A method and system for intelligent review of compliance of construction engineering bidding based on a multi-modal large model

Abstract

This invention discloses an intelligent review method and system for compliance of construction project bidding based on a multimodal large model, belonging to the field of intelligent review. The method includes: acquiring a multimodal dataset of the construction project to be reviewed; transforming semantic feature representations in a unified semantic space, and obtaining a set of semantically related entities composed of multiple semantically related entities based on a correlation index indicating semantic proximity; performing consistency checks on first and second observation data bound to the same semantically related entity based on the set of semantically related entities to obtain a first review result; locating abnormal bidding nodes that deviate from the normal bidding range to obtain a second review result; and inputting the first review result and the abnormal bidding nodes as context into the multimodal large model to generate an intelligent review report. This invention avoids the problems of repeated bidding for the same component and local high similarity crowding out the global optimal solution.

Description

Technical Field

[0001] This invention relates to the field of intelligent review, and in particular to an intelligent review method and system for compliance of construction project bidding based on a multimodal large model. Background Technology

[0002] In the process of reviewing bidding for construction projects, the core data to be verified naturally falls into three information domains with vastly different representations: CAD drawings, which contain spatial geometric information composed of vector primitives, belong to the visual-spatial modality; bill of quantities, which contains semi-structured pricing data organized in tables, belong to the numerical-text modality; and tender document terms, which contain unstructured legal and technical requirements text, belong to the natural language modality.

[0003] Existing review methods have two main drawbacks: First, traditional methods rely primarily on keyword matching for retrieval when linking data across different data sources. Due to the significant differences in the representation of the three types of data, the association between component names in drawings and descriptions in the bill of quantities can easily break down when there are differences in wording (e.g., the drawing is labeled 'KZ-1' while the bill of quantities describes it as 'frame column'). This can lead to the omission of implicit compliance risks such as 'present on drawings but absent in the bill of quantities' or 'present in the bill of quantities but not in accordance with specifications'. Second, traditional methods generally use fixed static deviation thresholds for quantity deviations. This static approach cannot adapt to the differences in reasonable deviation ranges under different component types and different specification requirements (e.g., tolerances should be relaxed for low earthwork measurement accuracy, while tolerances should be tightened for high steel structure processing accuracy), resulting in false positives and false negatives.

[0004] Furthermore, in terms of bid compliance review, traditional static reviews typically only focus on whether the unit price exceeds the price limit, failing to accurately identify hidden risks caused by unbalanced bidding. If bidders increase the unit price of projects completed earlier and decrease the unit price of projects completed later, they can recover funds in advance by utilizing the time value of money, thus obtaining implicit financial benefits while keeping the total bid unchanged. Additionally, directly entrusting compliance judgment to large language models for probabilistic reasoning can easily lead to unstable and reproducible judgment results. Moreover, when the project scale is extremely large and the drawings contain millions of vector elements, constructing a complete topology graph globally can cause memory overflow, limiting the deployability of the algorithm. Summary of the Invention

[0005] One of the objectives of this invention is to provide an intelligent review method and system for compliance of construction project bidding based on a multimodal large model, in order to solve the problems of missed implicit compliance risks due to the broken association of cross-modal data caused by differences in wording, the unreproducible compliance judgment caused by direct reliance on probabilistic reasoning of large models, and memory overflow in ultra-large-scale drawing parsing in the existing technology.

[0006] This invention is achieved through the following technical solution: a method and system for intelligent review of compliance in construction project bidding based on a multimodal large model, comprising the following steps: acquiring a multimodal dataset of the construction project to be reviewed, wherein the multimodal dataset includes: first observation data distributed in different information domains, used to indicate the spatial geometric information of building components; second observation data used to indicate the pricing information of building components; and third observation data used to indicate the technical specification information constraining the pricing information of building components; converting the first observation data, the second observation data, and the third observation data into semantic feature representations in a unified semantic space, and binding semantic feature representations pointing to the same building component to the same semantic space according to the correlation index indicating the closeness of semantics. A unique identifier is used to obtain a set of semantically related entities consisting of multiple semantically related entities; based on the set of semantically related entities, a consistency check is performed on the first observation data and the second observation data bound to the same semantically related entity to obtain a first review result; the technical specification information indicated by the third observation data is converted into a set of deterministic constraints, and the quotation information indicated by the second observation data is substituted into the set of deterministic constraints for solution and verification to locate abnormal quotation nodes that deviate from the normal quotation range to obtain a second review result; the first review result and the abnormal quotation nodes are used as context input to a multimodal large model, and the multimodal large model generates an intelligent review report for the construction project to be reviewed.

[0007] Furthermore, at least one of the first observation data, the second observation data, and the third observation data is obtained by the processing device from at least one of a bidding and tendering transaction platform, an electronic tender document library, or a server communicatively connected to the processing device.

[0008] Furthermore, parsing the first observation data into an attribute graph model also includes a fragmentation process performed when the size of the first observation data exceeds a preset memory threshold: dividing the first observation data into multiple spatial sub-regions based on a spatial index, and setting an overlapping buffer zone between two adjacent spatial sub-regions; performing feature extraction and primitive conversion on each spatial sub-region to obtain local drawing structure data and storing it in an external temporary cache; when constructing the semantic association entity, deduplicating and merging redundant building components in the overlapping buffer zone of adjacent spatial sub-regions, and splicing the local drawing structure data of each spatial sub-region into the complete attribute graph model.

[0009] Furthermore, the width of the overlapping buffer band is greater than the outer rectangle size of the largest building component in the first observation data, so that the building component spanning the boundary of the spatial sub-region is fully extracted in at least one spatial sub-region.

[0010] Furthermore, the deduplication and merging process includes: when the distance between the centroids of two building components extracted from two adjacent spatial sub-regions is not greater than the preset spatial positioning tolerance, and the object categories and material properties of the two components are consistent, they are determined to be the same building component and merged into one node. The topological edges related to the building component in the two spatial sub-regions are merged into the merged node. The spatial attributes of the merged node are taken as the weighted average of the spatial attributes of the two merged nodes, and the cross-sectional dimension attribute is taken as the original value of either of the two merged nodes or the larger value.

[0011] Further, the first, second, and third observation data are converted into semantic feature representations in a unified semantic space, including: parsing the first observation data into an attribute graph model, wherein the attribute graph model uses building components as nodes and the spatial topological relationships between building components as edges, wherein each node is associated with the object category, spatial coordinates, geometric dimensions, and material properties of its respective building component, and each edge is associated with the spatial topological relationship type and spatial distance between the corresponding two building components; parsing the second observation data into a list structure data, wherein the list structure data records the code, name, and feature description information of each sub-item; and using a multimodal large model text encoder, mapping the attributes of the nodes in the attribute graph model, the feature description information of the sub-items in the list structure data, and the text paragraphs in the third observation data to the semantic feature representations of the same dimension in the unified semantic space.

[0012] Further, parsing the first observation data into an attribute graph model includes: reading vector metadata from the first observation data; extracting closed polygon boundary lines from the vector metadata and identifying corresponding object categories based on layer labels; calculating the bounding boxes of the closed polygon boundary lines and constructing a spatial adjacency matrix based on the spatial intersection relationship between the bounding boxes of different building components, wherein the spatial adjacency matrix is ​​used to represent the hierarchical relationship and connection topology between building components.

[0013] Further, binding semantic feature representations pointing to the same building component to the same unique identifier includes: calculating the cosine of the angle between two semantic feature representations in the unified semantic space to obtain the correlation index indicating the semantic closeness between them, and constructing a correlation degree measurement matrix; establishing semantic feature representations with correlation degree indices greater than a preset threshold as candidate associations, and binding the candidate associations to the same unique identifier to form the semantic association entity; taking the maximization of the sum of the correlation degree indices of all selected associations as the optimization objective, and applying a one-to-one constraint, wherein the one-to-one constraint indicates that each building component corresponding to the first observation data is associated with at most one sub-item corresponding to the second observation data, and each sub-item is associated with at most one text paragraph corresponding to the third observation data, and vice versa; solving under the optimization objective and the one-to-one constraint to obtain the globally optimal association scheme; and binding the mutually associated first observation data, second observation data, and third observation data to the same unique identifier according to the globally optimal association scheme.

[0014] Furthermore, the globally optimal association scheme is obtained by solving an integer linear programming problem, where the decision variable is constrained to take the value of zero or one, indicating whether the two corresponding objects are associated or not.

[0015] Further, the first review result is obtained through the following steps: for a specified semantically related entity, based on the geometric parameters indicated by the bound first observation data, and according to the morphological mapping relationship corresponding to the corresponding object category, the theoretical value of the first physical quantity of the building component is obtained by inversion; the declared value of the first physical quantity of the building component is extracted from the bound second observation data; compliance is determined based on the consistency deviation between the theoretical value of the first physical quantity and the declared value of the first physical quantity, and cases of data breakage or numerical deviation are recorded in the first review result, wherein the data breakage indicates that a building component existing in one observation data has no corresponding related item in another observation data, and the numerical deviation indicates that the consistency deviation exceeds the allowable range.

