale fusion feature vectors in step S130, in some examples of embodiments of this application, firstly, in the weighted vector relationship graph, with the target graph node as the center, the number of graph nodes within the preset search radius is counted, and the local node distribution density of the target graph node is calculated based on the ratio of the statistical result to the area of ​​the circular search region corresponding to the preset search radius.

[0136] In engineering drawings, the distribution of vector primitives in different areas often shows significant differences. For example, in areas where pipelines intersect, structural details are shown, or where local structural elements are concentrated, the number of primitives is often large, and the connections are more complex; while in areas such as wall edges, beam edges, or straight extensions of pipelines, the distribution of primitives is relatively sparse. To enable the system to identify the complexity of the local environment in which the target vector node is located, this implementation method preferably uses the target vector node as the center, counts the number of surrounding vector nodes within a preset search radius, and calculates the local node distribution density by combining this with the area of ​​the corresponding search region. Through this processing, the density index output by the system is no longer just a simple node count, but a scaled representation that reflects the degree of primitive arrangement within a unit area.

[0137] For example, target graph nodes Local node distribution density It can be obtained by calculation using the following formula:

[0138] ;

[0139] In the formula, Represents the target graph node Other graph nodes in the weighted vector relationship graph The distance between their centroids; Pi; Indicates the preset search radius; Indicates nodes other than the target graph In addition, to the target graph node The number of graph nodes whose distance is less than the preset search radius. Using the density calculation method described above, the primitive arrangement around the target graph node can be measured relatively stably, providing a basis for the adaptive adjustment of subsequent fusion weights.

[0140] Then, the length feature or cross-sectional scale feature in the geometric attribute features of the target graph node is extracted as its feature size, and the adaptive fusion weight is calculated based on the local node distribution density and the feature size; wherein, the adaptive fusion weight is jointly determined by the sensitivity control function based on the local node distribution density and the relative size ratio of the feature size in the global reference dimension.

[0141] In some implementations, the local node distribution density primarily reflects the environmental complexity of the region where the graph node is located, while the feature size reflects the scale attribute of the graph node itself within the local structure. For graph nodes located in dense regions, if a fixed ratio is still used for neighborhood feature fusion, the important representation of the target graph node itself is easily diluted by a large number of adjacent features; for small primitives with strong representational dependencies, relying entirely on their own features may result in a lack of sufficient contextual semantic information. Therefore, this implementation preferably considers both the local node distribution density and the feature size of the target graph node simultaneously, using both to determine the adaptive fusion weight, so that different graph nodes can obtain a weight allocation that is adapted to their local environment and their own scale during the fusion process.

[0142] For example, adaptive fusion weights It can be obtained by calculation using the following formula:

[0143] ;

[0144] In the formula, Represents the target graph node Adaptive fusion weights; is the base of the natural logarithm; Hyperparameters for controlling sensitivity to density changes; The preset density baseline threshold; For target graph nodes Feature dimensions; This is the global reference feature size.

[0145] In the above formula, the first part reflects the change in local node distribution density relative to the baseline threshold, while the second part reflects the relative proportion of the target graph node's own feature size to the global reference feature size. Through the combined effect of these two parts, graph nodes located in complex regions and with large-scale attributes can retain a higher proportion of their own features during fusion, while smaller or more locally dependent primitives can absorb more neighborhood information, thus making the fusion result more compatible with the local structural attributes of the graph nodes.

[0146] Then, based on the topological associations in the weighted vector graph, the target graph nodes are determined. adjacency graph node set .

[0147] Here, the set of adjacent graph nodes is not determined through a simple two-dimensional spatial nearest neighbor search, but rather relies on the connection relationships in the aforementioned weighted vector relationship graph. In other words, only graph nodes that have a valid topological association with the target graph node in the graph structure are included in the set of adjacent graph nodes. The underlying principle is that engineering drawings often contain elements that are spatially close but physically do not belong to the same component representation unit, such as intersecting pipelines at different elevations or adjacent but unconnected local annotation boundaries. By determining the adjacency range based on topological association rather than simple spatial distance, the possibility of irrelevant elements entering the fusion range can be reduced, making the source of neighborhood information more consistent with the actual structural relationships.

[0148] Furthermore, in the adjacency graph node set When not empty, use adaptive fusion weights For target graph nodes Basic semantic feature vectors and adjacency graph node set The target graph node is generated by weighted feature aggregation based on the mean of the basic semantic feature vectors of each graph node. Corresponding multi-scale fusion feature vector .

[0149] For example, multi-scale fusion feature vectors It can be obtained by calculation using the following formula:

[0150] ;

[0151] In the formula, Represents the target graph node The corresponding multi-scale fusion feature vector; Represents the target graph node The basic semantic feature vector; Represents the set of nodes in an adjacency graph Middle Adjacency Graph Nodes The basic semantic feature vector; This represents the total number of nodes in the adjacency graph node set.

[0152] In some implementations, the system first performs mean aggregation on the basic semantic feature vectors of the adjacency graph node set to obtain an overall semantic representation of the neighborhood. Then, it weights and fuses this overall semantic representation of the neighborhood with the basic semantic feature vector of the target graph node itself according to adaptive fusion weights. Through this process, the multi-scale fused feature vector obtained by the target graph node retains its own semantic features while incorporating contextual information provided by adjacent graph nodes with which it has effective topological associations. In this way, the fusion result can better reflect the semantic position of the target graph node in the local structured network.

[0153] Furthermore, in the adjacency graph node set When empty, the target graph node will be... Basic semantic feature vector The corresponding multi-scale fusion feature vector is determined. Through this process, when the target graph node has no effective adjacent nodes in the current weighted vector relationship graph, the system can still directly use its basic semantic feature vector as the output result, thereby ensuring that the generation process of the multi-scale fusion feature vector remains complete and consistent in isolated primitive scenarios.

[0154] In this embodiment, based on the basic semantic representation of graph nodes, the local node distribution density, the node's own scale attributes, and topological adjacency relationships are combined to achieve adaptive fusion of graph node features. This enables the generated multi-scale fused feature vector to more fully reflect the target graph node's own attributes and its local structural environment, which is beneficial to improving the consistency and stability of the semantic representation of graph nodes in complex engineering drawings.

[0155] Figure 3 A flowchart illustrating an example of the operation of determining standard entries corresponding to component entities by performing hierarchical graph matching in a method according to an embodiment of this application is shown.

[0156] like Figure 3 As shown, in step S310, based on the contract list data and the preset engineering quantity calculation specifications, the category attributes and measurement features of each standard item contained in the contract list data are extracted, a hierarchical vocabulary map with a tree-like hierarchical structure is constructed, and each standard semantic vector is generated and mapped to a unified semantic space.

