A method for engineering quantity summarization based on large model and large data analysis intelligence

Abstract

The application relates to the technical field of building engineering informatization, and discloses a method for intelligently collecting engineering quantities based on large models and big data analysis, which comprises the following steps: constructing a semantic vector mapping space of historical data and generating a standardized historical feature library; analyzing a building information model and calling semantic vector mapping space standardized component attributes to generate an engineering graph containing connection probability weights; searching for a benchmark project cluster in the historical feature library to extract ontological constraints, logically authenticating the probability connection attributes in combination with connection frequencies, and converting probability edges into certain connection edges; calling a large language model to convert engineering change instructions into a structured graph query script, screening target nodes and aggregating explicit and implicit engineering quantities. Finally, double closed-loop verification based on statistical distribution and model quality feedback is performed. The application repairs model defects through a probability graph and historical statistical constraints, solves the problem of fuzzy instruction analysis, and realizes accurate collection of engineering quantities.

Description

Technical Field

[0001] This invention relates to the field of information technology in building engineering, specifically a method for summarizing engineering quantities based on large models and big data analysis intelligence. Background Technology

[0002] Building Information Modeling (BIM) technology has been widely applied in engineering cost management, enabling the automated extraction of quantities by analyzing the geometric dimensions and attribute information of components in the model. Existing mainstream quantity calculation software relies on the accuracy and standardization of model data, calculating explicit indicators such as component volume, surface area, and length through Boolean operations and rule-matching algorithms. This deterministic rule-based calculation model requires the modeling process to strictly adhere to geometric and topological constraints to ensure the correctness of the connections between components.

[0003] In actual engineering projects, due to limitations in the accuracy of modeling software or human error, building information models (BIMs) commonly contain geometric defects such as gaps between components, minor overlaps, or non-manifold structures. Traditional rigid geometric algorithms cannot identify adjacency relationships with minor errors, leading to invalid deduction relationships or omissions in the calculation of hidden quantities such as contact area. Furthermore, engineering change orders and design modification instructions are often issued in unstructured natural language text format. Existing cost estimation software relies on keyword matching or manual adjustments, which cannot parse semantically ambiguous change orders. Directly applying general-purpose large language models, lacking engineering domain ontology constraints and contextual parameters, easily results in erroneous query scripts or data indexes.

[0004] Existing geometric accuracy deficiencies and semantic parsing obstacles mean that the current quantity aggregation process still requires significant manual intervention for model repair and instruction conversion, hindering the efficiency and accuracy of cost analysis. Existing technical solutions fail to effectively utilize the statistical patterns of historical engineering big data to help correct topological errors in the current model, and lack an automated closed-loop verification mechanism for calculation results, making it difficult to meet the high accuracy and automation requirements of cost data in complex engineering projects. Therefore, this invention proposes a method for quantity aggregation based on large models and big data analysis intelligence to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for summarizing engineering quantities based on large models and big data analysis intelligence, which solves the problems of data loss caused by geometric defects in models and low efficiency in processing engineering change instructions in traditional engineering quantity calculation.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for summarizing engineering quantities based on large-scale models and big data analysis intelligence, comprising:

[0007] This includes constructing a semantic vector mapping space for historical data and generating a standardized historical feature library based on semantic distance filtering.

[0008] During the construction process, classification items are read from the historical database, and a domain embedding model is invoked to map classification items and standard terms into high-dimensional feature vectors. The Euclidean distance between the high-dimensional feature vectors corresponding to the classification items and the standard vectors corresponding to the standard terms is calculated. The standard terms corresponding to the minimum Euclidean distance are found by traversing the standard engineering ontology library. It is then determined whether the minimum Euclidean distance is less than or equal to the semantic validity threshold. If the determination result is yes, the mapping relationship between classification items and standard terms is established and stored in the historical feature library; if the determination result is no, the data is removed as noise.

[0009] In the method, the building information model is parsed, the semantic vector mapping space is invoked to map component attributes to standard terms, and the components are instantiated as attribute graph nodes. An engineering graph containing probabilistic connectivity attributes and connectivity probability weights is generated based on the grid geometric distance.

[0010] Specifically, non-geometric attributes of physical components in the Building Information Model (BIM) are extracted, and the distance between the feature vectors corresponding to the non-geometric attributes and the standard vectors in the semantic vector mapping space is calculated to determine standardized component attributes. Simultaneously, physical components are instantiated as component nodes in the attribute graph, and the axis-aligned bounding boxes of the component entities are calculated. For component node pairs coarsely filtered by the axis-aligned bounding boxes, the minimum geometric distance between the entity meshes is calculated. When the minimum geometric distance is less than the fine tolerance, probabilistic edges are generated between component nodes, and the connection probability weights of the probabilistic edges are calculated based on the minimum geometric distance. These connection probability weights are set to decay exponentially with increasing minimum geometric distance.

[0011] The method further retrieves benchmark project clusters from a standardized historical feature library, extracts ontology constraints, and logically assigns weights to probabilistic connection attributes in the engineering graph based on statistically derived connection logic constraints, converting probabilistic edges into deterministic connection edges. This step includes extracting global feature parameters of the current project and mapping them to the current project feature vector, calculating the cosine similarity between the current project feature vector and the historical project feature vectors recorded in the historical feature library. Historical projects with cosine similarity higher than the similarity screening threshold are selected to form benchmark project clusters, and component connection frequencies are extracted from the benchmark project clusters as ontology constraints.

[0012] In the rights confirmation calculation, the connection probability weights of the probabilistic edges are obtained, the connection frequencies of physical connections between different types of components in the benchmark project cluster are read, and the product of the connection probability weights and connection frequencies is calculated to obtain the rights confirmation criterion function value. If the rights confirmation criterion function value is greater than the rights confirmation threshold, the data type of the probabilistic edges is updated to a defined connection edge, and it is marked as an automatically repaired edge.

[0013] Furthermore, the method combines ontology constraints and attribute default values ​​to construct a prompt context, and calls a large language model to convert engineering change instructions into a structured graph query script. The specific steps are: receiving the engineering change instruction text in natural language form; extracting high-frequency attributes as attribute default values ​​from the baseline project cluster in the historical feature library; constructing a prompt context containing graph data, attribute default values, and the engineering change instruction text; and calling a large language model to generate a structured graph query script containing matching path clauses, filtering condition clauses, and attribute update clauses.

[0014] In this context, the default attribute values ​​in the structured graph query script are used to complete the parameters not mentioned in the engineering change instruction text in the attribute update clause, ensuring the completeness of the graph database query.

[0015] Subsequently, the method executes a structured graph query script to filter target component nodes and aggregates explicit and implicit engineering quantities based on topology propagation. By running the structured graph query script to filter the set of target component nodes and aggregating the explicit geometric attributes of each node in the target component node set, and by performing Boolean operations to determine connecting edges through traversal, the area integral or overlapping volume integral of the geometric contact surface between the component node and its adjacent node entities is calculated to obtain the implicit engineering quantity; finally, the explicit geometric attributes and implicit engineering quantities are summed to obtain the total engineering quantity.

[0016] Finally, the method performs a dual closed-loop verification based on statistical distribution and model quality feedback, outputting the final summary result of the engineering quantities. When verifying the calculation results, the engineering quantity unit index for the current project is calculated, and a standard score for the engineering quantity unit index is calculated based on the mean and standard deviation of the benchmark project cluster index; if the absolute value of the standard score exceeds the preset standard deviation range, the result is deemed abnormal. When verifying model quality, the proportion of automatically repaired edges to the total number of connected edges is statistically analyzed, and it is determined whether the proportion exceeds a set quality threshold; if it does, a model quality warning is output, prompting a check of the modeling accuracy of the building information model.