[0016] Furthermore, the morphological mapping relationship is based on the geometric conversion rules determined by the object category. Different object categories correspond to different geometric conversion rules. The morphological mapping relationship converts the geometric dimension parameters indicated by the bound first observation data into the physical quantities of the corresponding building components. The consistency deviation is obtained by calculating the ratio of the absolute value of the difference between the theoretical value and the declared value of the first physical quantity to the theoretical value of the first physical quantity.

[0017] Further, the compliance determination includes: comparing the consistency deviation degree with the adaptive deviation threshold; determining the corresponding semantically associated entity to be compliant when the consistency deviation degree is not greater than the adaptive deviation threshold; and recording the deviation value when the consistency deviation degree is greater than the adaptive deviation threshold. The adaptive deviation threshold is obtained by the combined effect of the industry benchmark allowable deviation rate, the object type correction coefficient, and the standard semantic correction coefficient. The object type correction coefficient varies with the object category of the building component, with a looser tolerance for coarse-type objects and a tighter tolerance for precise-type objects. The standard semantic correction coefficient varies with the semantic parsing result of the multimodal large model on the bound third observation data, with a tighter tolerance when the third observation data indicates precision control semantics and a looser tolerance when it indicates lenient measurement semantics.

[0018] Further, converting the technical specification information indicated by the third observation data into a deterministic constraint set includes: inputting the third observation data into a multimodal large model, locating condition trigger words, numerical boundary words, and penalty clauses in the third observation data through dependency parsing; converting the technical specification information into symbolic logical expressions based on the condition trigger words and the numerical boundary words; and constructing the deterministic constraint set from the symbolic logical expressions.

[0019] Furthermore, the deterministic constraint set includes: individual boundary constraints and joint proportional constraints. The individual boundary constraints indicate that the quotation data of the specified object does not exceed the corresponding unit price limit, and the joint proportional constraints indicate that the sum of the quotations of the controlled object set does not exceed a preset proportion of the total quotation.

[0020] Furthermore, individual boundary constraints also indicate that the quotation data of the specified object is not lower than the corresponding unit price lower limit, so as to apply a two-way boundary constraint to the quotation data of the specified object.

[0021] Furthermore, the process of locating abnormal pricing nodes that deviate from the normal pricing range includes: substituting the pricing data into the deterministic constraint set, and recording pricing data that does not meet the deterministic constraint set as out-of-bounds anomalies; performing unbalanced pricing analysis based on the deviation index of each pricing data from the benchmark reference value, and recording objects that are determined to be unbalanced pricing by the unbalanced pricing analysis as unbalanced pricing violations;

[0022] The out-of-bounds anomaly and the unbalanced bidding violation are jointly regarded as the abnormal bidding node; the deviation index is obtained by calculating the ratio of the difference between the bidding data and the benchmark reference value to the benchmark reference value, and the benchmark reference value is the benchmark market unit price obtained from historical bidding data, which indicates the normal price level of the corresponding object.

[0023] Furthermore, the unbalanced pricing analysis includes: assigning time decay weights to each object based on the sequential position of its corresponding process on the construction timeline, wherein the time decay weights decrease as the construction time of the corresponding process progresses later; calculating a directional imbalance index based on the deviation index, the time decay weights, and the budget proportion weight determined by the budget proportion for each object, wherein the directional imbalance index indicates the degree to which the pricing is directionally higher for preceding objects and lower for subsequent objects in the construction sequence; and determining that the corresponding pricing scheme has an unbalanced pricing when the directional imbalance index exceeds a preset sensitivity threshold.

[0024] Furthermore, the time decay weight is a discount factor that decays exponentially with the increase of the construction time position of the corresponding process, and the discount factor reaches its maximum value when the construction time position is the starting time.

[0025] Furthermore, the directional imbalance index is obtained by positively accumulating the deviation index of each object in the preceding object set according to the corresponding weight, and negatively subtracting the deviation index of each object in the following object set according to the corresponding weight. The preceding object set is the set of objects with earlier construction time, and the following object set is the set of objects with later construction time.

[0026] Furthermore, the unbalanced pricing analysis also includes: calculating a discrete imbalance index based on the deviation of each object's deviation index relative to the weighted average deviation rate, the budget proportion weight, and the time decay weight, wherein the discrete imbalance index indicates the degree of dispersion of the pricing deviation of each object; obtaining a comprehensive risk score based on a weighted combination of the directional imbalance index and the discrete imbalance index; determining a compliant pricing when the comprehensive risk score is not greater than a first risk threshold, determining a mild imbalance warning when the comprehensive risk score is greater than the first risk threshold but not greater than a second risk threshold, and determining an unbalanced pricing violation when the comprehensive risk score is greater than the second risk threshold, wherein the first risk threshold is less than the second risk threshold.

[0027] Furthermore, the generation of the intelligent review report includes: filling the object number and deviation data associated with the abnormal quotation node, as well as the deviation data in the first review result, into a preset structured assessment template to generate a structured risk data stream; the multimodal large model reads the structured risk data stream and matches it with the corresponding penalty clauses in the third observation data to generate the intelligent review report in natural language; the intelligent review report includes four dimensions: violation point location, quantitative analysis, corresponding clause support, and risk impact prediction. The violation point location is used to identify objects with compliance risks, the quantitative analysis is used to provide deviation data for the corresponding objects, the corresponding clause support is used to cite the corresponding penalty clauses in the third observation data, and the risk impact prediction is used to make a prediction of the consequences of the corresponding violations in combination with the bidding business background.

[0028] Another aspect of the present invention provides an intelligent review system for compliance of construction project bidding based on a multimodal large model, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements any of the intelligent review methods for compliance of construction project bidding based on a multimodal large model as described above.

[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0030] 1. This invention maps heterogeneous data belonging to different modalities to the same measurable semantic space and establishes associations with an association degree index that indicates the degree of semantic proximity, so that the association depends on semantic meaning rather than literal wording, thereby eliminating the modal gap and effectively avoiding the omission of implicit compliance risks; and by replacing the greedy pairwise matching with the globally optimal one-to-one association, it avoids the problem of the same component being repeatedly contested and the local high similarity crowding out the global optimal solution.

[0031] 2. This invention employs an adaptive deviation threshold that dynamically adjusts based on object type and specification semantics for consistency verification. This allows for dynamic tightening or loosening of tolerance based on the component and specification requirements, significantly reducing false alarms and missed alarms compared to conventional single fixed thresholds. Furthermore, by introducing a time decay weight to quantify the time value of money, unbalanced bidding behavior that exploits time differences in construction sequence can be effectively identified and quantitatively warned. The invention also uses a discrete imbalance index to capture extreme over-bidding that is masked by the mutual cancellation of positive and negative deviations, and finally achieves fine-grained risk classification through a comprehensive risk score.

[0032] 3. This invention places the large model only at the end of the review process, where it is solely responsible for translating unstructured text into symbolic expression and generating reports from data to language. Specific violation determinations are all completed by deterministic mathematical calculations. This division of labor—'mathematical calculations determine the conclusions, and the large model provides the expression'—ensures that compliance determinations are deterministic and reproducible, thus avoiding the risk of unreliable probabilistic outputs. Furthermore, through the mechanism of overlapping buffer strip segmentation and edge consistency maintenance, the algorithm can quickly process ultra-large-scale drawings with limited computing resources, and the results of segmentation processing are completely consistent with the results of global processing in terms of the entire topological relationship, thereby ensuring deployability and scalability. Attached Figure Description

[0033] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0034] Figure 1 This is a flowchart of the method provided in Embodiment 1 of the present invention.

[0035] Figure 2 This is a schematic diagram comparing the memory scalability of the drawing fragmentation mechanism provided in Embodiment 1 of the present invention.

[0036] Figure 3 This diagram illustrates the impact of component integrity extraction rate on Embodiment 1 of the present invention.

[0037] Figure 4 This is a heatmap of the cosine similarity metric matrix provided in Embodiment 1 of the present invention.

[0038] Figure 5 This is a comparison chart of total matching weights provided in Embodiment 1 of the present invention.

[0039] Figure 6 The adaptive deviation threshold provided in Embodiment 1 of the present invention is shown as a surface plot of the values ​​of the adaptive deviation threshold under different component types.

[0040] Figure 7 This is a scatter plot for consistency deviation verification provided in Embodiment 1 of the present invention.

[0041] Figure 8 This is a comparison chart of the false alarm / false negative rates of the adaptive threshold and the traditional fixed threshold provided in Embodiment 1 of the present invention.

[0042] Figure 9 This is a discount factor curve diagram under different discount rates provided in Embodiment 1 of the present invention.