[0157] In some implementations, engineering cost contract lists typically have a clear hierarchical organizational structure, such as being divided into professional categories, sub-projects, itemized projects, and detailed entries. The system first performs structured parsing on the imported contract list data, extracting the name, category, measurement rules, material attributes, specifications, and other business characteristics of each standard item. Based on this, a hierarchical lexical graph with hierarchical relationships is constructed. The hierarchical lexical graph can express the semantic hierarchy, business affiliation, and differences in measurement attributes between different standard items, thereby transforming the originally discrete list text data into a business semantic system that can be structured for retrieval and constraint verification.

[0158] Building upon this, the system further utilizes engineering domain language models to semantically encode each standard item, generating standard semantic vectors mapped to a unified semantic space. Since the nodes on the drawing side have already obtained multi-scale fusion feature representations in a unified semantic space through the aforementioned steps, by vectorizing the standard entries in the same semantic space, the drawing side objects and business side entries can be semantically compared and matched in the same feature space, thereby ensuring that the semantic association analysis between the two has a consistent data representation basis.

[0159] In step S320, the cosine similarity between the multi-scale fusion feature vector of each graph node and each standard semantic vector is calculated to generate a matching score matrix composed of each graph node and each standard entry.

[0160] In some implementations, the system traverses the multi-scale fused feature vectors of each graph node on the drawing side and the standard semantic vectors of each standard entry on the business side, and evaluates the degree of matching between the graph node and the standard entry in a unified semantic space by calculating the cosine similarity between the two. For example, graph nodes... Compared with standard entries Match score between It can be obtained by calculation using the following formula:

[0161] ;

[0162] In the formula, Represents graph nodes Compared with standard entries Cosine similarity matching score between them; Represents graph nodes Multi-scale fused feature vectors; Indicates standard entries Standard semantic vectors; The L2 norm of a vector is used to represent the vector's L2 norm.

[0163] Through the above similarity calculation, the system can form a matching score matrix composed of each graph node and each standard item. Since cosine similarity mainly reflects the consistency of vector directions and weakens the influence of absolute numerical magnitude, this method can stably characterize the degree of fit between the geometric and textual semantics contained in the graph nodes and the business semantics of the standard items.

[0164] In step S330, the set of graph nodes in the weighted vector relation graph is used as the first vertex set, and the set of standard entries in the hierarchical vocabulary graph is used as the second vertex set. A bipartite graph matching model is constructed based on the matching score matrix.

[0165] In some implementations, to transform the matching problem between graph nodes and standard entries into a structured problem, the system uses the set of graph nodes in the weighted vector graph as the first vertex set and the set of standard entries in the hierarchical vocabulary graph as the second vertex set, constructing a bipartite graph matching model based on the matching score matrix. For graph node-standard entry pairs whose matching scores meet preset association conditions, candidate matching edges can be established between the corresponding vertices, with the matching score used as the weight of the candidate edge. In this way, the candidate correspondence between drawing-side objects and business-side entries is uniformly expressed as a weighted bipartite graph structure.

[0166] Through the above processing, the originally scattered graph primitive semantic matching task is organized into a graph theory model with a clear set of vertices and candidate edges. This model not only preserves the matching strength between each graph node and each standard item, but also provides a unified data structure expression for subsequent searching, filtering and verification under business constraints.

[0167] In step S340, a greedy search algorithm with unique mapping constraints for graph nodes is executed. Matching candidate edges are selected from the bipartite graph matching model in descending order of matching scores. After determining the mapping relationship between graph nodes and standard entries, other candidate edges connected to the graph node are removed from the search space to determine the candidate standard entries corresponding to each graph node.

[0168] In some implementations, the system executes a greedy search algorithm with a unique mapping constraint on the bipartite graph matching model. Specifically, the system traverses the candidate edges in descending order of their matching scores. Once the mapping relationship between a graph node and a certain standard entry is determined, other candidate edges connected to that graph node are removed from the search space, leaving only one candidate standard entry for that graph node. For example, the unique mapping constraint on the graph node can be expressed as:

[0169] ;

[0170] In the formula, Represents graph nodes Compared with standard entries The mapping decision variable takes a value of 1 to indicate that a mapping is established, and a value of 0 to indicate that a mapping is not established. This indicates the total number of standard entries participating in the matching; Represents any graph node; This represents the set of graph nodes (i.e., the first set of vertices) in the weighted vector relation graph.

[0171] In the above formula, the constraint ensures that the same drawing node corresponds to at most one candidate standard entry during the entry assignment process, thereby avoiding the drawing node being repeatedly assigned to multiple business entries. Since there are some auxiliary lines or noise in the engineering drawings that do not need to be extracted, this constraint allows the mapping sum of some drawing nodes to be 0; at the same time, this constraint does not restrict multiple drawing nodes from pointing to the same standard entry, so that the initial mapping relationship between drawing nodes and candidate standard entries can be determined in a relatively orderly manner.

[0172] In step S350, the hierarchical consistency check is performed on the candidate standard entries corresponding to each graph node using the hierarchical relationship of the vocabulary graph. After the check passes, the topological associations in the weighted vector relationship graph are combined to aggregate multiple spatially adjacent graph nodes that correspond to the same candidate standard entry into a component entity, thereby determining the standard entry corresponding to the component entity composed of graph nodes.

[0173] In some implementations, after the system obtains candidate standard entries for each graph node through greedy matching, it does not directly output the final attribution result. Instead, it further utilizes the hierarchical relationships in the hierarchical lexical graph to perform a hierarchical consistency check. If there is a significant contradiction between the candidate standard entry of a graph node and its professional category, the attribution of its parent entries, or the hierarchical relationship of its entries, the candidate result can be eliminated or remarked as an object to be further judged. This check process can reduce hierarchical attribution conflicts caused by relying solely on local semantic similarity.

[0174] For graph nodes that pass the hierarchical consistency check, the system further returns to the weighted vector relationship graph. Combining the topological relationships between graph nodes, it performs aggregation processing on multiple spatially adjacent, topologically connected graph nodes corresponding to the same candidate standard item, forming component entities with continuous geometric boundaries and unified business affiliation. The resulting component entities are no longer discrete line segments or local primitives, but aggregated objects that simultaneously possess drawing structural continuity and item affiliation. Through this processing, candidate mapping relationships at the graph node level are elevated to standard item affiliation results at the component entity level, making the correspondence between drawing-side objects and contract list items clearer.