[0017] This invention provides a method for summarizing engineering quantities based on large-scale models and big data analysis intelligence. It has the following beneficial effects: 1. This invention solves the problem of inaccurate judgment of component connection relationships caused by geometric modeling errors in Building Information Modeling by constructing an engineering graph containing probabilistic connection attributes and connection probability weights, and combining the connection frequency of historical benchmark project clusters for logical weighting. Furthermore, by combining the connection probability weights generated by grid geometric distance with the statistical regularity of historical data, uncertain geometric proximity relationships are transformed into definite logical connections. This can automatically repair the model topology without manual intervention, thereby ensuring the integrity and accuracy of implicit engineering quantity calculation.

[0018] 2. This invention combines high-frequency attributes of historical benchmark project clusters as default values ​​and constructs a prompt context with engineering change instructions. This solves the problem of structured script generation errors caused by missing parameters when the large language model processes fuzzy engineering instructions. Furthermore, by introducing prior attribute constraints from historical data, it automatically completes unmentioned filtering conditions or updates parameters during the generation of structured graph query scripts, ensuring the closed loop of graph database query logic and the accuracy of execution. This achieves automated conversion from unstructured text to accurate engineering quantity calculation.

[0019] 3. This invention constructs a standardized historical feature library based on a semantic vector mapping space and performs dual closed-loop verification based on statistical distribution and model quality feedback, overcoming the data silos caused by non-standard naming of historical data and the lack of verification basis for calculation results. Semantic alignment is achieved by calculating the Euclidean distance of high-dimensional feature vectors. This method can aggregate the statistical features of heterogeneous projects as a benchmark. The dual verification based on this can both identify statistical anomalies in the calculation results through standard scores and reverse-evaluate the modeling quality of the building information model through the proportion of automatically repaired edges, providing multi-dimensional credibility assurance for quantity surveying. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the overall process of the method of the present invention; Figure 3 This is a schematic diagram comparing the accuracy and recall of different semantic recognition methods of the present invention; Figure 4 This is a schematic diagram of the decay curve of the connection probability as a function of the gap in this invention; Figure 5 This is a schematic diagram showing the dual-axis comparison of the time taken to summarize the engineering quantities and the error rate in this invention; Figure 6 This is the system confidence convergence curve of the present invention with the project iteration cycle.

[0021] Among them, 10 is the semantic alignment module; 20 is the graph construction module; 30 is the constraint extraction module; 40 is the logic compilation module; 50 is the calculation execution module; 60 is the verification feedback module; 200 is the data server; 201 is the graph storage node; 202 is the vector storage node; and 203 is the historical database. Detailed Implementation

[0022] The technical solutions in 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 2 This invention provides a system for summarizing engineering quantities based on large models and big data analysis. The engineering quantity summarization system is deployed in a computing environment and performs engineering quantity calculations through data standardization, graph reconstruction, and logical constraint injection.

[0024] The hardware environment of the quantity summary system includes a data server 200, a graph storage node 201, a vector storage node 202, and a historical database 203. The data server 200 is equipped with a processor and storage media. The storage media records computer instructions, and the processor executes these instructions to implement data processing logic. The data server 200 communicates with the graph storage node 201, the vector storage node 202, and the historical database 203 via a data bus. The graph storage node 201 stores the building model data with attribute graph structures. The vector storage node 202 stores the high-dimensional feature vectors of historical terms and the feature vectors of projects. The historical database 203 stores the cleaned historical quantity list and design specifications.

[0025] The engineering quantity summary system includes collaboratively operating logical units: semantic alignment module 10, graph construction module 20, constraint extraction module 30, logic compilation module 40, calculation execution module 50, and verification feedback module 60.

[0026] The semantic alignment module 10 establishes a mapping between the original data and the standard ontology library. The semantic alignment module 10 reads the classification items from the historical database 203 and calls the domain embedding model to convert the classification items into feature vectors. The semantic alignment module 10 calculates the Euclidean distance between the feature vectors and the standard term vectors, maps the classification items to the standard terms with the smallest distance, and removes noisy data based on a semantic threshold to generate a historical feature library.

[0027] The graph construction module 20 parses the model and instantiates the components as nodes, calculates the AABB and mesh geometric distance, generates probabilistic edges when the distance is less than the fine tolerance, and determines the connection probability weights accordingly.

[0028] The constraint extraction module 30 retrieves the benchmark cluster based on the similarity of the project feature vectors, combines the connection logic constraints and probability weights obtained from statistics to determine the probabilistic edges, and converts them into deterministic connection edges.

[0029] The logic compilation module 40 combines change instructions and attribute constraints to build prompts, and calls the large language model to generate a structured graph query script that includes node filtering and attribute update operations.

[0030] The calculation and execution module 50 executes a script to filter nodes, aggregates explicit attributes, and calculates the topological integral of the contact surface to output implicit engineering quantities. The verification and feedback module 60 determines anomalies based on the standard score of the engineering quantity index; if the proportion of repair probability edges exceeds a threshold, an early warning is output.

[0031] See attached document Figure 1 This invention provides a method for summarizing engineering quantities based on large models and big data analysis, comprising the following steps: Construct a semantic vector mapping space for historical data and generate a standardized historical feature library; This step aims to address the issues of non-standard naming and heterogeneous classification in historical engineering data. Its core principle is to map discrete text strings to a continuous high-dimensional vector space, using geometric distance in the vector space to represent the semantic similarity between terms, thereby overcoming the dependence of traditional keyword matching on complete character consistency.

[0032] In practice, the system calls a fine-tuned BERT domain-specific word embedding model to map the original entries and standard terms into high-dimensional feature vectors. Semantic similarity is represented by calculating the Euclidean distance between these vectors. The dimensionality of the high-dimensional feature vectors... It is usually set to 256 dimensions or 512 dimensions to balance computational efficiency and semantic expressiveness.

[0033] Calculate the Euclidean distance between the high-dimensional feature vector and the standard vector. The smaller the Euclidean distance, the closer their semantics are. The formula for calculating the Euclidean distance is as follows: ; in, Indicates semantic Euclidean distance. This indicates a non-standard original entry. Indicates standard terminology, This represents the high-dimensional feature vector corresponding to the non-standard original entry. This represents the standard vector corresponding to the standard terminology. The dimension of a high-dimensional feature vector is represented by its length. and They represent the vector at the th order. Dimensional component values.

[0034] The system traverses the standard engineering ontology library to find the standard terminology with the smallest Euclidean distance. If the distance is greater than the semantic validity threshold, the system will proceed. Then remove noise; if it is less than or equal to the semantic validity threshold... Then a mapping is established and stored in the historical feature database.

[0035] This step involves parsing the Building Information Model (BIM) and generating an engineering graph containing probabilistic connectivity attributes. It transforms the traditional hierarchical BIM model into an attribute graph structure suitable for graph computation. Addressing the minor gaps commonly found in 3D modeling due to floating-point precision or human error, this method introduces a probabilistic connectivity mechanism. Instead of directly determining a break, it preserves the possibility of a connection for subsequent logical judgment.

[0036] First, physical components in the Building Information Model (BIM) are instantiated as nodes in the attribute graph. After coarse filtering using axis-aligned bounding boxes (AABB), the minimum geometric distance of the mesh is calculated. If the distance is less than the fine tolerance... The formula for generating probabilistic edges and assigning connection probability weights that decay exponentially with distance is as follows: ; in, Represents component nodes With component nodes The connection probability weights between them, with values ​​ranging from 1 to 2. ; Indicates the distance attenuation coefficient. The value of is determined by the model accuracy requirements, and is usually taken as [0.5, 2.0]. A larger value indicates greater sensitivity to distance errors; Represents component nodes With component nodes Minimum geometric distance between solid meshes; This indicates a potential neighbor pair.