[0043] Figure 10 This is a comparison chart of the unbalanced price identification and detection capabilities provided in Embodiment 1 of the present invention.

[0044] Figure 11 The price deviation rate distribution along the process time axis is provided for Embodiment 1 of the present invention.

[0045] Figure 12 This is a schematic diagram of the three-level judgment distribution of the comprehensive risk score provided in Embodiment 1 of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0047] Example 1

[0048] This embodiment discloses an intelligent review method for compliance of construction project bidding based on a multimodal large model. Figure 1 The flowchart of the method in this embodiment is shown. As can be seen from the flowchart, this embodiment includes the following steps:

[0049] Step 1: Obtain the multimodal dataset of the bidding for the construction project to be reviewed. The dataset includes: CAD drawing files, a PDF bill of quantities, and the bidding technical specifications text.

[0050] Among them, construction projects pending review refer to specific construction projects that are in the bidding or tender review stage and require verification of the compliance of their tender documents and the reasonableness of their quotations. These may include building construction projects, municipal infrastructure projects, highway and bridge projects, or decoration and renovation projects.

[0051] A multimodal dataset for bidding and tendering refers to a set of heterogeneous data sources distributed across multiple completely different information domains, carrying the core information of the bidding and tendering process for a construction project. In other words, the core data to be verified in the review of construction project bidding and tendering is distributed across three information domains with vastly different representational forms:

[0052] CAD drawings contain spatial geometric information and belong to the visual-spatial modality composed of vector primitives; bill of quantities contains semi-structured project pricing data and belongs to the numerical-text modality organized in tabular form; tender document clauses contain unstructured legal and technical requirements text and belong to the natural language modality.

[0053] CAD drawing files are electronic construction drawings that describe the design content of various disciplines of a building in the form of vector primitives. They record the geometric and attribute data of points, lines, polylines, layers, and blocks that make up the building components, and carry the spatial geometry and topological information of the building components. For example, CAD drawing files can be construction drawings for architecture, structure, plumbing, electrical and other disciplines in DWG or DXF format.

[0054] A bill of quantities is a tabular document prepared by the tendering party or the tenderer to list all physical quantities of the proposed project and their corresponding prices. It is usually submitted in PDF format and includes information such as the bill of quantities code, project name, characteristic description, unit of measurement, quantity of work and comprehensive unit price for each sub-item of the project.

[0055] The technical specifications for bidding refer to the unstructured natural language clauses in the bidding documents that stipulate the materials, processes, quality standards, measurement rules, price adjustment boundaries, payment nodes, and penalties for rejected bids. They serve as the legal basis for determining whether bidding behavior is compliant.

[0056] In this embodiment, CAD drawing files, bill of quantities, and bidding technical specifications can be obtained by the review system from the bidding transaction platform, electronic bid document library, or a server connected to it.

[0057] It should be noted that, due to the significant differences in the representation of the three types of data mentioned above, traditional review methods mainly rely on keyword collision for cross-data source correlation retrieval. This leads to a break in the correlation between component names in drawings and descriptions in the list, for example, the drawing may label it as 'KZ-1' while the list describes it as 'frame column'. This can result in overlooking implicit compliance risks such as 'present on drawings but absent in the list' or 'present in the list but not conforming to specifications'. This embodiment eliminates this modal gap between heterogeneous data through a cross-modal semantic alignment mechanism in subsequent steps.

[0058] In addition, to avoid memory overflow during large-scale CAD drawing parsing, this step also includes a drawing sharding caching mechanism for large-scale data scenarios. When the scale of a construction project is extremely large and the CAD drawing contains millions of vector elements, directly constructing a complete drawing topology map globally can lead to excessive memory overhead and even memory overflow. To ensure the deployability and scalability of this solution in real-world engineering scenarios, a spatial sharding and edge consistency maintenance mechanism needs to be designed, enabling the algorithm to process ultra-large-scale CAD drawings with limited computing resources, while ensuring that the result after sharding is consistent with the result of global processing in terms of topological relationships. The core idea of ​​this mechanism consists of two steps:

[0059] The first step is to divide the super-large drawing into several sub-regions that can be processed independently according to their spatial location, with the size of each sub-region controlled within the range that computing resources can bear. The second step is to identify and remove duplicate components in the boundary areas of adjacent regions after the segmentation process is completed, and to re-establish the cross-boundary topological connection relationship, thereby merging the local topological maps of each region into a complete global topological map.

[0060] In this embodiment, the drawing fragment caching mechanism may specifically include:

[0061] When the size of the CAD drawing file to be parsed exceeds the preset memory threshold, the spatial R-tree index is activated to divide the drawing into regions, cutting the complete CAD drawing into multiple spatial sub-drawing areas with overlapping boundaries.

[0062] Parallel feature extraction and primitive transformation are performed on each spatial sub-drawing area, and the generated local drawing structure data is stored in an external temporary cache.

[0063] When constructing semantic anchor points, redundant primitives on overlapping boundaries are merged to achieve deduplication and splicing of local drawing structure data.

[0064] For spatial partitioning, this algorithm adopts a regularized spatial partitioning strategy, defining the minimum two-dimensional bounding rectangle of the entire CAD drawing as the global spatial range, and setting an upper limit on the maximum number of entities processed in a single operation. When the total number of entities in the drawing exceeds this upper limit, the global spatial range is divided into several sub-regions. To ensure that components located at the boundaries of adjacent sub-regions are not cut off or omitted due to partitioning, an overlap buffer zone of a certain width is set between two adjacent sub-regions.

[0065] For example, in this embodiment, the minimum two-dimensional bounding rectangle of the entire CAD drawing is defined as:

[0066] Let the maximum number of entities processed in a single operation be [value]. When the total number of entities in the drawings At that time, Divided into Sub-regions:

[0067]

[0068] The width of the overlap buffer zone between adjacent sub-regions is... This can be achieved through the following boundary definition:

[0069]

[0070] in, The width parameter of the overlapping buffer zone should be greater than the outer rectangle dimension of the largest component in the drawing to ensure that any component crossing the boundary is fully extracted in at least one sub-region. Understandably, it is precisely because of the width of the overlapping buffer zone... Only by ensuring that the size is greater than the maximum component size can we prevent a situation where a component is cut in two by the boundary but cannot be fully identified in any sub-region. Figure 2 This diagram illustrates a comparison of the memory scalability of the drawing fragmentation mechanism in this embodiment. Figure 3 This diagram illustrates the effect of the width parameter of the overlapping buffer band on the component integrity extraction rate in this embodiment.

[0071] After the piecewise computation is completed, the same physical component may be repeatedly extracted from two adjacent sub-regions within the boundary overlap zone. To eliminate this ghosting and restore the correct topological relationship, component deduplication and topological merging are required. By comparing the spatial centroid coordinates and attribute information of the components extracted from the two sub-regions, it is determined whether they are the same physical component. If the spatial centroids are close enough and the component category and material properties are completely consistent, they are determined to be the same component, merged into a single node, and the topological edges related to this component in the two sub-regions are uniformly merged into the new merged node.

[0072] For example, in this embodiment, the spatial centroid coordinates of the component geometry are defined as:

[0073] For adjacent areas and Components extracted and When both of the following conditions are met, the components are considered to be the same physical component and are merged:

[0074] One is the geometric nearest neighbor condition. ,in, The first requirement is extremely small spatial positioning tolerance; the second is the condition of consistency in category and attributes. and .

[0075] Understandably, spatial proximity alone is insufficient to determine that they are the same component (e.g., two adjacent columns of different materials). Therefore, it is necessary to simultaneously combine the conditions of consistent category and material properties to reliably determine that the two extracted results are actually the ghost images of the same physical component.

[0076] The new node after the merger The spatial attributes are updated to a weighted average of the attributes of the two source nodes:

[0077]

[0078] in, and These are the geometric parameter vectors of the two source nodes. This weighted averaging process is suitable for correcting small deviations in spatial coordinates. For deterministic properties such as the cross-sectional dimensions of components, in practical implementation, the original value of either source node or a larger value should be used to avoid unnecessary accuracy loss introduced by the averaging process. After merging, the topological adjacency matrix is ​​updated synchronously, and the topological edges across the boundaries are reconstructed. This allows for the seamless integration of local topology maps of each region into a complete global topology map, providing high-quality input data for subsequent cross-modal semantic alignment.

[0079] Step 2: Perform spatial-text bidirectional mapping on the bidding multimodal dataset, extract the feature representation of the same building component in different modalities, map them uniformly and bind them to the same semantic anchor point of the building component, forming a semantic anchor point set.

[0080] Among them, space-text bidirectional mapping refers to the process of converting CAD drawing information belonging to the spatial geometric modality and list text and standard text belonging to the natural language modality into calculable and measurable mathematical expressions in a unified semantic space, and establishing a precise relationship between the three in this unified space.