[0175] Furthermore, to enable the hierarchical graph matching results to be used for subsequent reliability assessment, the system calculates the matching confidence level for each standard item based on the matching scores between the graph nodes that ultimately pass the hierarchical consistency check and the standard items. Specifically, for any standard item... After passing the aforementioned unique mapping constraints and hierarchical consistency checks, it will ultimately be mapped to this standard entry. All graph nodes constitute the target graph node set. And record the number of graph nodes in the target graph node set as Among them, when multiple spatially independent component entities all correspond to the same standard entry. At that time, the graph nodes contained in multiple component entities are collectively included in the target graph node set. .

[0176] For the target graph node set Each graph node in Get the nodes of the graph Compared with standard entries The matching score calculated in step S320 Considering that the theoretical range of cosine similarity is [-1, 1], we can first perform linear normalization on the matching score to obtain the normalized matching score. :

[0177] ;

[0178] in, The value range is [0, 1].

[0179] Furthermore, for entries mapped to the same standard entry The normalized matching scores corresponding to each graph node are averaged and aggregated to obtain the matching confidence score corresponding to the standard item. For example, it can be represented as:

[0180] ;

[0181] In the formula, Indicates standard entries The corresponding matching confidence level; This indicates the final mapping to the standard entry. The number of nodes in the graph; Represents graph nodes Compared with standard entries The normalized matching score between them.

[0182] When standard entries When there is only one graph node, its matching confidence is determined by the normalized matching score corresponding to that graph node; when the standard entry When dealing with multiple graph nodes, aggregating the normalized matching scores for each node can comprehensively reflect the overall matching degree between the graph-side objects that form the standard entry and the business semantics. For standard entries that do not form effective node mappings, their matching confidence can be set to zero and they can be marked as entries to be verified, or they can be excluded from the set of standard entries with completed mappings. This achieves the conversion from graph node-level matching scores to standard entry-level matching confidence, providing a unified input basis for subsequent reliability assessments based on matching confidence and reconciliation discrepancies.

[0183] This embodiment completes the process from the structured representation of contract list items and the semantic similarity calculation between graph nodes and standard items, to the constraint matching and component entity aggregation. This enables discrete graph nodes to be grouped into component entities with clear standard item affiliations based on semantic information, hierarchical relationships, and topological associations, thereby forming a relatively stable correspondence between the structural objects on the drawing side and the list items on the business side.

[0184] Regarding the implementation details of calculating the target extracted engineering quantity in step S150, in some examples of embodiments of this application, on the one hand, when the measurement rule corresponding to the standard item is identified as length measurement, the length features of each graph node in the component entity are extracted and summed, and then the sum of the common edge length, contact length or overlap length corresponding to the undirected connecting edges between the connected graph nodes inside the component entity is deducted, and the length extracted engineering quantity is obtained by topological deduplication calculation.

[0185] In actual engineering cost calculation scenarios, for linear components such as pipes, cables, skirting boards, and others, length is typically used as the measurement standard. These components are often represented by multiple discrete geometric elements in engineering drawings. Designers may use methods such as endpoint connection, partial overlap, or short line splicing to represent continuous components. If the lengths of each discrete geometric element are simply added together, it is easy to double-count the contact or overlapping sections between geometric elements, leading to an overestimation of the quantity. Therefore, in this implementation, the system introduces a deduplication process based on graph structure connection relationships within the component entity. This deducts the common edge length, contact length, or overlap length between graph nodes belonging to the same component entity and having undirected connections with each other, making the length result closer to the actual effective length under the corresponding measurement standard.

[0186] For example, length is used to extract the quantity of work. It can be obtained by calculation using the following formula:

[0187] ;

[0188] In the formula, Indicates the length of the extracted work volume; Represents the set of graph nodes that constitute the component entity; Represents graph nodes Length characteristics; This represents the total deduplication length obtained after merging duplicate segments among the common side length, contact length, or overlap length of each connected graph node within the component entity.

[0189] In the above formula, the first part is used to obtain the sum of the length characteristics of each node within the component entity, and the second part is used to subtract the repeated length measurements after segment merging. It should be noted that by merging repeated segments first and then uniformly subtracting, duplicate subtractions can be avoided in scenarios common in CAD drawings where multiple elements partially overlap or where three or more elements are repeatedly expressed. Through this processing, the length measurement results can maintain good measurement consistency in scenarios with segmented element representation and complex overlapping representations.

[0190] On the other hand, when the measurement rule corresponding to the standard item is identified as area measurement, a closed topological loop is searched in the set of graph nodes corresponding to the component entity based on the connectivity of the weighted vector relationship graph. According to the direction attribute and inclusion relationship of the topological loop, the searched topological loop is divided into outward contour boundaries and inward hole boundaries. Then, the base area enclosed by the outward contour boundaries is calculated, and the sum of the invalid areas enclosed by each inward hole boundary is subtracted to obtain the area extraction quantity.

[0191] For planar components such as floor slabs, room areas, and green areas, they are typically enclosed by a set of boundary elements connected end-to-end in drawings. The enclosed area may also contain elevator shafts, pipe shafts, structural openings, or other hollow areas that should not be included in the effective area. In some implementations, the system can extract all closed topological loops within component entities belonging to the same area entry based on connectivity search of a weighted vector graph. To distinguish which closed boundaries represent effective outer contours and which represent internal openings, the system further classifies them based on the directional attributes of the topological loops and the inclusion relationships between different topological loops. Closed boundaries not surrounded by other topological loops and with a large enclosing area can be identified as outward contour boundaries; other boundaries located inside outward contours and forming independent closed areas can be identified as inward opening boundaries. The above loop finding and inclusion relationship analysis process can be implemented using graph traversal algorithms, geometric inclusion determination algorithms, or other suitable boundary recognition methods.

[0192] For example, area extraction of engineering quantity It can be obtained by calculation using the following formula:

[0193] ;

[0194] In the formula, This represents the base area calculated from the outward contour boundary; Indicates by the first Invalid area calculated from the boundary of an inward-facing hole; This represents the total number of inward-facing hole boundaries.

[0195] In the above formula, the system can perform area calculations separately for the outward contour boundary and each inward hole boundary, for example, by using a polygon area calculation method to obtain the area of ​​the region enclosed by the corresponding boundaries. Subsequently, by subtracting the invalid area corresponding to all inward hole boundaries from the base area, a valid area result conforming to the area measurement caliber can be obtained. Through this processing, area measurement no longer relies solely on a single closed boundary, but can simultaneously consider the combined influence of the outer contour and the internal hollow region, thus making the area result of complex planar components more complete.

[0196] Furthermore, the engineering quantity calculation results that match the measurement rules corresponding to the standard items are determined as the target extracted engineering quantities; the engineering quantity calculation results include length extracted engineering quantities or area extracted engineering quantities.