[0037] The ontological constraints of the historical baseline cluster are extracted, and the probabilistic connectivity attributes are logically weighted. This step utilizes the statistical patterns of historical data to verify the spatial relationships of the current model. The principle is that although the geometric data of the current model may contain errors, the construction logic of similar buildings is relatively stable.

[0038] Extract the global feature parameters (including building type, seismic resistance level, and climate zone) of the current project to be calculated, and map them to the current project feature vector. Calculate the cosine similarity between the current project feature vector and the feature vectors of historical projects, using the following formula: ; in, Represents cosine similarity. This represents the feature vector of the current project. This represents the feature vector of a historical project.

[0039] Set a similarity screening threshold (e.g., 0.85), and select historical items with a cosine similarity higher than this threshold to form a baseline item cluster. Calculate the prior probability (i.e., connection frequency) of physical connections between different types of components within the baseline item cluster. Define the rights determination criterion function: ; in, Indicates the type in the benchmark project cluster Types in component and benchmark project clusters Connection frequency of components. Setting a weighting threshold. (Usually 0.6). If This indicates that despite the existence of a geometric gap, the two should logically be connected. In this case, the data type of the probabilistic edge is updated to a defined connection edge, and the edge is marked as an auto-repair edge.

[0040] Combine ontology constraints to convert engineering change instructions into graph database query scripts; This step utilizes a large language model as a logic compiler. It receives engineering change instructions in natural language form. High-frequency attributes are extracted from the baseline project cluster as attribute default value constraints.

[0041] Construct a context containing graph data, attribute default values, and change instructions, and generate MATCH (matching path), WHERE (filtering condition), and SET (attribute update) clauses. Attribute default value constraints are used to complete parameters not mentioned in the instructions in the SET clause.

[0042] The script executes a graph database query and aggregates implicit engineering quantities based on topology propagation. It then runs a script to filter target nodes, directly aggregates explicit attributes, and calculates geometric intersections and unions by traversing connecting edges and performing Boolean operations, thus deriving the implicit engineering quantities. The formula for calculating the total engineering quantities is as follows: ; in, Indicates the total amount of work; Represents the set of target component nodes; Represents component nodes The explicit geometric properties (such as its own volume); Represents component nodes The set of adjacent nodes; This represents the implicit engineering quantity function, specifically implemented by calculating component nodes. With component nodes The integral of the geometric contact surface area or the integral of the overlapping volume of the solid.

[0043] Perform a dual closed-loop validation based on statistical distribution and model quality feedback. First, verify the calculation results: calculate the unit quantity index for the current project. Based on the law of large numbers, this index should follow a normal distribution. Calculate the standard score: ; in, Represents standard scores. This indicates the unit index of the current project's workload. and These are the mean and standard deviation of the benchmark cluster indicators, respectively. If (i.e., exceeding the range of 3 standard deviations), the result is judged as abnormal.

[0044] Next, verify the model quality: Statistically analyze the proportion of automatically repaired edges to the total number of connected edges. Set a quality threshold. (For example, 15%). If the percentage of automatically repaired edges is greater than... If so, a model quality warning will be output.

[0045] See attached document Figure 1 The semantic space construction method based on a domain-specific embedding model of the present invention includes the following steps: A corpus and ontology database were constructed. During data cleaning, stop words and pure unit symbols were removed, while alphanumeric specification identifiers were retained. The cleaning rules were set as follows: remove meaningless stop words and pure unit symbols (such as kg, m). 2 (e.g., C30, Φ25), but retains the specification markings consisting of a mixture of letters and numbers (e.g., C30, Φ25).

[0046] A domain-specific word embedding model was built and fine-tuned based on a pre-trained language model architecture; the BERT model based on the Transformer encoder architecture was selected as the base network. The base network is typically configured with a 12-layer Transformer encoder, a hidden layer dimension of 768, and 12 attention heads.

[0047] The base network is fine-tuned for domain adaptation using a pre-trained corpus. The training task employs a full-word masking strategy, with mask probabilities set. The threshold is 15%. This means randomly masking 15% of the words in the input sequence and updating the model weights by minimizing the prediction error. The principle is to force the model to infer the missing words using engineering context, thereby encoding the co-occurrence relationships of engineering terms into the network parameters.

[0048] A domain-specific word embedding model is used to map text entries into fixed-dimensional semantic feature vectors; the model is then fine-tuned by taking the text to be processed as input. The model's output layer generates a dynamic sequence of word vectors with a dimension of [missing information]. ,in For sequence length, The hidden layer dimension is [dimensional]. To obtain a fixed-length vector that can represent the semantics of the entire phrase or sentence, instead of using only the first tag ([CLS]), an average pooling strategy is employed. This strategy can synthesize the semantic contribution of each word in the sequence, exhibiting stronger robustness. The formula for generating the fixed-dimensional semantic feature vector is as follows: ; in, This represents the generated fixed-dimensional semantic feature vector, whose dimension... Fixed at 768 dimensions; This indicates the effective sequence length of the input text entry after word segmentation, truncation, and padding. Indicates the first position in the input sequence The corresponding vector output by each word identifier in the last hidden layer of the model.

[0049] To address the retrieval latency issue in large-scale vector databases, a high-dimensional semantic metric space and vector index are constructed. An approximate nearest neighbor index based on Hierarchical Navigation Small World Graph (HNSW) is then developed. The maximum number of node connections in the HNSW algorithm is set. (For example ) and search depth during construction (For example ).

[0050] See attached document Figure 1 The semantic matching and data cleaning method based on high-dimensional spatial geometric characteristics provided by this invention includes the following steps: Candidate standard term sets are retrieved based on vector indexes; for each non-standard original entry to be processed in the historical feature library, an approximate nearest neighbor search is performed in the vector index constructed from the standard ontology library.

[0051] We perform precise Euclidean distance calculation based on L2 normalization. In the vector space generated by deep learning, the magnitude of a vector contains word frequency or context length information, while semantic similarity is mainly reflected in the consistency of vector direction. Therefore, directly calculating the Euclidean distance of the original vectors introduces non-semantic errors. To address this, we first process the feature vectors of non-standard original entries... and the first in the candidate standard term set standard vectors After L2 normalization, the result is projected onto the unit hypersphere. The formula for calculating the L2 norm is: ; in, It is a vector In the Components in each dimension For dimensions.

[0052] Then, the Euclidean distance between the normalized feature vectors and each candidate standard vector is calculated. The distance calculation formula is as follows: ; in, This indicates that the non-standard original entry is related to the first... The semantic distance between candidate standard terms is strictly limited to the range [0,2]. The feature vector representing a non-standard original entry; Indicates the first term in the set of standard terms. A standard vector.

[0053] Noise removal is performed based on a minimum distance criterion and an adaptive threshold strategy; the calculated distance set is then used. Filter out the minimum value and their corresponding standard terms . This is the most likely semantic match object at present.

[0054] To identify and remove "dirty" data with no corresponding relationships in historical data, a semantic validity threshold is introduced. Noise filtering is performed. The determination was performed using statistical methods: a subset of labeled samples was selected as the validation set, the distance distribution of correct matches was calculated, and the 95th percentile of this distribution was taken as the validation set. Typically, the range of this threshold is [range missing]. .

[0055] like The system determines that the semantic distance between the non-standard original entry and any term in the standard library is too large, and it belongs to semantic noise, so it directly performs the removal operation.

[0056] like The system determines that a valid match exists and proceeds to the next verification step.