[0081] A semantic anchor point for a building component refers to a ternary association formed by binding together drawing elements, bill of quantities items, and specification paragraphs describing the same physical building component (e.g., a beam, a column, or a slab) after semantic alignment. It is the basic calculation unit for all subsequent consistency checks and compliance reviews.

[0082] The semantic anchor set refers to the set of semantic anchors of all successfully established building components in the project to be reviewed.

[0083] To eliminate the modal gap between heterogeneous data, the three types of data are first transformed into a unified mathematical expression, and then the globally optimal association match is found in a unified semantic space, thereby replacing the traditional retrieval mechanism that relies solely on keyword collision.

[0084] In this embodiment, in order to transform a complex engineering drawing into a structured model that can be traversed and calculated by an algorithm, the CAD drawing file is parsed, its spatial topology information is extracted and expressed as an attribute graph model, including:

[0085] Treat each identifiable building component (e.g., beam, column, slab, wall, etc.) in the drawing as a node in the drawing;

[0086] The spatial relationship between components (e.g., inclusion, support, adjacency, etc.) is treated as an edge in the diagram;

[0087] Each node is accompanied by key attribute information of the component, including component category, three-dimensional spatial coordinates and geometric dimension parameters, and material properties, thereby completely transforming the visual information in the drawings into structured data records;

[0088] Record the spatial topological relationship type between two components and their Euclidean distance on each edge, thereby preserving the spatial context information between components.

[0089] Among them, the attribute graph model refers to a mathematical model used to carry the spatial topology information of CAD drawings in a structured manner.

[0090] For example, in this embodiment, the parsed CAD drawing can be represented as a directed attribute graph:

[0091] ,

[0092] in For a set of nodes, It is a set of directed edges.

[0093] Node set Each node in Represents a building component entity. Given the total number of nodes in a CAD component, the attribute vector of this component entity is defined as follows: .

[0094] in, The component category is used to identify the type of component represented by the node, such as beam, column, slab, wall, etc.

[0095] Let be the coordinate position and geometric dimensions of the component in three-dimensional space, where , , The center coordinates of the component , , These are the width, height, and depth of the enclosure box for this component, respectively. This refers to the material properties marked on the drawings, such as C30 concrete, HRB400 steel bars, etc.

[0096] Directed edge set Each edge in Representative components and The spatial topological relationship between them, where, It refers to the spatial topology type, such as containment, support, or adjacency relationships; For components and The Euclidean distance between them. It can be understood that, through the attribute graph model, the components and their spatial relationships, originally existing visually in drawings, are losslessly transformed into structured data that the algorithm can directly read and traverse.

[0097] In this embodiment, in order to accurately reconstruct the component entities and their spatial adjacency relationships from vector primitives, the parsing of the CAD drawing file specifically includes:

[0098] Read vector image metadata from CAD drawing files;

[0099] Extract the boundary lines of closed polygons from the metadata of the vector graphics and identify the corresponding building component types based on the layer tags;

[0100] Calculate the bounding box of the closed polygon boundary line, and construct the component spatial adjacency matrix based on the spatial intersection relationship of the bounding boxes of different components to represent the hierarchy and connection topology between components.

[0101] Vector graphics metadata refers to the original drawing data stored in CAD drawing files in the form of geometric objects such as points, lines, arcs, and polylines. A closed polygon boundary line is a closed curve formed by connected geometric elements, representing the outer contour of a component, and serves as the geometric basis for separating independent components from the drawing.

[0102] A bounding box is the smallest rectangular region that can completely enclose the boundary line of a closed polygon. It is determined by the extreme coordinates of the boundary line along each coordinate axis and is used to describe the spatial range occupied by a component in a simplified rectangular approximation. A component spatial adjacency matrix is ​​a matrix used to record whether there is a spatial intersection, containment, or adjacency relationship between any two components. Its rows and columns correspond to the nodes of each component, and the values ​​of the matrix elements reflect the type of topological relationship between the corresponding two components.

[0103] Understandably, by determining whether different component enclosures intersect or contain each other in space, the hierarchical subordination and connection topology between components can be inferred purely from geometric relationships without relying on the text annotations on the drawings. Examples include the support relationship between a column and the beam it supports, and the adjacency relationship between a floor slab and its surrounding beams.

[0104] After structuring the CAD drawings, in this embodiment, in order to unify text information belonging to different modalities into a single measurable vector space, it is also necessary to transform the bill of quantities and technical specification texts into vector representations in the same semantic space, including:

[0105] The table area is located and the text is extracted from the bill of quantities table to obtain the bill of quantities code, project name and feature description information of each sub-item, and generate bill of quantities structure data.

[0106] The bidding technical specification text is segmented into paragraphs and entity recognition is performed to extract the process text paragraphs containing specific component construction techniques, material requirements and implementation standards.

[0107] By using a multimodal large model text encoder, the feature description text of the list items and the process paragraphs in the technical specification text are mapped to high-dimensional semantic feature vectors of the same dimension.

[0108] The bill of quantities structure data refers to the data records organized by sub-items after the bill of quantities table is structured and parsed. Each record contains the bill of quantities code, project name, and feature description information for that sub-item. The text encoder is a functional component in a multimodal large model used to convert a piece of natural language text into a fixed-dimensional numerical vector. Its output vector can reflect the semantic meaning of the original text in numerical space.

[0109] For example, in this embodiment, the bill of quantities sub-items The vectorized representation can be obtained by the following formula:

[0110]

[0111] in, For the list of The feature description text of the item can be, for example, 'precast reinforced concrete column, C30 concrete', etc. A text encoder for a multimodal large model; the output is A high-dimensional semantic feature vector.

[0112] Similarly, in the technical specification text, the first... Section process section The vectorized representation can be obtained by the following formula:

[0113]

[0114] in, For the first Section construction process and material requirements text. It is understandable that, since the list of items and specification sections are encoded using the same encoder... The mapping places the two within the same measurable high-dimensional semantic space, thus laying the foundation for subsequent calculations of their semantic closeness.

[0115] After all three types of data have been transformed into a unified mathematical expression, in this embodiment, in order to establish a precise correlation between them, the semantic similarity between the list feature description information and the process text paragraphs and the drawing component names is calculated based on the drawing structure data, including:

[0116] Since the attributes of CAD component nodes contain unstructured text information (such as material names and component category labels), their attributes are textualized and then mapped into semantic vectors through a text encoder.

[0117] By calculating the cosine similarity between CAD component vectors and list item vectors, and between list item vectors and specification paragraph vectors, we can measure their semantic similarity and thus construct a cross-modal similarity measurement matrix.

[0118] Cosine similarity is an index that measures the semantic closeness of two vectors by calculating the cosine of the angle between them in a high-dimensional space. Its core idea is that the smaller the angle between two vectors in a high-dimensional space, the closer the semantic information they carry, and the greater the possibility of matching.

[0119] For example, in this embodiment, CAD components The attribute vector can be denoted as:

[0120] ,

[0121] in, This is a string describing the attributes of the component in text form.

[0122] The cosine similarity matrix between CAD components and bill of quantities items, and between bill of quantities items and specification paragraphs, can be calculated using the following formula:

[0123]

[0124]

[0125] in, For the first The CAD component and the first Semantic similarity between the sub-items of the list; For the first Each list item and the first The semantic similarity between standard paragraphs; the numerator is the dot product of two vectors, and the denominator is... Let be the Euclidean norm of the vector. Understandably, because the calculation of this similarity depends on the semantic meaning of the text rather than the literal wording, even if the drawing is labeled 'KZ-1' and the list is described as 'frame column', the encoded vectors of the two are still highly similar in semantic space, and thus can be correctly associated. Figure 4 A heatmap of the cosine similarity metric matrix in this embodiment is shown.

[0126] After obtaining the similarity between all pairs, in this embodiment, in order to find a set of overall optimal association schemes from a global perspective, drawing components, list items, and process texts with similarity greater than a preset threshold are associated with the same unique identifier as semantic anchors for building components, including:

[0127] The optimization objective is to maximize the sum of similarities of all matching pairs, while imposing a one-to-one physical constraint: each CAD component can be associated with at most one list item, each list item can be associated with at most one specification paragraph, and vice versa.

[0128] Solve this constrained optimization problem to obtain the globally optimal three-way matching scheme;

[0129] The successfully matched CAD components, list items, and specification paragraphs are bound to the same unique identifier to form semantic anchors for building components.

[0130] Establishing associations with similarity exceeding a preset threshold as matching relationships essentially involves finding a set of globally optimal association schemes. It's important to note that, compared to a locally greedy approach that directly pairs items with similarity exceeding the threshold, this embodiment considers the sum of similarities across all matching pairs from a global perspective. This avoids the problem of multiple items repeatedly competing for the same component, or local high-similarity matching crowding out the globally optimal solution. Essentially, this is an integer linear programming problem, and a variant of the bipartite graph maximum weight matching problem.