[0197] In some implementations, the system selects the result matching its measurement caliber from the corresponding engineering quantity calculation results according to the measurement rules corresponding to the standard item, and determines it as the target extracted engineering quantity. For length measurement items, the target extracted engineering quantity can be determined from the length result after topological deduplication; for area measurement items, the target extracted engineering quantity can be determined from the effective area result after deducting holes. Through this processing, the geometric expression of the component entity on the drawing side can be converted into an engineering quantity result consistent with the measurement caliber of the business item.

[0198] This embodiment enables differentiated quantity calculations for different measurement rules. For linear components, the system can deduct duplicate measurements based on the connection relationships between nodes in the drawing; for planar components, the system can determine the effective area based on the relationship between closed boundaries and internal holes. Therefore, the resulting extracted quantities can better correspond to the measurement requirements of different standard items and improve the consistency between the geometric representation on the drawings and the quantity results.

[0199] Regarding the implementation details of outputting the reconciliation report in step S150, in some examples of embodiments of this application, firstly, each standard item is taken as the target standard item in turn, the matching confidence level corresponding to the target standard item is obtained, and the predetermined quantity of work, the budget benchmark price and the contract execution price corresponding to the target standard item are parsed from the contract list data.

[0200] In actual engineering cost reconciliation scenarios, the quantity results extracted from the drawings alone are insufficient for item-level verification. Therefore, the system prioritizes acquiring the matching confidence score of each completed item for each standard item, calculated using the aforementioned node-level matching scores. Furthermore, it parses the contract list data to obtain the corresponding predetermined quantity, budget benchmark price, and contract execution price for that standard item. The predetermined quantity reflects the target value pre-agreed in the contract or list, the budget benchmark price reflects the price reference from the quota or budget side, and the contract execution price reflects the actual unit price during project execution. By simultaneously acquiring this information, a unified data foundation can be provided for item-level discrepancy analysis and reliability assessment.

[0201] For example, each standard item is sequentially designated as the [number]th [item]. There are 1 target standard item, and the target extraction engineering quantity corresponding to the target standard item is recorded as . Following the aforementioned method for calculating matching confidence, the matching confidence of the target standard item is obtained based on the matching scores corresponding to each graph node ultimately mapped to the target standard item. .

[0202] Then, the absolute difference between the target extracted quantity and the predetermined quantity of work corresponding to the target standard item is calculated as the quantity deviation, and the absolute difference between the budget benchmark price and the contract execution price is calculated as the price deviation, so as to constitute the reconciliation deviation of the target standard item.

[0203] In some implementations, the system performs absolute value processing on the differences between the target extracted quantity of work and the predetermined quantity of work, as well as the differences between the budget benchmark price and the contract execution price, so that the deviation measurement can simultaneously cover both over-reporting and under-reporting. Through this processing, item-level reconciliation deviations are no longer limited to a single quantitative difference, but simultaneously reflect deviations in both the quantity and price dimensions. For example, the predetermined quantity of work corresponding to the target standard item is parsed from the contract list data. Budget benchmark price and contract execution price Calculate the target extraction quantity respectively With the planned amount of work Deviations in the amount of work and the benchmark price in the budget With contract execution price The price discrepancy between the two constitutes the reconciliation discrepancy of the target standard item.

[0204] Then, the comprehensive credibility score corresponding to the target standard item is calculated through the preset reliability assessment function. The comprehensive credibility score is based on the matching confidence level and is obtained by exponential decay penalty according to the degree of deviation of the quantity deviation from the preset quantity tolerance correction parameter and the degree of deviation of the price deviation from the preset price tolerance correction parameter.

[0205] For example, the overall reliability score corresponding to the target standard item is calculated using the following reliability assessment function. :

[0206] ;

[0207] In the formula, Indicates the first A comprehensive credibility score for each target standard item; This indicates the matching confidence level corresponding to the target standard item; Indicates the target extraction quantity; Indicates the planned amount of work; Indicates the budget benchmark price; Indicates the contract execution price; These are preset tolerance correction parameters for the quantity of work; This is the preset price tolerance correction parameter; For the natural constant An exponential function with base 0.

[0208] In the aforementioned evaluation function, the matching confidence level serves as a fundamental term in the calculation, while quantity deviation and price deviation are adjusted for the overall credibility score through exponential decay terms. When the matching results of items on the drawing side are relatively stable, and both quantity and price deviations are within a small range, the resulting overall credibility score is relatively high. Conversely, when the matching confidence level is low, or when the quantity and price deviations increase significantly relative to their respective tolerance correction parameters, the overall credibility score decreases accordingly. This processing allows for the unified inclusion of item-level identification uncertainty and business-side quantity and price differences into a single scoring result.

[0209] Subsequently, the quantity deviations of each target standard item and the corresponding budget benchmark price are weighted and summarized in terms of funding to obtain the overall funding deviation of the project. The products of the predetermined quantity of each target standard item and the corresponding budget benchmark price are then summed globally to obtain the total budget benchmark value in terms of funding. The ratio of the overall funding deviation of the project to the total budget benchmark value is used as the overall reconciliation deviation rate of the entire project.

[0210] In some implementations, since different standard items may correspond to different units of measurement, such as length, area, volume, or quantity, the quantity deviations of different items cannot be directly added together using the same caliber. Therefore, the system preferably introduces a budget benchmark price to convert the quantity deviations of each item into deviations in terms of funding, and then performs project-level aggregation under a unified funding caliber. For example, the overall reconciliation deviation rate for the entire project... It can be obtained by calculation using the following formula:

[0211] ;

[0212] In the formula, This indicates the total number of target standard items included in the reconciliation analysis.

[0213] Through the above processing, the system can uniformly map the deviations of different items at the quantity level to the financial dimension for comparison and summarization, thereby obtaining a global reconciliation deviation rate that reflects the overall deviation of the project. This indicator considers both the differences in the quantity of each item and the price weight of the corresponding item, making it more suitable for summarizing deviations at the overall project level.

[0214] Next, the overall credibility score will be determined. Items falling below a preset safety threshold are marked as high-risk items, and their overall credibility scores are adjusted accordingly. The evaluation results were determined by the individual high-risk items and the overall reconciliation deviation rate.

[0215] In some implementations, the system can preset a security threshold and mark target standard items with an overall credibility score below the threshold as high-risk items. In this way, the item-level scoring results can be further transformed into explicit risk screening results, enabling the system to not only output continuous score values ​​for each item, but also to further identify abnormal items that need to be prioritized for verification. Thus, the item-level overall credibility score, the screened high-risk items, and the project-level global reconciliation deviation rate together constitute the reconciliation evaluation result.

