A BIM-driven engineering cost prediction method and system

Abstract

This invention relates to the field of data processing technology, specifically to a BIM-driven method and system for predicting engineering costs. The method includes the following steps: acquiring text fields and geometric material information of BIM components; extracting multi-dimensional semantic feature datasets of components; extracting the semantic features of the bill of quantities structure and quota codes to generate multi-dimensional semantic feature sets for items; generating a mapping relationship table based on cosine similarity matching and anomaly labeling; correcting the mapping relationship by clustering functional geometric materials; and inputting the cost model to output the engineering cost prediction result by combining quantity and price. In this invention, the quantification of semantic features between component attributes and bill of quantities items enables computationally driven association establishment, reducing reliance on experience. Combining geometric proportions and material codes forms multi-dimensional constraints, distinguishing semantically similar but attribute-different matching relationships. Deviation quantification achieves mapping correction, reducing the frequency of manual verification, enhancing computational consistency and prediction stability, and improving the reliability of the results.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a BIM-driven method and system for predicting engineering costs. Background Technology

[0002] The field of data processing technology mainly involves the collection, organization, storage, calculation, and analysis of various types of engineering and business data. Its core aspects include the structured representation of multi-source data, the correlation mapping between data, the data calculation process based on established rules or models, and the output format of results. In engineering construction-related scenarios, data processing revolves around bill of quantities, component attribute information, price information, and historical project data. By uniformly encoding, associating, matching, and calculating the data, it provides basic support for subsequent analysis and decision-making.

[0003] Traditional engineering cost forecasting methods involve cost estimators manually counting the number of engineering components based on two-dimensional design drawings and written descriptions during the construction process. They then apply the list of items one by one according to the quota rules, calculating the quantities of the project manually or with the help of spreadsheets based on the component type, size, material type, and construction method. The calculation results are then matched with the unit prices of labor, material, and machinery costs in the cost information database. Finally, the overall cost forecast is completed through manual summarization and adjustment. In this process, engineering information and cost data are mostly in the form of scattered tables or documents, and the correspondence between component information and price data depends on manual establishment. The forecasting process is mainly based on historical project data and empirical rules.

[0004] Existing technologies rely on manual statistics and adjustments, resulting in scattered component information and inconsistent attribute descriptions. This lack of standardized guidelines for matching items with components easily leads to mismatches. The calculation process depends on empirical rules, failing to effectively consider component geometric features and material differences, resulting in unstable prediction results. Data updates and verification involve repetitive work and are inefficient. Furthermore, the technology cannot automatically identify, accurately match, or correct differences between similar components and construction methods, causing cost deviations to accumulate and leading to poor prediction accuracy. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a BIM-driven engineering cost prediction method and system.

[0006] To achieve the above objectives, this invention employs a BIM-driven engineering cost prediction method, comprising the following steps: S1: Obtain the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, use BERT to perform text encoding to extract multi-dimensional feature vectors of component text, and generate a multi-dimensional semantic feature dataset of components. S2: Based on the multidimensional semantic feature dataset of the components, extract the structural semantic phrases and key quota codes of the bill of quantities items, use a word vector model to calculate the feature vectors of the items, and construct a multidimensional semantic feature set of the items; S3: Call the multidimensional semantic feature set of the entries and the multidimensional semantic feature dataset of the components, calculate the matching degree based on cosine similarity, filter the component entry pairs whose matching degree exceeds the matching threshold, mark abnormal matching pairs, and generate a mapping relationship table; S4: Call the mapping relationship table, calculate the cluster center coordinates according to the component functional classification number, geometric scale and material code, calculate the entry vector center coordinates, quantify the matching deviation, perform offset vector analysis and adjust the mapping, and generate a correction mapping relationship; S5: Based on the corrected mapping relationship, input the mapping data into the cost calculation model, combine the number of components and material prices, perform a cost summary calculation, and generate the project cost prediction result.

[0007] As a further aspect of the present invention, the component multidimensional semantic feature dataset includes component functional semantic categories, construction attribute feature labels, material semantic attribution identifiers, and geometric morphological feature parameters; the item multidimensional semantic feature set includes list item semantic types, quota code semantic orientations, and pricing attribute feature identifiers; the mapping relationship table includes component item correspondence, matching validity identifiers, and similarity level distinction results; the corrected mapping relationship includes deviation correction coefficients, consistency test results, and optimized correspondence indexes; and the engineering cost prediction results include sub-item cost results, overall engineering cost total value, and cost composition ratio indicators.

[0008] As a further aspect of the present invention, the steps for obtaining the component multidimensional semantic feature dataset are as follows: S111: Obtain the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, perform word segmentation, perform symbol mapping on the name field, map the material code to a value and normalize it with the component geometric feature ratio, and generate a component text structure code set. S112: Based on the component text structure encoding set, perform word frequency statistics and position calculation on the word encoding sequence, map the word to semantic vector, weight the structural parameter vector, calculate the cosine similarity, filter samples with a consistency benchmark value, align the text semantic vector and the structural parameter vector and concatenate them to generate a component multidimensional semantic vector matrix. S113: Call the component multidimensional semantic vector matrix, perform dimensional normalization and numerical normalization on the component vectors, perform index mapping based on the component identifier, arrange and verify the vector set, and generate a component multidimensional semantic feature dataset.