[0131] For example, in this embodiment, the cross-modal three-way aligned integer linear programming optimization model can be represented by the following equation:

[0132]

[0133] The following constraints must be met:

[0134]

[0135] Here, the objective function represents the sum of the semantic similarities of all selected matching pairs; CAD components With list of sub-items Matching decision variables between them This indicates that the two are matched and associated. This indicates that the two are not related; For list items With standard paragraphs Matching decision variables between them This indicates that the two are matched and associated; This represents the total number of nodes in the CAD component. Total number of items in the list; To standardize the total number of paragraphs. Figure 5 A comparison chart of the total matching weights of global integer linear programming and local greedy thresholding method in this embodiment is shown.

[0136] Understandably, the first two equations in the above constraints ensure that each CAD component matches at most one list item, and each list item is matched by at most one CAD component; the latter two equations ensure that each list item matches at most one specification paragraph, and each specification paragraph is matched by at most one list item. The decision variables are constrained to integers of 0 or 1, reflecting the physical rationale of a one-to-one matching principle of 'either associated or not associated'.

[0137] After the optimization solution is completed, if and This indicates a CAD component. List of items With standard paragraphs The three were successfully linked.

[0138] For example, in this embodiment, the first semantic anchors of building components It can be defined as:

[0139]

[0140] in, For the CAD component node associated with this anchor point, For the list items associated with this anchor point, This refers to the canonical paragraph associated with this anchor point. All successfully established anchor points. Together, they form a set of semantic anchors. It can be understood that these semantic anchors of a building component bind together the characteristic representations of the same physical component across three modalities, serving as the basic computational unit for all subsequent consistency checks and compliance reviews.

[0141] Step 3: Based on the semantic anchor set, traverse the semantic anchors of each building component, extract the corresponding drawing design parameters, list feature descriptions and construction process requirements, perform cross-modal consistency comparison, and output the first review result of data breakage or numerical deviation.

[0142] Cross-modal consistency comparison refers to the process of verifying whether the geometric parameters in the drawings are consistent with the engineering quantity data filled in the bill of quantities in terms of physical and mathematical logic within the framework of each semantic anchor point. Its purpose is to discover compliance issues such as false reporting or omission of engineering quantities.

[0143] The first review result refers to the verification conclusion output through cross-modal consistency comparison, which records data breakage or numerical deviation. Data breakage refers to a situation where a building component exists in one modality but cannot be found in another modality; for example, a component is drawn in the drawings but not listed in the bill of quantities. Numerical deviation refers to a situation where the difference between the theoretical quantity of the same component in the drawings and the quantity declared in the bill of quantities exceeds the allowable range.

[0144] In this embodiment, in order to reverse-engineer the theoretically required physical engineering quantity of the component from the geometric information of the drawings, for any specified semantic anchor point of the building component, the theoretical value of the design engineering quantity of the component is calculated from the associated drawing data. ,include:

[0145] Select the corresponding geometric volume or area calculation formula based on the type of component (such as beam, column, slab, etc.);

[0146] Substituting the dimensional parameters extracted from the CAD drawings into the calculation formula yields a theoretical engineering quantity independently derived from the drawing information, which serves as a benchmark for verifying the declared quantity in the bill of quantities.

[0147] Among them, the theoretical value of the engineering quantity in the drawing design refers to the physical engineering quantity that the component should theoretically have, which is derived solely from the geometric dimension parameters provided by the CAD drawings and inversely calculated according to the geometric calculation rules of the corresponding component category. This theoretical value is completely independent of the data submitted in the bill of quantities and is derived solely from the drawings, thus serving as an objective and reliable verification benchmark.

[0148] For example, in this embodiment, for anchor points Utilizing CAD nodes Spatial geometric parameters Based on the geometric shape mapping function of building components Inversion calculation of its theoretical physical engineering quantity It can be obtained through the following formula:

[0149]

[0150] in, According to component category A defined geometric calculation function; different categories of components correspond to different geometric calculation rules.

[0151] For example, if If it is a 'concrete rectangular column', then That is, the concrete volume of the column is calculated by multiplying the cross-sectional width by the cross-sectional height and then by the column height. This can be understood as a geometric shape mapping function. It acts as a bridge between abstract geometric parameters and specific physical engineering quantities, enabling the algorithm to apply appropriate measurement rules to components of different shapes such as beams, columns, slabs, and walls.

[0152] After obtaining the theoretical quantities of work, it is necessary to compare them with the actual quantities declared in the bill of quantities. In this embodiment, the declared quantity of the component is extracted from the associated bill of quantities entry. That is, list items The declared project quantity filled in .

[0153] It should be noted that, due to the allowable reasonable loss rates or minor differences caused by measurement standards between construction drawings and the actual bill of quantities, the deviation threshold should not be a fixed static constant. The reasonable deviation range varies depending on the type of component and the specific specifications; for example, earthwork engineering, due to its lower measurement accuracy, allows for larger deviations, while steel structures, due to their high processing precision requirements, should allow for smaller deviations. Furthermore, if the technical specifications explicitly require settlement based on actual quantities or strict control of losses, the allowable deviation range should be correspondingly tightened.

[0154] Therefore, in this embodiment, to avoid false positives and false negatives caused by hard-coded thresholds, this algorithm designs an adaptive deviation threshold function, which can dynamically adjust the allowable deviation range according to the specific engineering component type and the semantic content of the specification clauses, including:

[0155] The component type correction coefficient is used to distinguish between extensive and precision components. For extensive components such as earthwork, a positive correction is given to relax the tolerance, while for precision components such as fine decoration and steel structure, a negative correction is given to tighten the tolerance.

[0156] Semantic signals regarding precision control in the normative text are captured by using normative constraint correction coefficients;

[0157] By combining the industry benchmark allowable deviation rate with the two correction coefficients mentioned above, an adaptive deviation threshold for the current anchor point is obtained.

[0158] Among them, the adaptive deviation threshold refers to the tolerance boundary that is dynamically adjusted according to the component type and specification semantics, and is used to determine whether the deviation of the engineering quantity constitutes a violation.

[0159] For example, in this embodiment, the adaptive deviation threshold It can be calculated using the following formula:

[0160]

[0161] in, This is the industry benchmark allowable deviation rate, for example, it can be 3%; This is a component type correction factor, the value of which is related to the component category. Relatedly, for extensive structural components such as earthwork... For precision components such as fine decoration and steel structure And its value range is constrained as follows: To ensure threshold factor Always positive; To standardize the constraint correction coefficients, their values ​​depend on the large model's understanding of the standard text. The semantic parsing results, when When it contains semantics that emphasize precision, such as 'settlement based on actual costs' and 'strict control of losses'... When it contains semantics that allow for loose measurement, such as 'estimate' or 'provisional estimate' Its value range is also constrained. . Figure 6 The surface plots showing the adaptive deviation threshold values ​​for different component types in this embodiment are illustrated.

[0162] It is understandable that this is achieved through a component type correction factor. Correction factor for normative constraints Product-based adjustment, adaptive deviation threshold Based on the actual situation in two dimensions, namely "what kind of component" and "how the specification requires", the system can dynamically tighten or loosen the standards on the basis of industry benchmarks. This avoids both oversensitivity to false alarms for rough components and leniency in underreporting for precision components.

[0163] After determining the adaptive threshold, the consistency deviation between the theoretical quantities and the quantities declared in the bill of quantities can be calculated, and compliance can be determined accordingly. In this embodiment, the theoretical values ​​of the quantities designed in the drawings are calculated. Compared with the declared project quantity The relative deviation rate between them is determined, and it is judged whether it is greater than the preset engineering quantity error threshold, including:

[0164] Calculate the absolute value of the difference between the theoretical quantity of work and the declared quantity of work, and then divide it by the theoretical quantity of work to obtain the consistency deviation.

[0165] The consistency deviation is compared with the adaptive deviation threshold. If the deviation is not greater than the threshold, the engineering quantity corresponding to the anchor point is determined to be compliant. If the deviation is greater than the threshold, the semantic anchor point of the current building component is marked as a data deviation violation and the deviation value is recorded.

[0166] For example, in this embodiment, the consistency deviation is... It can be calculated using the following formula:

[0167]

[0168] in, These are the theoretical physical engineering quantities obtained by inverting the geometric parameters from CAD drawings; For list items The quantity of work declared in the bill of quantities. It is understandable that the absolute value of the difference in the numerator means that regardless of whether the quantity declared in the bill of quantities is higher or lower than the theoretical quantity in the drawings (i.e., whether it is false or omitted), the deviation will be taken into account; by dividing by the theoretical quantity of work, the deviation becomes a dimensionless relative quantity, so as to make it possible to use a unified standard for judgment between components of different quantities.