[0216] Then, by combining the global reconciliation deviation rate and each high-risk item, a reconciliation report containing discrepancy information and risk warnings is generated; the global reconciliation deviation rate, quantity deviation, and price deviation are used as discrepancy information, and high-risk items are used as risk warnings.

[0217] In some implementations, the final reconciliation report generated by the system can be output in the form of structured tables, database records, visual pages, or audit working papers. The reconciliation report can include both a global reconciliation deviation rate at the project level and detailed item-level deviations in quantities and prices, while also identifying high-risk items. This output allows for the unified organization and display of item-level discrepancy information and project-level overall deviation information, resulting in reconciliation results with good readability and ease of verification.

[0218] This embodiment enables the comprehensive organization of item-level differences, project-level deviations, and risk items, allowing the differences between the target extracted quantities and the contract list data to be output in a structured manner. The resulting reconciliation report can more completely reflect the reconciliation results at both the item and project levels.

[0219] Figure 4 This diagram illustrates the system operation mechanism of an example of a method for extracting and reconciling quantities based on vector feature analysis of engineering drawings, according to an embodiment of this application. The system operation mechanism can be divided into four logical functional blocks: an input area, a core processing area, a scheduling and control area, and an output area.

[0220] like Figure 4 As shown, the input area is used to acquire vector CAD drawings and contract and budget lists. The data stream from the vector CAD drawings is input into the core processing area and sequentially passes through a vector parsing and topology mapping module, a semantic feature decomposition module, and a scale-adaptive fusion module to complete the structured representation of underlying graphic elements, graphic-text semantic parsing, and multi-scale feature representation. The data stream from the contract and budget lists is simultaneously input into the core processing area and, together with the feature information generated from the drawings, undergoes hierarchical graph matching processing in the hierarchical vocabulary matching module to determine the attribution relationship between component entities and standard entries.

[0221] After the core processing is completed, the relevant processing results are transmitted to the scheduling and control area. The quality control and rule engine module in the scheduling and control area can perform summary calculations on the geometric attributes on the drawing side based on standard items and their corresponding measurement rules, and output the quantity extraction results. The reliability assessment and risk detection module can receive the item matching results and quantity calculation results, and perform variance analysis and reliability assessment in conjunction with the corresponding data in the contract and budget lists.

[0222] The output area is used to output the results of the quantity extraction, as well as reconciliation reports and risk warnings. The reconciliation reports and risk warnings reflect item-level discrepancies and corresponding risk information. Figure 4 The dashed links in the diagram indicate that there may be feedback relationships between the scheduling and control area and the core processing area, in order to support the correction, constraint or review of the core processing process.

[0223] To verify the effectiveness and performance limits of the engineering quantity extraction and reconciliation method based on vector feature analysis of engineering drawings proposed in this application in real engineering scenarios, the technical effects of this application are explained below in conjunction with specific test datasets and comparative experiments. The dataset constructed for the test includes 500 sets of real engineering CAD drawings, specifically covering 200 sets of architectural floor plans, 150 sets of structural reinforcement drawings, and 150 sets of electromechanical installation drawings. The number of native vector nodes in each drawing ranges from 10,000 to 1.5 million, which can better simulate complex and changing actual engineering conditions.

[0224] To conduct objective, multi-dimensional comparisons, three comparative benchmarks were established during the testing process. The first benchmark was a manual processing benchmark, in which senior cost engineers used mainstream quantity calculation software for semi-automatic selection and extraction. This benchmark served as a reference for extraction accuracy, but it involved significant manual time consumption. The second benchmark was a traditional rasterization algorithm benchmark, which employed a conventional image instance segmentation model. After exporting the CAD drawings as high-resolution images, pixel-level target detection and region calculation were performed. The third benchmark was the method provided in the embodiments of this application, which directly read the original vector files of the engineering drawings, extracted the original vector elements to construct a weighted vector relationship graph, and combined it with text annotations to perform semantic feature decomposition and multi-scale adaptive fusion processing.

[0225] Furthermore, the evaluation metrics used in this comparative experiment mainly include mean relative error (MRE), overall recognition accuracy (F1-Score), and overall processing time. Peak memory consumption and sensitivity to drawing density were also incorporated as evaluation dimensions. Through these multi-dimensional evaluation metrics, the aim is to comprehensively assess the extraction accuracy, computational efficiency, and system robustness of different processing methods under conditions of large-scale drawings and high drawing density.

[0226] Figure 5 This diagram illustrates a comparative simulation of the error evolution in quantity extraction using different methods under varying drawing density. The comparative experiment employs a dual-axis evaluation system, where the horizontal axis represents drawing density (i.e., the number of nodes per square meter), the left main vertical axis represents the average relative error (%), and the right secondary vertical axis represents the topological ambiguity rate. Figure 5 The paper compares the performance of traditional image rasterization-based extraction methods (corresponding to AI-Raster in the figure) and the method provided in the embodiments of this application (corresponding to Proposed-Vector in the figure) under different drawing density conditions.

[0227] Combination Figure 5The simulation data shows that in simple scenarios where the components on the drawing are sparsely distributed (e.g., fewer than 50 nodes per square meter), both the traditional rasterization method and the method in this embodiment maintain a low average relative error, within the range of 3% to 4%. However, as the density of the drawing increases (e.g., the introduction of high-density drawings such as electromechanical installation drawings with more than 200 nodes per square meter), a large number of leaders, annotations, and geometric contours visually overlap. Under this complex condition, the traditional rasterization method, relying on pixel-level segmentation, suffers significant impact on its boundary features, increasing the topological ambiguity rate to around 0.4, resulting in an average relative error exceeding 12%.

[0228] Conversely, the method in this application does not forcibly convert the drawing into a raster image. Instead, it directly extracts native vector elements to construct a weighted vector relationship graph and performs multi-scale adaptive fusion processing by combining local node distribution density and topological association. This processing mechanism based on the native primitive topology can reduce interference caused by visual occlusion and pixel overlap. Therefore, as... Figure 5 As shown, even in high-density regions, the topological ambiguity rate of the method in this embodiment remains stable at a level close to 0, and the average relative error remains stable at around 2%. This simulation result demonstrates that the vector topology and multi-scale feature fusion mechanism in this embodiment possesses strong anti-interference capability and robustness when dealing with complex high-density drawing scenarios.

[0229] Figure 6 This diagram illustrates a comparative simulation of the processing time evolution of different methods under varying drawing complexity. The simulation experiment uses the number of primitive nodes as an example. The logarithmic scale (Log Scale) is used as the horizontal axis to measure the complexity of the drawing, and the end-to-end single-drawing processing time (in seconds) is used as the vertical axis. The figure compares the single-drawing time trends of the traditional raster extraction method (corresponding to AI-Raster in the figure) and the method provided in the embodiments of this application (corresponding to Proposed-Vector in the figure).