[0057] Ambiguity verification based on distance ratios: In engineering terminology, there are many semantically similar but differently defined terms. To ensure high confidence in automatic mapping, a nearest neighbor ratio test is performed to obtain the second smallest value in the distance set. Calculate the distance ratio between the best match and the second-best match. : ; Set ambiguity threshold This threshold is typically set to a value of .

[0058] like This indicates that the distance between the best and second-best matches is not significant, indicating a risk of semantic ambiguity. The system marks this entry as pending manual review and does not include it in the automated rule base.

[0059] Only when and When both conditions are met, the mapping relationship is confirmed to be valid and unique, and the non-standard original entries are mapped to standard terms. And store it in a structured historical feature database.

[0060] See attached document Figure 1 The ontology library adaptive update mechanism based on clustering analysis and incremental learning provided by this invention includes the following steps: Density clustering analysis is performed on outlier data identified as noise or ambiguity; non-standard original entries that fail to be mapped are temporarily stored in a pending buffer. The system periodically executes a density-based clustering algorithm (DBSCAN) on the buffer data. This algorithm does not require a preset number of clusters. The cluster radius parameter is set. and minimum neighborhood number Cluster radius parameter Set as semantic validity threshold 0.8 to 1.0 times.

[0061] Minimum number of neighborhood points The value range is usually set to This means that the system only recognizes a cluster of new words as a valid candidate cluster when at least 5 to 10 unidentified entries are clustered within a very small semantic range; otherwise, they are considered outlier noise.

[0062] Calculate the centroid vector of the new word clusters and generate candidate standard terms; for each cluster generated by clustering that meets the density requirement... To represent this new concept mathematically, a representative vector needs to be extracted. The arithmetic mean (centroid) of all eigenvectors within the cluster is calculated. This centroid vector has the smallest sum of geometric distances to all points within the cluster, thus providing the best representativeness. The formula for calculating the centroid vector is as follows: ; in, Indicates the first The centroid vector of a new word cluster; Indicates the number of samples contained in the cluster; Indicates the sample index in the cluster; This represents a sample vector.

[0063] Calculate the sample vector in each cluster With the centroid vector The Euclidean distance is used to select the original text name corresponding to the sample with the closest distance as the recommended standard name for the new concept, which is then submitted for manual review.

[0064] Perform incremental index updates and ontology repository expansion; if approved, confirm the term as a new standard term. The system adds the new term to the standard engineering ontology repository and updates the corresponding centroid vector. Insert it into the existing vector index structure.

[0065] To ensure the continuity of online system services, the dynamic insertion logic of the HNSW index is executed. The specific process is as follows: starting from the top-level entry node of the index graph, a greedy search algorithm is used to find the distance to the centroid vector. The set of nearest nodes. As a new node, it is inserted into the graph and a bidirectional connection is established between it and the set of nearest nodes found. The number of connections is determined by the index, which determines the maximum number of connections a node can have. Control (e.g.) This allows for local updates to the graph structure while maintaining global small-world navigation properties.

[0066] The model based on concept drift detection is retrained periodically. As engineering standards evolve, the semantic context of terminology changes. The system continuously monitors the cumulative number of newly added standard terms. .

[0067] Set retraining threshold ratio (Usually 5%). When hour( (This is the total number of terms in the standard library), triggering the model fine-tuning process. New terms and their associated contextual corpora are added to the training set, incrementally training the word embedding model, and updating the model parameters to adapt to the new semantic distribution.

[0068] See attached document Figure 1 The method for converting a hierarchical building information model into graph database nodes according to the present invention includes the following steps: The system parses the Building Information Model (BIM) source files and extracts physical entity component objects; it reads BIM files conforming to the Industry Foundation Class (IFC) standard. The system traverses the data dictionary of the model files, filtering out all component instances that inherit from IfcProduct and have solid geometric representations, including walls, columns, beams, slabs, and pipe segments. Simultaneously, it extracts the globally unique identifier (GUID) for each entity component, which serves as the unique primary key for the graph database node.

[0069] Discretized mesh reconstruction is performed on the parametric geometric representation. Components in the BIM model are mathematically defined using parametric scans or constructed solid geometry. This implicit representation is difficult to directly use for distance calculations between components of arbitrary shapes. To unify the calculation basis, it needs to be converted into an explicit triangular mesh.

[0070] The system calls the computational geometry kernel component to read the shape representation of the component. Based on the preset linear deflection accuracy, the continuous parametric surface is divided into discrete triangular patches.

[0071] For each component node Its geometry is reconstructed into a set of vertices. The vertex set is represented as: ; in, Indicates the first The three-dimensional coordinates of each vertex The number of vertices.

[0072] Specifically, when generating vertex coordinates, the local placement properties of the components must be parsed, and the local coordinates relative to the parent component must be converted into global world coordinates through matrix transformation to ensure that all components are spatially analyzed in the same coordinate system.

[0073] Calculate the axis-aligned bounding box of the component entity and inject spatial properties, then traverse the generated vertex set in the world coordinate system. Find the extreme values ​​of the component along the three principal axes of X, Y, and Z. The axis-aligned bounding box is formed by the minimum coordinate point. and the maximum coordinate point The definition and calculation formula are as follows: ; ; in, , , Vertices The coordinate components.

[0074] Calculated and The bounding box is stored as a vector attribute in the graph nodes. The system also calculates the diagonal length of the bounding box. This length will subsequently be used to determine the baseline scale for adaptive tolerance.

[0075] The non-geometric attributes of the components are loaded based on the semantic alignment results. After completing the structured transformation of the geometric information, the semantic attributes of the components are loaded. The original BIM attribute set is read, and the semantic vector alignment method of multi-source heterogeneous historical data is called to map non-standard attribute values ​​to the standard terminology in the standard feature library.

[0076] The standardized attribute set (including component category, material name, and cross-sectional dimensions) is written into the node attributes of the graph database in the form of key-value pairs. At this point, each node in the graph database becomes a computational unit with precise geometric boundaries and standardized semantic definitions.

[0077] See attached document Figure 1 The present invention provides a two-level spatial adjacency detection method from coarse to fine screening, comprising the following steps: Construct a global spatial partition index and perform a coarse filtering based on AABB; in large architectural models containing a massive number of components, set the edge length of the mesh cells. This value is typically set to 1.5 to 2.0 times the longest side of the bounding box of the largest component in the model. The axis-aligned bounding box of each component is mapped to the spatial mesh. For components that cross mesh boundaries, multiple mapping is performed, meaning their IDs are simultaneously registered in the buckets of all mesh cells covered by their bounding boxes to prevent missed boundary detections. The system only performs subsequent detection on component pairs within the same mesh cell.

[0078] For the retrieved potential neighbor pairs Calculate the minimum Euclidean distance between their axis-aligned bounding boxes. This distance is geometrically equivalent to the magnitude of the projection gap vectors of the two cuboids along the three principal axes. The calculation formula is as follows: ; in, Indicates the directions of the coordinate axes: x, y, z; Represents component nodes The enclosing box in Minimum value on the axis; Represents component nodes The enclosing box in Maximum value on the axis; Its function is that when the projections of two bounding boxes on a certain axis overlap, the distance component of that axis is 0.

[0079] Setting coarse screening tolerance ,like If the two components are determined to be physically separated, they are directly excluded; if , and mark them as candidate adjacent pairs.

[0080] To address the issue of low efficiency in traversing complex meshes, a hierarchical bounding box tree is constructed for the triangular mesh of each component. This embodiment uses a top-down midpoint partitioning method to construct a binary tree. Each tree node contains an AABB bounding box and a pointer to its child nodes, while the leaf nodes contain the specific triangular facet indices.