[0169] Furthermore, compliance determination rules can be defined as indicator functions. :

[0170]

[0171] in, This indicates that the deviation in the quantity of work corresponding to the anchor point is within the adaptive allowable range, and is therefore deemed compliant. This indicates that the deviation exceeds the adaptive threshold. This item is marked as 'abnormal deviation in project quantity' and requires further manual verification. It will be included in the first review result as a data deviation violation. Understandably, this is achieved using an adaptive threshold. The deviation judgment rule in this embodiment can dynamically tighten or loosen the tolerance according to the specific engineering components and specification clauses. Compared with the judgment method using a single fixed threshold, it significantly reduces the probability of false alarms and false alarms. Figure 7 A scatter plot of the consistency deviation verification in this embodiment is shown; Figure 8 A comparison chart of false alarm / false negative rates between the adaptive threshold in this embodiment and the traditional fixed threshold is shown.

[0172] It should be noted that if a certain modality is found to be unique while other modalities are missing during the process of traversing semantic anchors, that is, there is no matching item to match it, it will be recorded as a data break, which, together with the above-mentioned numerical deviation, constitutes the first review result.

[0173] Step 4: Use a multimodal large model to parse the bidding technical specification text, identify the unstructured price adjustment boundary conditions, material price limit rules and payment node logic, and convert the unstructured price adjustment boundary conditions, material price limit rules and payment node logic into symbolic logical expressions, thereby generating a deterministic set of mathematical constraint equations.

[0174] Among them, unstructured price adjustment boundary conditions refer to the financial rules in the bidding documents, expressed in natural language, that constrain bid prices, including: upper and lower limits on the unit price of individual materials or processes, and proportional restrictions on the total price of several sub-items.

[0175] Symbolic logical expressions refer to precise expressions composed of mathematical symbols and variables, translated from the aforementioned natural language financial rules. A system of mathematical constraint equations is a set of equations and inequalities, composed of all symbolic logical expressions, used to constrain the range of values ​​for the bid price vector.

[0176] It's important to note that the core idea of ​​this step is to use the large model as a translator between unstructured text and structured mathematical expressions, ensuring that subsequent numerical calculations are based on precise symbolic logic. This design makes the final compliance judgment result deterministic and reproducible, thus avoiding the instability issues that may arise from directly relying on the probabilistic output of the large model.

[0177] In this embodiment, to accurately locate the financial rules that need to be transformed from unstructured text, the bidding technical specification text is input into a multimodal large model. Dependency parsing is used to locate conditional trigger words, numerical boundary words, and penalty clauses in the text, including:

[0178] By leveraging the natural language understanding capabilities of large models, dependency parsing is performed on standard texts to analyze the dominance and subordination relationships among the components of a sentence.

[0179] Based on the syntactic structure, locate conditional trigger words (e.g., words that indicate constraint relationships such as 'should not exceed' and 'must not be lower than'), numerical boundary words (e.g., specific price limits, percentage values, etc.), and punitive clauses (e.g., expressions such as rejection of bids, deduction of points, and penalties for breach of contract) in the text.

[0180] The located semantic components are extracted and used as the basis for constructing symbolic constraints in the future.

[0181] Among them, dependency parsing refers to the processing method of analyzing the dependency relationships between words in a natural language sentence to reveal the grammatical structure of the sentence; condition trigger words are words in a sentence that indicate that a certain constraint condition is about to appear; numerical boundary words are words in a sentence that represent the numerical boundary of the constraint; and punitive clauses are clauses that stipulate the consequences of violations.

[0182] After locating the above semantic components, constraints can be extracted and mathematical constraint equations can be constructed based on conditional trigger words and numerical boundary words. This includes extracting two typical financial constraints: one is the individual unit price limit constraint set for certain specific materials or processes, and the other is the combination ratio constraint that requires the total price of certain sub-items to not exceed a certain percentage of the total price.

[0183] For example, in this embodiment, let the unit price of all items in the bill of quantities be composed of a vector:

[0184] ,

[0185] in, The transpose symbol for a matrix. This represents the total number of items in the list.

[0186] Individual unit price limit constraints can be expressed as:

[0187]

[0188] in, For list items The unit price of the bid; The unit price is the upper limit specified in the tender documents for this item. This represents the set of all sub-items subject to price limits. It should be noted that, in a more general case, this constraint may also include a lower price limit, expressed as... ,in For the corresponding project's price floor, This sets a price ceiling, thereby imposing a two-way boundary constraint on the unit price.

[0189] Furthermore, the combination ratio constraint can be expressed as:

[0190]

[0191] in, For list items The amount of work; A set of specific process sub-items subject to combination ratio constraints; The upper limit of the ratio specified in the bidding documents; the right side of the equation This represents the total price obtained by multiplying the quantity and unit price of all items in the bill of quantities and summing them up. The whole represents the set of controlled items. The total price must not exceed the total quoted price. times.

[0192] It is understandable that this combined proportional constraint is essentially a multivariate joint constraint, and its more general form can be written as: ,in, For a controlled set of subheadings, The weighting coefficient (corresponding to the quantity of work in the above proportional constraints) is the weighting coefficient. (The conversion factor, which is determined together with the upper limit of the proportion) This is a maximum limit on the total amount.

[0193] A deterministic set of mathematical constraint equations is formed by all individual unit price limits and combined ratio constraints, which serves as the basis for subsequent price calculation and dispersion determination.

[0194] Step 5: Extract the pricing data from the bill of quantities table, input it as an independent variable into the mathematical constraint equation system to solve for extreme values ​​and calculate the pricing dispersion, locate abnormal pricing nodes that deviate from the normal pricing range, and form the second review result.

[0195] The quotation data refers to the comprehensive unit price data filled in for each item in the bill of quantities. Abnormal quotation nodes refer to items in the bill of quantities that, after solution and calculation, are determined to deviate from the normal quotation range. The second review result refers to the verification conclusion output through constraint checks and unbalanced pricing analysis, recording violations of boundary-crossing anomalies and unbalanced pricing.

[0196] In this embodiment, to first identify quotations that explicitly violate the price limit rules, the unit prices of each item in the quotation data are substituted into a system of mathematical constraint equations to check whether all constraints are satisfied, including:

[0197] Substitute the unit prices of each item in the quotation data into the mathematical constraint equations;

[0198] Verify that each unit price satisfies all individual boundary constraints and multivariate joint constraints;

[0199] Record the unit price of the sub-item that does not meet the constraints as an out-of-bounds exception.

[0200] It should be noted that out-of-bounds anomalies can only identify bids that explicitly exceed the price limit, but cannot identify hidden risks where every unit price is within the price limit, but the overall bidding structure shows a speculative tendency. The latter is the unbalanced bidding problem that this step further aims to address.

[0201] Imbalanced pricing is a common compliance risk in the bidding process: bidders increase the unit price of projects expected to be completed earlier and decrease the unit price of projects to be completed later, thereby gaining hidden financial benefits by leveraging the time value of money to recover more funds in advance without changing the total bid. Traditional static review methods only focus on whether the unit price exceeds the price limit and cannot identify this kind of financial maneuvering based on the construction sequence.

[0202] In order to identify such hidden risks, this embodiment obtains a set of preliminary work procedures for the construction project to be reviewed. and post-process project collection And retrieve the corresponding benchmark market unit price vector from the historical bidding database, including:

[0203] Based on the sequential position of each item in the bill of quantities along the construction timeline, they are divided into sets of preliminary work items. Collection of later process items ;

[0204] The benchmark market unit price for each item is retrieved from the historical bidding database to form a benchmark market unit price vector.

[0205] Among them, the "early-stage process item set" refers to the set of sub-items with relatively earlier construction times; the "late-stage process item set" refers to the set of sub-items with relatively later construction times; the benchmark market unit price refers to the reference unit price that reflects the normal market price level of each sub-item, obtained based on historical bidding data statistics, and is denoted as... .

[0206] After obtaining the division of pre- and post-processing and the benchmark market unit price, in this embodiment, the deviation between the actual price and the benchmark market unit price of each item in the list is calculated, an unbalanced pricing analysis model is constructed, and the distribution characteristic index of the deviation is calculated to identify whether it belongs to the type of unbalanced pricing with a heavy pre-processing and light post-processing.

[0207] It's important to note that the core of unbalanced bidding lies in the time value of money; the earlier the same amount of money is received, the higher its value. Therefore, comparing only the absolute deviation of unit prices is insufficient; the timing of the deviation must also be considered. Bidders tend to raise the unit price of projects completed earlier to receive higher progress payments sooner, and lower the unit price of projects completed later. This is because the losses from later payments have a smaller impact after discounting, allowing them to profit from the time value of money while keeping the total bid unchanged.

[0208] To mathematically quantify this time-dimensional financial game precisely, this embodiment introduces a time value of money discount factor. The core principle is that for sub-items with earlier construction times, the unit price deviation has a greater weight after discounting; for sub-items with later construction times, the weight of the unit price deviation naturally decreases after discounting.