[0230] like Figure 6 As shown, based on the simulation distribution scatter plot and the fitted trend line, it can be seen that for nodes with fewer than [number missing] nodes... For drawings of low to medium complexity (such as conventional architectural floor plans), the processing time of the method in this application embodiment is generally stable between 10 and 20 seconds, which is roughly equivalent to the processing efficiency of traditional raster extraction methods. However, when the drawing complexity is high and the number of nodes exceeds a certain threshold... When dealing with orders of magnitude larger (e.g., densely packed basement utility layouts), the processing time of the method in this embodiment exhibits a rapid, non-linear increase, with some highly complex samples taking more than 60 seconds to process. At this point, its processing speed is slightly lower than that of traditional raster extraction methods with relatively fixed forward inference time complexity.

[0231] The non-linear increase in processing time mentioned above is related to the processing mechanism of graph theory network algorithms. The method in this embodiment constructs a weighted vector relation graph and performs hierarchical graph matching with exclusive mapping constraints in the non-image domain to determine matching combinations between component entities and standard entries. As the size of the topology nodes increases, the space complexity and computational time overhead of the system when performing graph matching searches increase accordingly.

[0232] Under the aforementioned processing time overhead, the method of this application embodiment can achieve high accuracy in topological relationship resolution and component extraction. The comparative experiment shows that the method of this application embodiment can achieve a certain balance between computational efficiency and extraction accuracy.

[0233] Therefore, although the vector feature parsing method proposed in this application demonstrates good performance in terms of extraction accuracy and topological integrity, it also faces objective performance limitations and technical trade-offs in practical engineering applications. For example, constructing a weighted vector relationship graph requires global node association, resulting in high memory consumption when processing extremely large and complex drawings. Simultaneously, the algorithm is somewhat dependent on the drawing standardization of the underlying CAD drawings. When faced with a large number of fragmented, non-standard drawings (such as numerous independent micro-short lines) or non-standard local abbreviations, the search space for graph matching and the mapping cost of the initial lexical graph will increase accordingly. However, overall, this application improves the accuracy of vector topological relationship parsing and the stability of component extraction in complex engineering drawing scenarios by allowing for certain computational resource overhead, thereby meeting the requirements for geometric relationship accuracy and result reliability in quantity extraction and reconciliation processing.

[0234] In summary, the embodiments of this application achieve cross-level semantic association from two-dimensional discrete lines to standard business lists through the construction of weighted vector relationship graphs, multi-scale feature fusion of text and graphics, hierarchical vocabulary matching, and comprehensive reconciliation and evaluation models. In further extended implementation scenarios, the technical architecture of this application can be deepened in multiple dimensions: for example, incremental vector analysis algorithms can be introduced to dynamically adapt to frequent engineering design changes; or the extracted structured vector component features can be mapped to the BIM (Building Information Modeling) system to achieve three-dimensional comprehensive control of quantity, price, and schedule; in addition, the hierarchical vocabulary graph can be flexibly expanded to adapt to multinational engineering measurement standards, and machine learning prediction models can be introduced in combination with historical project data, thereby providing advance cost risk warning and decision support for the entire digital engineering cost business.

[0235] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0236] Figure 7 A structural block diagram of an example of a quantity extraction and reconciliation system based on vector feature analysis of engineering drawings, according to an embodiment of this application, is shown.

[0237] like Figure 7 As shown, the engineering quantity extraction and reconciliation system 700 based on engineering drawing vector feature analysis includes a drawing list extraction unit 710, a vector relationship construction unit 720, a semantic feature fusion unit 730, a standard item matching unit 740, and an engineering quantity reconciliation evaluation unit 750.

[0238] The drawing list extraction unit 710 is used to obtain the engineering drawing files to be processed and the corresponding contract list data, and to extract the original vector element set and the text annotation set corresponding to the original vector element set from the engineering drawing files.

[0239] The vector relationship construction unit 720 is used to construct a weighted vector relationship graph by taking each native vector element in the native vector element set as a graph node and constructing the graph based on the spatial geometric relationship and connection status between each native vector element.

[0240] The semantic feature fusion unit 730 is used to extract the geometric attribute features corresponding to each graph node in the weighted vector relationship graph, and combine them with the corresponding text annotations located in the neighborhood of the graph node to perform semantic feature decomposition processing on the geometric attribute features and the corresponding text annotations to obtain the basic semantic feature vector corresponding to the graph node; according to the local node distribution density of the graph node and the topological association in the weighted vector relationship graph, the basic semantic feature vector is subjected to scale adaptive fusion processing to generate a multi-scale fusion feature vector mapped to a unified semantic space.

[0241] The standard entry matching unit 740 is used to construct a hierarchical vocabulary graph based on contract list data. Based on the multi-scale fusion feature vector, the topological association in the weighted vector relationship graph, and the vocabulary hierarchy relationship in the hierarchical vocabulary graph, hierarchical graph matching is performed between the weighted vector relationship graph and the hierarchical vocabulary graph to determine the standard entries corresponding to the component entities composed of graph nodes.

[0242] The quantity reconciliation assessment unit 750 is used to summarize and calculate the geometric attribute features corresponding to each drawing node in the component entity mapped to the standard item according to the measurement rules corresponding to the standard item, to obtain the target extracted quantity of work, and to perform reconciliation analysis between the target extracted quantity of work and the contract list data; to perform reliability assessment based on the matching confidence of each standard item and the corresponding reconciliation deviation, and to output a reconciliation report containing difference information and risk warnings based on the assessment results.

[0243] In some embodiments, this application provides a non-volatile computer-readable storage medium storing one or more programs including execution instructions. These execution instructions can be read and executed by electronic devices (including but not limited to computers, servers, or network devices) to perform the steps of the above-described method for extracting and reconciling quantities of engineering drawings based on vector feature analysis.

[0244] In some embodiments, this application also provides a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium. The computer program includes program instructions that, when executed by a computer, cause the computer to perform the steps of the above-described method for extracting and reconciling quantities based on vector feature analysis of engineering drawings.

[0245] In some embodiments, this application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of a quantity extraction and reconciliation method based on engineering drawing vector feature parsing.

[0246] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0247] The electronic device in this application embodiment can be an engineering cost processing terminal, a workstation, a server, or other electronic device with data processing capabilities.