[0081] Calculate two component meshes and Minimum geometric distance between When performing a recursive traversal of two tree nodes, the algorithm first checks if the bounding boxes of the two tree nodes intersect: if they do not intersect and the distance is greater than the currently known minimum distance, the search for that branch is stopped (pruning); if they intersect or are very close, the algorithm continues to the child nodes, up to the leaf node level. At the leaf node level, the shortest Euclidean distance between pairs of triangular faces is calculated. The mathematical definition of the minimum geometric distance is: ; in, and They belong to the component mesh respectively and Triangular facets; Represents two triangular faces and The shortest distance in three-dimensional space is calculated by solving the extremum problem between geometric primitives, specifically using the Separated Axis Theorem (SAT) or the Voronoi domain method.

[0082] An adaptive fuzzy tolerance is applied to generate probabilistic connection edges. After obtaining the precise distance, fuzzy logic is introduced to handle the inaccurate capture error commonly found in BIM modeling. A fine tolerance is then set. This tolerance value is adaptively obtained by querying the preset component type tolerance configuration table. For example: For prefabricated assembled components (such as PC wall panels), set ; For cast-in-place concrete components (such as beams and slabs), set .

[0083] like Then, connecting edges are created in the graph database. Simultaneously, based on the Gaussian radial basis function (RBF) principle, the connection probability weight of this edge is calculated. This model assumes that the modeling error follows a normal distribution, and that the closer the distance, the higher the confidence level of the actual connection. The probability weight formula is as follows: ; in, The minimum geometric distance; The bandwidth parameter for controlling the weight decay rate is set to [value]. .when hour, ;when hour, This reflects the low confidence characteristic of edge connectivity.

[0084] See attached document Figure 1 The method for establishing graph connections based on integrated geometric features and semantic logic of the present invention includes the following steps: Calculate the effective overlap ratio of the component contact surfaces by traversing the meshes of the two components that constitute potential adjacency pairs. and Determining contact solely based on distance is insufficient; it also requires verification using the normal direction of the facets. The system identifies a set of facets that simultaneously meet the following two conditions: The distance from the center point of the patch to the opposite mesh is less than the fine tolerance. ; normal vector of the face Normal vector of the nearest face to the other dot product (That is, the two sides are facing each other, and the included angle is greater than 143 degrees).

[0085] Calculate the total area of ​​the selected contact surface set. The formula for calculating the effective overlap ratio is as follows: ; in, Indicates the effective overlap ratio; The number of contact surfaces required to satisfy the distance and normal conditions; For the first The area of ​​each contact surface; and Representing component nodes respectively and component nodes The total surface area of ​​the geometric mesh; express and Minimum component area.

[0086] This section uses The technical principle behind using a normalized denominator is to solve the problem of determining the connection between components of vastly different sizes (such as a small socket installed on a large wall). If the area of ​​the larger component is used as the denominator, the ratio will approach 0; while using the area of ​​the smaller component (socket) as the denominator can accurately reflect the tightness of their contact (close to 1.0).

[0087] To eliminate false connections that do not conform to building mechanics or construction logic (such as direct connections between water supply pipes and power cable trays), the system introduces semantic compatibility constraints.

[0088] Pre-configure a two-dimensional semantic compatibility matrix based on national building standards and design specifications The rows and columns of this matrix correspond to the component categories in the standard ontology library. Matrix elements The value represents the first Class component node and the first The logical confidence level for establishing connections between class components takes values ​​from discrete sets. : (Prohibited connections): such as structural columns and soft decorative fabrics; (Weak connection / Potential connection): such as wall and baseboard; (Strong connections): such as beams and slabs, columns and foundations. For the edges to be processed... The system looks up the semantic weights in a table. .

[0089] The final connection probability is calculated based on a multi-factor fusion model; the connection probability weights are then applied based on geometric distance. Effective overlap ratio and semantic weight The fusion process is then performed. The final connection probability is calculated using the following formula: ; in, Let be the final connection probability of the graph edges; For semantic weights; For connection probability weights; The effective overlap ratio. This is a balancing coefficient used to adjust the contribution of distance weight and area weight; Saturation coefficient: due to the effective overlap ratio Typically small, it needs to be checked using the saturation coefficient. Magnify it and pass it through the hyperbolic tangent function Map its nonlinearity to Range to prevent numerical overflow.

[0090] Perform probabilistic edge instantiation and graph pruning. Based on the calculated... Filter potential connection edges. Set a global confidence threshold. This threshold is typically set to 0.4. If The system determines that the connection is invalid, performs a pruning operation, and does not generate edge data in the graph database.

[0091] like The system instantiates the connection edge relationships in the graph database, and at the same time... Stored as the weight attribute of the edge, the minimum geometric distance Stored as a distance attribute.

[0092] See attached document Figure 1 The present invention provides a method for verifying the legality of engineering nodes based on graph topology coding, comprising the following steps: Construct the local topological context vector of the current project node for each component node in the graph to be assigned a weight. A first-order neighborhood subgraph is extracted centered on this subgraph. A weighted aggregation strategy based on an attention mechanism is employed, utilizing the calculated final connection probabilities. As the credibility weight of the edge, the semantic features of neighboring nodes are weighted and synthesized. Local topological context vector. The calculation formula is as follows: ; in, Represents component nodes The set of direct neighbor nodes in the current probability graph; Represents component nodes With component nodes The probability of the final connection between the edges; Represents component nodes Standard semantic vectors; To prevent numerical stability constants with a denominator of zero.

[0093] To generate a joint feature vector that integrates its own semantics and environmental features, this embodiment uses weighted linear interpolation to fuse the two types of information while maintaining the vector dimension unchanged. The formula for calculating the joint feature vector is as follows: ; in, Represents component nodes The joint eigenvectors; For component nodes Its own standard semantic vector; It serves as a feature balancing factor, used to adjust the weight ratio between its own attributes and environmental attributes. This indicates the L2 normalization operation.

[0094] Retrieve baseline structure clusters from the historical ontology repository; generate joint feature vectors As the query object, a search is performed in the historical ontology constraint database, and the hierarchical small-world graph (HNSW) indexing algorithm is used to retrieve and combine feature vectors. Find the nearest reference cluster centroid in the feature space. Set the search quantity parameter. (Usually the value is 1, i.e., finding the nearest neighbor). Calculate the joint feature vector. centroid of the best-matching benchmark cluster Structural differences between The formula for calculating structural dissimilarity is as follows: ; in, It is the cluster centroid with the smallest Euclidean distance to the current node's features in the historical database; For dimensions. It quantifies the degree of deviation between the current node and its connections and historical experience. If A value approaching 0 indicates that the current node's connection pattern conforms to conventional engineering logic; if A significantly larger value suggests the presence of topological connection errors or unconventional special constructions.

[0095] See attached document Figure 1 This invention not only relies on manually set explicit rules, but also utilizes statistical methods to extract latent topological logic from historical engineering data to construct an implicit ontology constraint library. This process includes the following steps: The system undergoes standardized topological mapping of historical sample data. It accesses a historical engineering database and selects delivered BIM models that have undergone manual verification as the training sample set. Using the aforementioned topological reconstruction technology, each historical BIM model is converted into a structured topological map. During this process, semantic normalization is enforced: the nodes are... The attribute is replaced with a unique category identifier from the standard semantic library.

[0096] The system extracts topological triples and performs frequent itemset statistics. It traverses each edge in the graph and extracts topological triples in the form of source node category, connection relationship, and target node category. The triple pattern is defined as follows: ,in and These represent the standard component categories of the nodes at both ends of the connecting edge. It indicates an established physical connection.

[0097] The system calculates the absolute observation frequency of each triplet pattern throughout the entire historical sample set. and source node category Total frequency of occurrence in all samples And set a minimum support threshold. Only retain those that meet the requirements. The triplet pattern is used as a candidate frequent itemset.