[0209] For example, in this embodiment, each process is defined. The midpoint of the estimated construction time is:

[0210] ,

[0211] in The time value of money discount factor is the total project duration. It can be calculated using the following formula:

[0212]

[0213] in, It is a natural constant; The industry's continuously compounded discount rate (or the social discount rate); This is the midpoint of the estimated construction time for this process. Figure 9 The discount factor curves for different discount rates in this embodiment are shown.

[0214] It should be noted that when (That is, the initial construction process of the project) When the discount factor reaches its maximum value, it means that the unit price deviation of this process is fully factored in; when When the work is larger (i.e., a later stage of construction), As the index declines, it reflects the natural depreciation of the value of forward funds.

[0215] Based on this, in this embodiment, the unit price deviation rate is defined. The relative deviation of the bid unit price from the market benchmark can be calculated using the following formula:

[0216]

[0217] in, The unit price is the bid price. This is the market benchmark unit price. It's understandable that... This indicates that the bid price is higher than the market price. This indicates a price lower than the market price. This relative deviation eliminates the impact of the absolute magnitude difference in unit prices of different sub-items, allowing the degree of price deviation for each sub-item to be compared on a uniform scale.

[0218] In this embodiment, the unbalanced pricing analysis model can be achieved by calculating the unbalanced pricing index. This is to identify unbalanced pricing where the front is heavier than the back.

[0219] For example, in this embodiment, the time-discounted imbalance pricing index It can be calculated using the following formula:

[0220]

[0221] in, Subheading The unit price deviation rate; Subheading The discount factor for the time value of money in the corresponding process; Subheading The weighting coefficient, determined by its proportion of the budget, can be obtained using the following formula:

[0222]

[0223] in, Subheading The amount of work; Its market benchmark unit price; the denominator is the total budget calculated for all sub-items based on the benchmark market unit price, thus ensuring that the sum of all weights is 1. It is understandable that... The larger the positive value, the more significant the net effect of the bidder's excess bids on the high-priced items in the early stage after discounting and weighting, and the stronger the tendency of unbalanced bidding.

[0224] It should be noted that, due to the discount factor Collection of preliminary processes The sub-items in the list are given greater weight, which affects the set of subsequent processes. The sub-items in the formula are assigned smaller weights. In effect, it is equivalent to accumulating positive deviations in earlier processes with high weights and deducting negative deviations in later processes with low weights.

[0225] Therefore, in an equivalent implementation, the unbalanced quote index It can also be explicitly grouped according to the previous process. Collection of later processes Grouping is expressed as:

[0226]

[0227] in, and Each of the bidders in the preliminary work projects and later process projects The actual quote; and These are the corresponding benchmark market unit prices; and These are weighting coefficients assigned based on the process timeline. It's understandable that the former measures the degree of imbalance by positively accumulating positive deviations in the early stages and negatively deducting deviations in the later stages; the latter achieves the same goal through the inherent difference in discounting weights. Both use the deviation structure of higher reporting in the early stages and lower reporting in the later stages as a distinguishing feature for unbalanced pricing.

[0228] In this embodiment, when the calculated unbalanced quote index Greater than the preset sensitivity threshold If the bidder's pricing scheme is found to be unbalanced and violates regulations, the corresponding item in the bill of quantities will be output as an abnormal pricing node. Figure 10 A comparison chart of the unbalanced price identification and detection capabilities in this embodiment is shown.

[0229] It should be noted that relying solely on the time-discounted unbalanced pricing index may lead to a situation where bidders significantly raise prices for some items in the early stages while simultaneously significantly lowering prices for others. This results in a weighted sum that cancels each other out, masking the actually very extreme price manipulation. To better capture the risk of excessive pricing dispersion, this embodiment introduces a pricing dispersion penalty model. Its core principle is to consider not only the directionality of pricing deviations (heavier at the beginning and lighter at the end) but also the degree of dispersion of these deviations across different items. In other words, the more dispersed and extreme the deviations, the higher the speculative risk, even if the weighted mean is close to zero.

[0230] For example, in this embodiment, the weighted dispersion of the quote deviation It can be calculated using the following formula:

[0231]

[0232] in, This is the weighted average deviation rate; The variance contribution of the deviation rate of a single sub-item relative to the weighted mean; The time discounting factor gives higher weight to extreme bids from earlier sub-items in the dispersion calculation. This is understandable, as dispersion... It measures the degree of dispersion of the price deviation of each sub-item from the overall value, even the directional index. Because the positive and negative deviations cancel each other out and approach zero, as long as there is extreme stretching in the previous sub-project, It can still be captured. Figure 11 The diagram shows the distribution of the quotation deviation rate along the process time axis in this embodiment.

[0233] Finally, in this embodiment, an unbalanced pricing comprehensive risk score is constructed by comprehensively considering the discount imbalance index (used to measure directional deviation) and dispersion (used to measure extreme deviation).

[0234] For example, in this embodiment, the unbalanced pricing comprehensive risk score It can be calculated using the following formula:

[0235]

[0236] in, and The preset risk control weighting coefficient is used to balance the contributions of directional bias and extreme bias in the final score.

[0237] The system performs a three-level judgment based on the solution results:

[0238]

[0239] in, and The preset risk level threshold, and Understandably, compared to relying solely on a single sensitivity threshold... The binary judgment method and the three-level judgment can set a transitional range of mild imbalance warning between compliance and non-compliance, thus providing reviewers with a more granular risk classification reference. Sub-items judged as having unbalanced pricing violation warnings, along with the aforementioned out-of-bounds anomalies, are all output as abnormal pricing nodes, jointly constituting the second review result. This is achieved by introducing a time discount factor. This scheme mathematically quantifies the extent to which the time value of money can be obtained, enabling the effective identification and quantification of unbalanced bidding behavior that exploits time differences in construction sequence. Figure 12 A schematic diagram of the three-level judgment distribution of the comprehensive risk score in this embodiment is shown.

[0240] Step 6: Feedback the abnormal pricing node and the first review result as contextual prompts to the multimodal big data model. The multimodal big data model then combines the bidding and tendering business background to generate a compliance intelligent review report.

[0241] The context of the prompts refers to organizing the structured review data output from the preceding steps into an input context that can be understood and continued by the multimodal big data model. The compliance intelligent review report is a comprehensive verification conclusion document generated by the multimodal big data model in natural language, intended for reviewers.

[0242] In this embodiment, to ensure that the final report combines the accuracy of mathematical conclusions with the readability of natural language expression, the specific steps for generating a compliance intelligent review report include:

[0243] The list number, deviation rate, unbalanced pricing index associated with the identified abnormal pricing nodes, as well as the deviation data in the first review results, are filled into the preset structured compliance assessment template to generate a structured risk data stream.

[0244] By using a multimodal large model to read structured risk data streams and matching them with corresponding breach of contract penalties, rejection clauses, or deduction standards in the bidding technical specifications, a compliance intelligent review report is generated in natural language.

[0245] Among them, the structured compliance assessment template refers to a pre-designed template used to carry various review conclusions in a regular field format. The structured risk data flow refers to the structured data sequence formed by filling in the various review conclusions according to the template fields.

[0246] It should be noted that this step places the large language model at the end of the review process rather than in the core judgment stage. Its responsibility is to generate a representation from data to language, while the specific violation determinations are all completed by the deterministic mathematical calculations in the preceding steps. This division of labor—where mathematical calculations determine the conclusions and the large language model provides the representation—leverages the advantages of the large language model in natural language organization and understanding of business context, while avoiding the unreliability risks associated with entrusting crucial judgments to probabilistic models.

[0247] In this embodiment, the compliance intelligent review report generated using natural language includes four dimensions: 'violation point location, quantitative analysis of mathematical features, corresponding clause support, and risk impact prediction'. Specifically, the violation point location dimension identifies the specific list item or building component with compliance risks; the quantitative analysis of mathematical features provides quantitative indicators such as the deviation rate and unbalanced pricing index corresponding to the violation; the corresponding clause support dimension cites the breach penalties, rejection clauses, or deduction standards in the technical specifications text corresponding to the violation; and the risk impact prediction dimension, combined with the commercial context of the bidding process, makes a prediction of the potential compliance consequences and commercial impact of the violation.

[0248] Understandably, the above four dimensions enable the review report not only to point out where the problem is, but also to explain how serious the problem is, which regulation has been violated, and what the consequences will be, thus providing a complete closed-loop basis for review decisions.

[0249] Example 2

[0250] This embodiment discloses an intelligent review system for compliance of construction project bidding based on a multimodal large model.

[0251] Specifically, the intelligent review system for compliance of construction project bidding based on a multimodal large model can be integrated into electronic devices, such as terminals and servers. Terminals can be mobile phones, tablets, smart Bluetooth devices, laptops, or personal computers; servers can be a single server or a server cluster composed of multiple servers.

[0252] In this embodiment, the intelligent review system for compliance of construction project bidding based on a multimodal large model can also be integrated into multiple electronic devices. For example, the intelligent review system for compliance of construction project bidding based on a multimodal large model can be integrated into multiple servers, and the method in embodiment 1 of this application can be implemented by multiple servers.