[0248] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0249] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0250] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for quantity extraction and reconciliation of engineering work based on vector feature analysis of engineering drawings, characterized in that, The method includes: Obtain the engineering drawing file to be processed and the corresponding contract list data, and extract the original vector element set and the text annotation set corresponding to the original vector element set from the engineering drawing file; Each native vector element in the set of native vector elements is used as a graph node, and a weighted vector relationship graph is constructed based on the spatial geometric relationship and connection status between each native vector element. For each graph node in the weighted vector relation graph, the geometric attribute features corresponding to the graph node are extracted, and combined with the corresponding text annotations within the neighborhood of the graph node, semantic feature decomposition processing is performed on the geometric attribute features and the corresponding text annotations to obtain the basic semantic feature vector corresponding to the graph node; according to the local node distribution density of the graph node and the topological association in the weighted vector relation graph, scale adaptive fusion processing is performed on the basic semantic feature vector to generate a multi-scale fusion feature vector mapped to a unified semantic space; A hierarchical vocabulary graph is constructed based on the contract list data. Based on the multi-scale fusion feature vector, the topological associations in the weighted vector relationship graph, and the vocabulary hierarchy relationships in the hierarchical vocabulary graph, hierarchical graph matching is performed between the weighted vector relationship graph and the hierarchical vocabulary graph to determine the standard entries corresponding to the component entities composed of the graph nodes. For each of the aforementioned standard items, the geometric attribute features corresponding to each drawing node in the component entity mapped to the standard item are summarized and calculated according to the measurement rules corresponding to the standard item to obtain the target extracted engineering quantity. The target extracted engineering quantity is then reconciled with the contract list data. A reliability assessment is performed based on the matching confidence level of each of the aforementioned standard items and the corresponding reconciliation deviation. Based on the assessment results, a reconciliation report containing discrepancy information and risk warnings is output.

2. The method for extracting and reconciling engineering quantities based on vector feature analysis of engineering drawings according to claim 1, characterized in that, The construction of a weighted vector relationship graph based on the spatial geometric relationships and connection states between each original vector element includes: For any two graph nodes in the original vector element set, extract the common edge length, contact length or overlap length, directional angle, and spatial centroid distance between the two graph nodes; Using a preset geometric association evaluation function, the connection edge weight between the two graph nodes is calculated by linearly weighting and summing the normalized ratio of the common edge length, contact length or overlap length relative to the global maximum representation length, the directional consistency gain term obtained by converting the directional angle, and the distance attenuation factor based on the spatial centroid distance. When the weight of the connecting edge is greater than a preset connectivity threshold, a connecting edge with corresponding weight is established between the two graph nodes to construct an initial weighted vector relationship graph. For any graph node in the initial weighted vector graph, when it is determined that there are adjacent graph nodes, the weighted topological centrality algorithm is used to calculate the centrality score of the graph node in the global topology by multiplying and aggregating the relative proportion of the edge weights between the graph node and each of its adjacent graph nodes in the sum of local adjacency weights with a decay factor based on the shortest path length of the inverse edge weight; when it is determined that there are no adjacent graph nodes, its centrality score is recorded as zero. Based on the centrality scores of each graph node and the spatial proximity relationships between graph nodes, graph nodes with centrality scores lower than a preset centrality threshold are identified as fragmented graph nodes. Geometric aggregation is then performed on the fragmented graph nodes that meet the preset merging conditions. The corresponding graph nodes and connecting edges are updated based on the aggregation results to generate a weighted vector relationship graph.

3. The method for extracting and reconciling engineering quantities based on vector feature analysis of engineering drawings according to claim 2, characterized in that, The semantic feature decomposition process is performed on the geometric attribute features and corresponding text annotations to obtain the basic semantic feature vectors corresponding to the graph nodes, including: Using a pre-trained language model for the engineering domain, word segmentation and entity recognition are performed on the corresponding text annotations located in the neighborhood of the graph node, extracting the set of text words associated with the graph node, and mapping each word in the set of text words to a corresponding text embedding vector; The length, cross-sectional scale, orientation angle, and layer identifier features of the graph nodes are extracted as the geometric attribute features, and the corresponding geometric attribute vectors are constructed. By using a linear mapping transformation with learning parameters, the geometric attribute vector is projected onto the same feature dimension space as the text embedding vector to obtain the geometric mapping vector; When the text word set is not empty, the text embedding vector corresponding to each word in the text word set is extracted and the mean feature vector is calculated. Based on the preset semantic fusion coefficient, the geometric mapping vector and the mean feature vector are weighted and fused to generate the basic semantic feature vector corresponding to the graph node. When the text word set is empty, the geometric mapping vector is directly determined as the basic semantic feature vector corresponding to the graph node.

4. The method for extracting and reconciling engineering quantities based on vector feature analysis of engineering drawings according to claim 3, characterized in that, The step of performing scale-adaptive fusion processing on the basic semantic feature vector based on the local node distribution density of the graph nodes and the topological associations in the weighted vector relationship graph to generate a multi-scale fused feature vector mapped to a unified semantic space includes: In the weighted vector relationship diagram, with the target graph node as the center, the number of graph nodes within a preset search radius is counted, and the local node distribution density of the target graph node is calculated based on the ratio of the statistical result to the area of ​​the circular search region corresponding to the preset search radius. The length feature or cross-sectional scale feature in the geometric attribute features of the target graph node is extracted as its feature size, and an adaptive fusion weight is calculated based on the local node distribution density and the feature size; wherein, the adaptive fusion weight is jointly determined by the sensitivity control function based on the local node distribution density and the relative size ratio of the feature size in the global reference dimension; Based on the topological associations in the weighted vector relation graph, determine the set of adjacent graph nodes of the target graph node; When the set of adjacent graph nodes is not empty, the adaptive fusion weight is used to perform weighted feature aggregation on the basic semantic feature vector of the target graph node and the mean of the basic semantic feature vectors of each graph node in the set of adjacent graph nodes, so as to generate the multi-scale fusion feature vector corresponding to the target graph node; when the set of adjacent graph nodes is empty, the basic semantic feature vector of the target graph node is determined as the corresponding multi-scale fusion feature vector.