[0098] To calculate the conditional connectivity probability and mutual information strength, the mutual information (PMI) theory is introduced to eliminate the statistical bias of the component's own popularity on the association strength. First, the conditional connectivity probability is calculated. , indicating that when there are components At that time, its connecting components are The probability is calculated using the following formula: ; in, This indicates the total number of component categories in the ontology library; is the Laplace smoothing coefficient, used to prevent zero probability problems caused by sample sparsity.

[0099] Calculate normalized mutual information strength The calculation formula is as follows: ; in, It is a component With components Joint probability of co-occurrence; and They are components With components Independent marginal probabilities; It is the total number of all connecting edges in the sample set.

[0100] To construct implicit constraint matrices and anomaly detection thresholds, and to automatically determine the decision boundaries, the system performs analysis on all candidate triples. One-dimensional K-Means clustering is performed on the values. The following threshold is determined based on the cluster centers: Strong correlation threshold : .when When a strong implicit rule is established that a connection must be made; mutual exclusion threshold. : .when When this occurs, it is established as an implicit rule that prohibits connections.

[0101] Finally, the high-confidence triples and their corresponding conditional connection probabilities will be extracted. The metadata layer stored in the graph database forms an implicit ontology constraint library.

[0102] See attached document Figure 1 This invention, based on the generation of geometric connection probabilities, introduces a Bayesian inference model and combines prior knowledge of historical ontology to perform secondary weighting of connection relationships, eliminating pseudo-connections that are geometrically overlapping but logically invalid, or completing implicit connections that are logically valid but geometrically separated. The invention includes the following steps: Obtain the logical prior confidence of the edge to be determined; for any probabilistic edge to be determined, the system first determines the confidence level based on the component nodes. and component nodes Based on the standard semantic category, retrieve the rule base constructed in the preceding steps, and synthesize the logical prior confidence of the connection relationship. The retrieval process follows a strategy that prioritizes mandatory norms over historical experience: first, it queries the two-dimensional semantic compatibility matrix. If the explicit rule is not defined (i.e., a neutral value that is neither 0 nor 1), then the corresponding conditional join probability is queried from the implicit ontology constraint library. The logic for calculating logical prior confidence is as follows: ; in, and These are component nodes. and component nodes Component categories; For a set of implicitly constrained triples; This represents conditional statements and logical operations; Indicates existence; This represents the maximum entropy state in the absence of any explicit normative constraints and historical data support.

[0103] Perform posterior probability calculation based on Bayesian evidence fusion; calculate the final connectivity probability obtained through geometric computation. Treating it as geometric observational evidence, and applying logical prior confidence. Treating these as semantic prior evidence, the posterior truth probability of the connectivity relationship is calculated using the Bayesian evidence fusion formula. The formula for calculating the posterior true probability is as follows: ; in, This represents the final probability of rights confirmation after combining the current physical situation with historical logic. The mathematical principle of this formula lies in converting two probability values ​​into logarithmic probabilities, linearly superimposing them, and then converting them back into probability space. Its physical effect is as follows: Bidirectional enhancement: when and At that time, the result It will clearly approach 1; Conflict suppression: When one of the two is high and the other is low, the result will tend to 0 or be suppressed, thus automatically filtering out the noise generated by accidental collisions.

[0104] The system calculates a cognitive conflict index and identifies anomalous connections; in practical engineering models, obvious inconsistencies between geometric representations and logical rules indicate potential modeling errors. The system also calculates the cognitive conflict index between observed probabilities and prior probabilities. The formula for calculating the conflict index is as follows: ; Set conflict threshold This threshold is typically set to 0.7. When When this happens, the system triggers the exception marking process: Hard collision error: If and For example, if a water pipe passes through an electrical cable tray, the edge is marked with the attribute type:CLASH in the graph and output as a collision check result in subsequent applications.

[0105] Logical gap: If and For example, if there is a tiny gap between the structural column and the floor slab, it is determined to be an omission in the modeling. In this case, the system does not create a physical connection edge, but instead generates a virtual edge with the attribute type:LOGICAL_SUGGESTION to prompt the designer to complete it.

[0106] The finalization of the topological graph and the instantiation of edge types. Based on... Based on the conflict detection results, the graph edges are finally instantiated. A weighting threshold is then set. A value of 0.8 is recommended. and At this point, the edge is officially recognized as a valid topological connection. Based on the component category attributes at both ends of the connection, the system consults a pre-defined domain classification set to further refine the edge's type label. : ; in: This represents a set of structural components, including: walls, columns, beams, slabs, and foundations; This refers to a set of electromechanical components, including: pipe sections, pipe fittings, and terminal equipment; Indicates other.

[0107] Ultimately, the edge data structure stored in the graph database includes: source node ID, target node ID, and connection type label. Posterior true probability And the original geometric distance. Through this step, the original geometric BIM model is reconstructed into a computable knowledge graph with semantic reasoning capabilities.

[0108] A global topological importance index is constructed. To quantify the topological importance of each component in the overall structural system, the system runs the PageRank algorithm on the entire graph. For any component node... Its importance score The iterative calculation formula is: ; in, This represents the total number of nodes in the graph; This is the damping coefficient (usually set to 0.85); To point to component nodes The set of nodes; For component nodes The number of out-degrees.

[0109] Calculated It will be stored as an inherent attribute of the node and used for entity disambiguation in subsequent natural language parsing and weight allocation in consistency monitoring.

[0110] See attached document Figure 1This invention provides a method for converting engineering requirements in natural language form into machine-executable logical instructions, comprising the following steps: Semantic parsing and intent slot filling for natural language requirements: The system receives engineering management instructions or contract terms from the user and performs semantic role-based intent parsing. A pre-trained language model (such as BERT) is used as the encoder to extract features from the input text. To accommodate the rigorous logic of the engineering domain, the intent logic frame contains four core slots for structured execution of business logic: Triggers: Identify time-related conditions or state transition events (e.g., "when...acceptance is passed"); Subjects: Identify the target engineering object of the operation (e.g., "first floor load-bearing column"); Actions: Identify specific instructions to be executed (e.g., "pay progress payment", "start the next process"); Constraints: Identify numerical or state-related limiting indicators (e.g., "strength grade > C30").

[0111] The system maps the input text to the above slots through a Conditional Random Field (CRF) or a sequence labeling layer to generate a structured set of intent instructions.

[0112] Graph-based entity disambiguation and context constraint injection are used to inject entities from the intent frame. The system accurately maps to the unique BIM component ID in the map and automatically completes the implicit engineering constraints, performing entity disambiguation and constraint injection.

[0113] Intent frame Text semantic vectors Joint eigenvectors of all nodes in the graph Similarity matching is performed. To improve the accuracy of core component identification, structural centrality is introduced as a weighting factor. The matching score calculation formula for graph enhancement is as follows: ; in, This represents a candidate building node in the graph; This is the joint feature vector of the nodes; This is the PageRank value of the node in the final weighted graph. This value is calculated based on the adjacency matrix of the graph and reflects the importance of the node in the topology. The centrality weighting coefficient.

[0114] Set matching threshold The system selects all. Node set As the object of operation.

[0115] Then, the system performs a context constraint injection operation. For Each build node in In the graph, perform a reverse search along the incoming edges where the edge relationship is a support relationship to find the set of all direct predecessor nodes. .like If not empty, the system will automatically check. The logical condition that all nodes are in the COMPLETED state is appended to the intent frame. In the slot.

[0116] The system constructs and logically compiles an Abstract Syntax Tree (AST); the enhanced intent frame is converted into a programmable AST. First, the system extracts all nodes involved in the graph and their dependencies, constructing a local dependency subgraph. The Kahn algorithm or depth-first search is applied to topologically sort this subgraph, thus transforming the mesh-like graph structure into a linear execution sequence.