[0253] In this embodiment, the server can also be implemented in the form of a terminal.

[0254] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent review of compliance in construction project bidding based on a multimodal large model, characterized in that, The intelligent review method includes: Obtain a multimodal dataset of the construction projects to be reviewed. The multimodal dataset includes: first observation data distributed across different information domains, used to indicate the spatial geometric information of building components; second observation data used to indicate the pricing information of building components; and third observation data used to indicate the technical specification information that constrains the pricing information of building components. The first observation data, the second observation data, and the third observation data are converted into semantic feature representations in a unified semantic space. Based on the correlation index that indicates the degree of semantic proximity, the semantic feature representations pointing to the same building component are bound to the same unique identifier, resulting in a set of semantically related entities composed of multiple semantically related entities. Based on the set of semantically associated entities, a consistency check is performed on the first observation data and the second observation data bound to the same semantically associated entity to obtain a first review result; The technical specification information indicated by the third observation data is converted into a set of deterministic constraints. The bidding information indicated by the second observation data is then substituted into the set of deterministic constraints for solution verification to locate abnormal bidding nodes that deviate from the normal bidding range, thereby obtaining the second review result. The first review result and the abnormal quotation node are used as context inputs to the multimodal large model, and the multimodal large model generates an intelligent review report for the construction project to be reviewed.

2. The intelligent review method for compliance of construction project bidding based on a multimodal large model as described in claim 1, characterized in that, The first observation data, the second observation data, and the third observation data are converted into semantic feature representations in a unified semantic space, including: The first observation data is parsed into an attribute graph model. The attribute graph model uses building components as nodes and the spatial topological relationship between building components as edges. The nodes are attached with the object category, spatial coordinates, geometric dimensions and material properties of the building component to which they belong. The edges are attached with the spatial topological relationship type and spatial distance between the corresponding two building components. The second observation data is parsed into a list structure data, which records the code, name and feature description information of each sub-item in units of sub-items; Using a multimodal large model text encoder, the attributes of nodes in the attribute graph model, the feature description information of sub-items in the list structure data, and the text paragraphs in the third observation data are respectively mapped to the semantic feature representation of the same dimension in the unified semantic space.

3. The intelligent review method for compliance of construction project bidding based on a multimodal large model as described in claim 1, characterized in that, Binding semantic feature representations pointing to the same building component to the same unique identifier includes: By calculating the cosine of the angle between the two semantic features in the unified semantic space, the correlation index indicating the semantic closeness between the two is obtained, and a correlation degree measurement matrix is ​​formed. The semantic features whose correlation index is greater than a preset threshold are identified as candidate associations, and the candidate associations are bound to the same unique identifier to form the semantic association entity. The optimization objective is to maximize the sum of the association degree indices of all selected associations, and a one-to-one constraint is applied, wherein each building component corresponding to the first observation data is associated with at most one sub-item corresponding to the second observation data, and each sub-item is associated with at most one text paragraph corresponding to the third observation data, and vice versa. The globally optimal correlation scheme is obtained by solving under the optimization objective and the one-to-one constraints. Based on the global optimal association scheme, the first observation data, the second observation data, and the third observation data that are related to each other are bound to the same unique identifier.

4. The intelligent review method for compliance of construction project bidding based on a multimodal large model as described in claim 1, characterized in that, The first review result is obtained through the following steps: For a specified semantically associated entity, based on the geometric parameters indicated by the first observation data bound to it, and according to the morphological mapping relationship corresponding to the corresponding object category, the theoretical value of the first physical quantity of the building component is obtained by inversion; Extract the first physical quantity declaration value of the building component from the bound second observation data; Compliance is determined based on the consistency deviation between the theoretical value and the declared value of the first physical quantity, and cases of data breakage or numerical deviation are recorded in the first review result. The data breakage indicates that a building component present in one observation data has no corresponding related item in another observation data, and the numerical deviation indicates that the consistency deviation exceeds the allowable range.

5. The intelligent review method for compliance of construction project bidding based on a multimodal large model as described in claim 4, characterized in that, The morphological mapping relationship is as follows: Based on the geometric conversion rules determined by the object category, different object categories correspond to different geometric conversion rules. The morphological mapping relationship converts the geometric dimension parameters indicated by the first observation data into the physical quantities of the corresponding building components. The consistency deviation is obtained by calculating the ratio of the absolute value of the difference between the theoretical value and the declared value of the first physical quantity to the theoretical value of the first physical quantity. The compliance determination includes: comparing the consistency deviation degree with the adaptive deviation threshold; determining that the corresponding semantically related entity is compliant when the consistency deviation degree is not greater than the adaptive deviation threshold; and recording the deviation value when the consistency deviation degree is greater than the adaptive deviation threshold. The adaptive deviation threshold is obtained by combining the industry benchmark allowable deviation rate, the object type correction coefficient, and the specification semantic correction coefficient. The object type correction coefficient varies with the object category of the building component, with a looser tolerance for coarse-grained objects and a tighter tolerance for fine-grained objects. The normative semantic correction coefficient is determined by the semantic parsing result of the multimodal large model on the bound third observation data, and the tolerance is tightened when the third observation data indicates precision control semantics and loosened when it indicates lenient econometric semantics.

6. The intelligent review method for compliance of construction project bidding based on a multimodal large model as described in claim 1, characterized in that, The step of converting the technical specification information indicated by the third observation data into a set of deterministic constraints includes: The third observation data is input into a multimodal large model, and the conditional trigger words, numerical boundary words, and penalty clauses in the third observation data are located through dependency parsing. Based on the conditional trigger words and the numerical boundary words, the technical specification information is converted into a symbolic logical expression; The symbolic logical expressions constitute the deterministic constraint set; The deterministic constraint set includes: individual boundary constraints and joint proportional constraints. The individual boundary constraints indicate that the quotation data for the specified object does not exceed the corresponding unit price ceiling. The joint proportion constraint indicates that the sum of the bid prices of the controlled object set does not exceed a preset proportion of the total bid price.

7. The intelligent review method for compliance of construction project bidding based on a multimodal large model as described in claim 1, characterized in that, The location of abnormal pricing nodes that deviate from the normal pricing range includes: Substitute the quoted data into the deterministic constraint set, and record the quoted data that does not meet the deterministic constraint set as an out-of-bounds exception; Based on the deviation index of each quotation data from the benchmark reference value, an unbalanced quotation analysis is performed, and objects that are determined to be unbalanced quotations through the unbalanced quotation analysis are recorded as unbalanced quotation violations. The out-of-bounds anomaly and the unbalanced pricing violation are collectively considered as the abnormal pricing node; The deviation index is obtained by calculating the ratio of the difference between the quoted price data and the benchmark reference value to the benchmark reference value. The benchmark reference value is a benchmark market unit price obtained from historical bidding data, indicating the normal price level of the corresponding object.

8. The intelligent review method for compliance of construction project bidding based on a multimodal large model as described in claim 7, characterized in that, The unbalanced pricing analysis includes: Based on the sequential position of the corresponding process on the construction timeline, a time decay weight is assigned to each object, and the time decay weight decreases as the construction time of the corresponding process is later. Based on the deviation index of each object, the time decay weight, and the budget proportion weight determined by the budget proportion, a directional imbalance index is calculated. The directional imbalance index indicates the degree to which the price is higher for the preceding object and lower for the following object in the construction sequence. When the directional imbalance index is greater than a preset sensitivity threshold, it is determined that the corresponding bidding scheme has an unbalanced bid. The time decay weight is a discount factor that decays exponentially with the increase of the construction time position of the corresponding process, and the discount factor reaches its maximum value when the construction time position is the starting moment.

9. The intelligent review method for compliance of construction project bidding based on a multimodal large model according to claim 1, characterized in that, The generation of the intelligent review report includes: The object number and deviation data associated with the abnormal quotation node, as well as the deviation data in the first review result, are filled into a preset structured assessment template to generate a structured risk data stream. The structured risk data stream is read from the multimodal large model, and the corresponding penalty clauses in the third observation data are matched to generate the intelligent review report in natural language. The intelligent review report includes four dimensions: violation location, quantitative analysis, supporting clauses, and risk impact prediction. The violation point location is used to identify objects with compliance risks, the quantitative analysis is used to provide deviation data for the corresponding objects, the corresponding clause support is used to cite the corresponding penalty clauses in the third observation data, and the risk impact prediction is used to make a prediction of the consequences of the corresponding violations in combination with the business background of bidding.

10. A smart review system for compliance of construction project bidding based on a multimodal large model, characterized in that, The intelligent review system includes: processor; The memory stores a computer program, which, when executed by a processor, implements the intelligent review method for compliance of construction project bidding based on a multimodal large model as described in any one of claims 1 to 9.

Images