5. The method for extracting and reconciling engineering quantities based on vector feature analysis of engineering drawings according to claim 4, characterized in that, The hierarchical lexical graph constructed based on the contract list data, according to the multi-scale fused feature vector, the topological associations in the weighted vector relationship graph, and the lexical hierarchy relationships in the hierarchical lexical graph, performs hierarchical graph matching between the weighted vector relationship graph and the hierarchical lexical graph to determine the standard entries corresponding to the component entities composed of the graph nodes, including: Based on the contract list data and the preset engineering quantity calculation specifications, the category attributes and measurement features of each standard item contained in the contract list data are extracted, the hierarchical vocabulary map with a tree-like hierarchical structure is constructed, and each standard semantic vector is generated and mapped to the unified semantic space. Calculate the cosine similarity between the multi-scale fusion feature vector of each graph node and the standard semantic vector of each graph node to generate a matching score matrix composed of each graph node and each standard entry. Using the set of graph nodes in the weighted vector relation graph as the first vertex set and the set of standard entries in the hierarchical vocabulary graph as the second vertex set, a bipartite graph matching model is constructed based on the matching score matrix. A greedy search algorithm with a unique mapping constraint for graph nodes is executed. Matching candidate edges are selected sequentially from the bipartite graph matching model in descending order of matching scores. After determining the mapping relationship between graph nodes and standard entries, other candidate edges connected to the graph node are removed from the search space to determine the candidate standard entries corresponding to each graph node. The hierarchical consistency check is performed on the candidate standard entries corresponding to each graph node using the hierarchical lexical hierarchy relationship in the hierarchical lexical graph. After the check passes, the topological associations in the weighted vector relationship graph are combined to aggregate multiple graph nodes that are spatially adjacent and correspond to the same candidate standard entry into a component entity, thereby determining the standard entry corresponding to the component entity composed of the graph nodes.

6. The method for extracting and reconciling engineering quantities based on vector feature analysis of engineering drawings according to claim 5, characterized in that, For each of the aforementioned standard items, the geometric attribute features corresponding to each drawing node in the component entity mapped to that standard item are summarized and calculated according to the measurement rules corresponding to that standard item to obtain the target extracted engineering quantity, including: When the measurement rule corresponding to the standard item is identified as length measurement, the length features of each graph node in the component entity are extracted and summed. Then, the sum of the common edge length, contact length or overlap length corresponding to the undirected connecting edges between connected graph nodes inside the component entity is subtracted. The length extraction quantity is obtained by topological deduplication calculation. When the measurement rule corresponding to the standard entry is identified as area measurement, a closed topological loop is searched in the graph node set corresponding to the component entity based on the connectivity of the weighted vector relationship graph, and the searched topological loop is divided into outward contour boundaries and inward hole boundaries according to the directional attributes and inclusion relationships of the topological loop. Calculate the base area enclosed by the outward contour boundary, and subtract the sum of the invalid areas enclosed by each of the inward hole boundaries to obtain the area extraction quantity. The engineering quantity calculation results that match the measurement rules corresponding to the standard entries are determined as the target extracted engineering quantities; the engineering quantity calculation results include length extracted engineering quantities or area extracted engineering quantities.

7. The method for extracting and reconciling quantities of engineering work based on vector feature analysis of engineering drawings according to claim 6, characterized in that, The process involves performing a reliability assessment based on the matching confidence level of each standard item and the corresponding reconciliation deviation, and outputting a reconciliation report containing discrepancy information and risk warnings based on the assessment results, including: Each of the aforementioned standard items is sequentially used as a target standard item. The set of graph nodes that are finally mapped to the target standard item after the hierarchical graph matching is determined. The matching score between each graph node in the set of graph nodes and the target standard item is obtained. The matching confidence level corresponding to the target standard item is calculated based on the aggregation result of each matching score. The predetermined quantity of work, the budget benchmark price and the contract execution price corresponding to the target standard item are parsed from the contract list data. The absolute difference between the target extraction quantity and the predetermined quantity of work corresponding to the target standard item is calculated as the quantity deviation, and the absolute difference between the budget benchmark price and the contract execution price is calculated as the price deviation, so as to constitute the reconciliation deviation of the target standard item. The comprehensive credibility score corresponding to the target standard item is calculated by a preset reliability evaluation function. The comprehensive credibility score is based on the matching confidence level and is obtained by exponential decay penalty according to the degree of deviation of the quantity deviation from the preset quantity tolerance correction parameter and the degree of deviation of the price deviation from the preset price tolerance correction parameter. The project's overall funding deviation is obtained by weighting and summing the deviations in the quantities of each target standard item against the corresponding budget benchmark price. The total budget benchmark value in the funding dimension is then obtained by globally summing the products of the predetermined quantities of each target standard item against the corresponding budget benchmark price. The ratio of the project's overall funding deviation to the total budget benchmark value is used as the overall reconciliation deviation rate of the entire project. The target standard items whose comprehensive credibility score is lower than the preset security threshold are marked as high-risk items, and each comprehensive credibility score, each high-risk item, and the global reconciliation deviation rate are determined as the evaluation results; By combining the global reconciliation deviation rate and each of the high-risk items, a reconciliation report is generated that includes the discrepancy information and risk warnings; wherein the global reconciliation deviation rate, the quantity deviation, and the price deviation are used as the discrepancy information, and the high-risk items are used as the risk warnings.

8. A system for extracting and reconciling quantities of engineering work based on vector feature analysis of engineering drawings, characterized in that: The system includes: The drawing list extraction unit is used to obtain the engineering drawing file to be processed and the corresponding contract list data, and to extract the original vector element set and the text annotation set corresponding to the original vector element set from the engineering drawing file. The vector relationship construction unit is used to take each native vector element in the native vector element set as a graph node and construct a weighted vector relationship graph based on the spatial geometric relationship and connection status between each native vector element. The semantic feature fusion unit is used to extract the geometric attribute features corresponding to each graph node in the weighted vector relationship graph, and combine them with the corresponding text annotations located in the neighborhood of the graph node to perform semantic feature decomposition processing on the geometric attribute features and the corresponding text annotations to obtain the basic semantic feature vector corresponding to the graph node; according to the local node distribution density of the graph node and the topological association in the weighted vector relationship graph, the unit performs scale adaptive fusion processing on the basic semantic feature vector to generate a multi-scale fusion feature vector mapped to a unified semantic space; The standard entry matching unit is used to construct a hierarchical vocabulary graph based on the contract list data, and perform hierarchical graph matching between the weighted vector relationship graph and the hierarchical vocabulary graph according to the multi-scale fusion feature vector, the topological association in the weighted vector relationship graph and the vocabulary hierarchy relationship in the hierarchical vocabulary graph, so as to determine the standard entry corresponding to the component entity composed of the graph nodes. The quantity reconciliation assessment unit is used to summarize and calculate the geometric attribute features corresponding to each drawing node in the component entity mapped to the standard item according to the measurement rules corresponding to the standard item, to obtain the target extracted quantity of work, and to perform reconciliation analysis between the target extracted quantity of work and the contract list data; to perform a reliability assessment based on the matching confidence of each standard item and the corresponding reconciliation deviation, and to output a reconciliation report containing difference information and risk warnings based on the assessment results.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the method as described in any one of claims 1-7.

10. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

Images