[0117] An AST is constructed based on a topological sequence, with the root node representing the contract definition and child nodes containing state variable declarations, event listeners, and logic functions. The logic compilation process involves constructing a state transition function. Its mathematical expression is as follows: ; in, This represents the global state vector of the current system. To trigger the event; This is the set of constraints automatically extracted from the topological relationships of the graph; A set of business constraints defined by the user; This represents the logical AND operation, meaning that all constraints must be satisfied simultaneously.

[0118] Code instantiation and formal verification. The generated AST is converted into code in the target programming language. Before code deployment, model checking techniques are introduced for formal verification. Linear temporal logic formulas are defined. To describe the security attributes of the system, for example, the attributes for structural security are described as follows: ; in, It is a sequential logic operator, meaning it holds true throughout all future time paths; It is a logical NOT operator; For logical implication operators; Represents component nodes It is already built; Indicates the component node Execute the payment action.

[0119] The system uses an SMT solver (such as Z3) or a model detector to traverse the state space and verify whether the generated code satisfies the above formula. The system will only output the final executable code package if the verification passes.

[0120] See attached document Figure 1 This invention includes runtime dynamic monitoring and self-evolution capabilities based on historical data, forming a dual closed loop of real-time execution correction and long-term ontology evolution, specifically including the following steps: Runtime physical-logical consistency monitoring: To convert unstructured point cloud data into structured component states, the system executes the following state determination logic: For each component node... Obtain its 3D geometric bounding box in the BIM model. Calculate the real-time point cloud density within this bounding box space. .like Determine the physical state If it already exists; otherwise, it is considered not started.

[0121] System defines consistency measurement function Used to quantify the current moment The execution deviation. The calculation formula is as follows: ; in, This indicates the total number of component nodes currently under active monitoring. The expected virtual state of the component as recorded in the smart contract or control script; This is an indicator function that takes the value 1 when the physical state and the logical state are inconsistent, and 0 otherwise; The importance weight of the component is directly reused from the PageRank value. .

[0122] when Exceeding the preset warning threshold When the threshold (usually set to 0.1, allowing non-critical deviations within 10%) is reached, the system determines that a virtual-real synchronization anomaly has occurred, automatically suspends the current execution process, and generates an anomaly report.

[0123] Interactive rights confirmation and feedback capture with manual intervention; for triggered anomalies or fuzzy connections marked as logical suggestions, the system provides an interactive interface for engineers to make decisions. The system captures user actions and converts them into standardized feedback signal scalars. The feedback signal is defined as follows: Positive confirmation: The user has adopted the system-generated suggested edge or confirmed the current execution status. .

[0124] Negative correction: The user deleted system-generated connection edges or forcibly modified the component state. .

[0125] Parameter fine-tuning: The user manually adjusted the connection properties. Based on the extent of the modification It can take values ​​between these ranges.

[0126] In a short-cycle closed loop, the system utilizes feedback signals to address the explicit semantic compatibility matrix. Online updates are performed. For areas where the system's prediction is less certain (lower probability), the weight adjustment should be larger if confirmed by the user (positive feedback). A stochastic gradient descent strategy is used to adjust the semantic weights. The adjustment formula is as follows: ; in, express Time component category and Semantic compatibility between them; The learning rate is used to control the speed at which the system accepts new knowledge. For user feedback values; This is a truncation function.

[0127] The system employs a long-term closed-loop process, periodically performing a full analysis of the implicit ontology constraint library. It also tracks the number of times each rule in the implicit rule base is validated in actual projects. and number of rejections The system calculates the global confidence index of the rule. The implicit rule is upgraded to an explicit ontology specification when the following migration conditions are met: ; in, The frequency threshold is set to 1000 times or more. The accuracy threshold is set to 0.95.

[0128] Once the rule migration is complete, in subsequent steps, the matrix elements corresponding to that rule will be locked to 1.0 (strong connection) or 0.0 (no connection) and marked with the STATIC attribute, no longer participating in short-cycle dynamic floating updates. Through this step, the system achieves automatic evolution from data-driven statistical regularities to expert-approved deterministic specifications.

[0129] To further verify the effectiveness of this invention in practical applications, this embodiment constructs an application scenario for intelligent aggregation of engineering quantities in a large commercial complex, and combines it with... Figures 3-6 Please provide a detailed explanation.

[0130] Specific application example: Intelligent aggregation of engineering quantities for a large commercial complex.

[0131] Scenario Construction: This embodiment selects a large commercial complex project as the test and verification scenario. The project model has a complex structure and contains a large amount of heterogeneous component data. To verify the effectiveness of the present invention in practical engineering applications, we conducted statistical and visual analysis of the core indicators during system operation. The specific results are as follows.

[0132] Experimental verification and effect comparison: Comparison of semantic recognition performance of multi-source heterogeneous data: In order to verify the accuracy of the system in recognizing engineering terms, the names of non-standard components in the project were selected as the test set, and three methods were used for testing: "traditional keywords / regular expressions", "general BERT model" and "embedded model of this invention".

[0133] Please refer to Figure 3 The figure shows a comparison of the accuracy and recall of different semantic recognition methods.

[0134] Detailed Data Analysis: Traditional keyword / regular expression methods: Due to their inability to handle fuzzy naming, they perform the worst, with an accuracy of only 65.2% and a recall of only 45%.

[0135] The general BERT model has improved performance, with an accuracy of 88.5% and a recall of 85.3%.

[0136] The embedded model of this invention achieves optimal performance thanks to fine-tuning for engineering applications. The data in the figure shows an accuracy of 98.5% and a recall rate of 97.2%.

[0137] Conclusion: Compared with traditional regular expression matching methods, the present invention improves the accuracy by more than 30% and doubles the recall rate in handling semantic alignment tasks, and can solve the problem of dirty data cleaning.

[0138] Connection Probability Decay Analysis Based on Geometric-Logic Fusion: To address the common modeling gap (floating error) problem in BIM models, this system introduces a probabilistic connection algorithm based on semantic weights. To verify the effectiveness of its logical weighting, we tested the trend of connection probability with geometric gap distance under different semantic overlap rates.

[0139] Please refer to Figure 4 The figure shows the connection probability decay curves under three typical overlap rates.

[0140] Explanation of the curves in the figure: Solid line (high overlap) ): Represents components with strong logical connections (such as structural columns and floor slabs). Data shows that even when the geometric gap increases from 0mm to 50mm, the connection probability remains above 0.9 (the curve is gentle). This indicates that the system has extremely strong robustness, can tolerate large modeling errors, and will not miss deduction relationships due to geometric separation.

[0141] Dotted lines (low overlap) ): Represents components with extremely weak logical connections (such as air ducts and structural beams). Data shows that when there are tiny gaps (>5mm), the connection probability drops rapidly to below 0.1.

[0142] Conclusion: As can be seen from the comparison of two-dimensional curves, the present invention can use semantic logic to retrieve the connected components, while using a distance threshold to ignore irrelevant components, thus achieving dual verification of geometry and logic.

[0143] Comprehensive comparison of efficiency and accuracy in quantity surveying: This embodiment compares the quantity surveying performance under three different modes: human expert group, mainstream BIM software and the system of this invention.

[0144] Please refer to Figure 5 The figure shows a dual-axis comparison of calculation time and overall error rate.

[0145] Detailed Data Analysis: Calculation time (left axis, bar chart): The human expert group took the longest, reaching 120 hours; mainstream BIM software took 8 hours; while the system of this invention only took 0.75 hours.

[0146] Overall error rate (right axis, line graph): The error rate of the human expert group is 4.5%; the error rate of mainstream BIM software soars to 12.0% because it cannot automatically handle non-standard modeling; the system of this invention controls the error rate at 0.8%.

[0147] Conclusion: The system of the present invention improves efficiency by more than 100 times while achieving the lowest error rate, which is significantly better than the existing traditional operation mode.

[0148] System self-evolutionary convergence performance: To verify the system's self-learning ability, the changes in system metrics were recorded during the four weeks of the project (Week 1-Week 4).

[0149] Please refer to Figure 6 The figure shows the convergence curve of system confidence as the project iterates.

[0150] Detailed Data Analysis: User negative feedback rate (left axis, solid line): It dropped rapidly from 15.2% in Week 1. It dropped to 5.5% in Week 2, 1.8% in Week 3, and finally to 0.4% in Week 4.

[0151] Implicit rule migration count (right axis, dashed line): As the feedback rate decreases, the number of rules automatically learned and migrated by the system steadily increases. It grows from 0 rules in Week 1 to 45 rules in Week 2, 92 rules in Week 3, and finally reaches a total of 120 rules in Week 4.

[0152] Conclusion: The data curves demonstrate that the system becomes more accurate with use and can achieve rapid convergence of model performance and knowledge solidification within 4 cycles through a closed loop of user feedback.

Claims

1. A method for summarizing engineering quantities based on large-scale models and big data analysis intelligence, characterized in that, Includes the following steps: Construct a semantic vector mapping space for historical data, and generate a standardized historical feature library based on semantic distance filtering; The building information model is parsed, the semantic vector mapping space is called to map the component attributes to standard terms, and the components are instantiated as attribute graph nodes. An engineering graph containing probabilistic connection attributes and connection probability weights is generated based on the grid geometric distance. In the standardized historical feature library, a benchmark project cluster is retrieved, ontology constraints are extracted, and logical weighting of the probabilistic connection attributes in the engineering graph is performed in combination with the statistically derived connection logic constraints, converting probabilistic edges into deterministic connection edges. By combining the ontology constraints and attribute default values, a prompting context is constructed, and the large language model is invoked to convert engineering change instructions into structured graph query scripts; The structured graph query script is executed to filter target component nodes, and explicit and implicit quantities of work are aggregated based on topology propagation. Perform a dual closed-loop verification based on statistical distribution and model quality feedback, and output the final summary result of the engineering quantity.

2. The method for summarizing engineering quantities based on large models and big data analysis intelligence according to claim 1, characterized in that, The construction of the semantic vector mapping space for historical data, and the generation of a standardized historical feature library based on semantic distance, specifically includes: Read the classification items from the historical database and call the domain embedding model to map the classification items and standard terms into high-dimensional feature vectors; Calculate the minimum Euclidean distance between the high-dimensional feature vector corresponding to the classification item and the standard vector corresponding to the standard term; Traverse the standard engineering ontology library to find the standard term corresponding to the minimum Euclidean distance, and determine whether the minimum Euclidean distance is less than or equal to the semantic validity threshold; If yes, then establish a mapping relationship between the classification items and standard terms and store it in the historical feature library; if not, then remove it as noise data.

3. The method for summarizing engineering quantities based on large models and big data analysis intelligence according to claim 1, characterized in that, The process of parsing the building information model, calling the semantic vector mapping space to map component attributes to standard terms, instantiating components as attribute graph nodes, and generating an engineering graph containing probabilistic connectivity attributes and connectivity probability weights based on grid geometric distance specifically includes: The standardized component attributes are assigned to the corresponding physical components, and the physical components are instantiated as component nodes in the attribute graph. The distance between the feature vector corresponding to the non-geometric attribute and the standard vector in the semantic vector mapping space is calculated to determine the standardized component attributes. Instantiate physical components as component nodes in the property graph, calculate the axis-aligned bounding box of the component entity, and calculate the minimum geometric distance between the entity meshes for the component node pairs after coarse screening by the axis-aligned bounding box. Determine whether the minimum geometric distance is less than the fine tolerance. If so, generate probabilistic edges between component nodes and calculate the connection probability weights of the probabilistic edges based on the minimum geometric distance.

4. The method for summarizing engineering quantities based on large-scale models and big data analysis intelligence according to claim 1, characterized in that, Retrieving benchmark item clusters from the standardized historical feature database and extracting ontology constraints specifically includes: Extract the global feature parameters of the current project and map them to the feature vector of the current project; Calculate the cosine similarity between the current project feature vector and the historical project feature vectors recorded in the historical feature database; Historical items with a cosine similarity higher than the similarity screening threshold are selected to form a benchmark item cluster, and the component connection frequency is extracted from the benchmark item cluster as an ontology constraint.

5. The method for summarizing engineering quantities based on large models and big data analysis intelligence according to claim 4, characterized in that, The step of logically determining the weights of probabilistic connection attributes in the engineering graph by combining statistically derived connection logic constraints and converting probabilistic edges into definite connection edges specifically includes: Obtain the connection probability weights of the probabilistic edges; Read the connection frequency of physical connections between different types of components in the benchmark project cluster, calculate the product of the connection probability weight and the connection frequency, and obtain the value of the determination criterion function; Determine whether the value of the right confirmation criterion function is greater than the right confirmation threshold; if so, update the data type of the probability edge to a determined connection edge and mark it as an automatically repaired edge.

6. The method for summarizing engineering quantities based on large models and big data analysis intelligence according to claim 1, characterized in that, The step of constructing a prompt context by combining the ontology constraints and attribute default values, and calling the large language model to convert engineering change instructions into a structured graph query script, specifically includes: Receive engineering change instruction text in natural language form, identify the target component category pointed to by the engineering change instruction text, and extract high-frequency attributes corresponding to the target component category from the benchmark project cluster of the historical feature library as the default value of the attribute; Construct a prompt context that includes image data, the default values ​​of the attributes, and the text of the engineering change instruction; The large language model is invoked to generate a structured graph query script containing matching path clauses, filtering condition clauses, and attribute update clauses based on the given prompt context.

7. The method for summarizing engineering quantities based on large models and big data analysis intelligence according to claim 6, characterized in that, The default values ​​of the attributes in the structured graph query script are used to complete the parameters not mentioned in the engineering change instruction text in the attribute update clause, ensuring the completeness of the graph database query.

8. The method for summarizing engineering quantities based on large-scale models and big data analysis intelligence according to claim 1, characterized in that, The process of executing the structured graph query script to filter target component nodes and aggregating explicit and implicit engineering quantities based on topology propagation specifically includes: Run the structured graph query script to filter the target component node set and aggregate the explicit geometric attributes of each node in the target component node set; By traversing the determined connecting edges and performing Boolean operations, the geometric contact surface area integral or overlapping volume integral of the target component node and the adjacent node entity are calculated to obtain the implicit engineering quantity. The total engineering quantity is obtained by adding the explicit geometric properties and the implicit engineering quantities.

9. The method for summarizing engineering quantities based on large models and big data analysis intelligence according to claim 4, characterized in that, The execution of the dual closed-loop verification based on statistical distribution and model quality feedback specifically includes the following steps: verifying the calculation results. Calculate the unit quantity index of the current project, and calculate the standard score of the unit quantity index based on the mean and standard deviation of the benchmark project cluster index. Determine whether the absolute value of the standard score exceeds the preset standard deviation range; if so, the result is deemed abnormal.

10. The method for summarizing engineering quantities based on large model and big data analysis intelligence according to claim 5, characterized in that, The execution of the dual closed-loop verification based on statistical distribution and model quality feedback also includes a step to verify model quality: The proportion of the number of automatically repaired edges to the total number of connected edges is counted, and it is determined whether the proportion is greater than the set quality threshold. If so, a model quality warning will be output, prompting you to check the modeling accuracy of the building information model.

Images