[0009] As a further aspect of the present invention, the steps for obtaining the multidimensional semantic feature set of the entries are as follows: S211: Based on the multidimensional semantic feature dataset of the component, call the term encoding sequence and structural parameters in the component text structure encoding set, perform encoding index mapping and positional relationship calculation, and perform vector mapping according to the geometric feature ratio and material code normalization result to generate an item structure semantic vector set; S212: Based on the item structure semantic vector set, according to the key quota coding semantic identifier, perform multi-dimensional semantic alignment calculation on the item vector, allocate weights according to the similarity distribution between dimensions, perform weight consistency verification and reconstruct the vector structure to generate an item semantic feature vector set. S213: Based on the set of semantic feature vectors of the items, perform pairwise similarity calculation on the item vectors, perform clustering and sequential arrangement according to the similarity results, perform structural consistency verification, and generate a multidimensional semantic feature set of the items.

[0010] As a further aspect of the present invention, the step of obtaining the mapping relationship table is as follows: S311: Call the item multidimensional semantic feature set and component multidimensional semantic feature dataset, extract the dimension components of the item vector and component vector, perform consistency check on the number of dimensions of the two types of vectors, perform dimension alignment for vectors with inconsistent number of dimensions, and generate a vector dimension alignment result set. S312: Based on the vector dimension alignment result set, extract the dimension component values ​​of the entry vector and component vector, and perform pairwise cosine similarity calculation on the entry vector and component vector using the formula: ; Calculate the similarity index of item components, filter the pairings of item components that exceed the matching threshold, and generate a candidate matching pairing set; in, Representing the The first entry and the first Similarity index between components Representing the The entry vector at the th entry vector in the th th The numerical values ​​of the components of the dimension, Representing the The component vector in the th... The numerical values ​​of the components of the dimension, Represents the total number of vector dimensions. Representing the The first entry and the first Semantic offset between components Represents the semantic offset attenuation coefficient; S313: Based on the candidate matching pairing set, extract the semantic offset values ​​of the entry vector and component vector in the pairing combination, compare the semantic offset values ​​with the offset threshold, perform an anomaly marking operation on the pairing combination whose semantic offset values ​​exceed the offset threshold, and establish a mapping relationship table.

[0011] As a further aspect of the present invention, the matching threshold is obtained by performing interval statistics on the cosine similarity values ​​between the entry vector and the component vector in the original sample data. The interval statistics are determined based on the similarity distribution range under a preset number of samples. The offset threshold is obtained by performing discrete statistics on the semantic offset values ​​of the entry vector and component vector in the original sample data. The discrete statistics are determined based on the mean and fluctuation range of the semantic offset values, and the value corresponding to the upper limit of the fluctuation range is used as the offset threshold.

[0012] As a further aspect of the present invention, the step of obtaining the correction mapping relationship is as follows: S411: Call the mapping table, extract the component functional classification number, geometric ratio value and material code in the paired combination, perform mean calculation on the geometric ratio value under the same functional classification number, perform encoding mapping conversion on the material code and calculate the weighted mean, and obtain the cluster center coordinate set; S412: Based on the cluster center coordinate set and mapping relationship table, extract the numerical values ​​of the dimension components of the item vector, calculate the category center vector for the dimension components of the item vector under the same functional classification number, calculate the category item center deviation, and perform a difference operation with the cluster center coordinate values ​​to generate a set of matching deviation quantification values. S413: Based on the set of matching deviation quantization values, extract the deviation of the category entry center and the coordinate values ​​of the cluster center, perform a weighted average calculation on the numerical differences, obtain the offset vector component values, perform a reverse correction operation on the entry vector coordinates, and establish a correction mapping relationship.

[0013] As a further aspect of the present invention, the steps for obtaining the engineering cost prediction results are as follows: S511: Based on the correction mapping relationship, the mapping data is matched item by item with the component coding rule, the component quantity parameter associated with each mapping item is retrieved, and a weighted operation is performed on the component quantity parameter and the correction coefficient in the mapping data to generate a component correction quantity set; S512: Based on the set of component correction quantities, collect the material price data of the corresponding component type, perform a weighted calculation on the material price values ​​and component correction quantities item by item, and perform numerical aggregation on the weighted results of similar components to obtain the total component cost value. S513: For the total component cost value, input the component cost value into the summary rules set by the cost calculation model, perform sequential integration and pricing caliber alignment processing on the cost value sequence, and generate the project cost prediction result.

[0014] The BIM-driven engineering cost prediction system is used to execute the aforementioned BIM-driven engineering cost prediction method, and the system includes: The component feature extraction module obtains the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, and uses BERT to perform text encoding to extract multi-dimensional feature vectors of component text, generating a multi-dimensional semantic feature dataset of components. The item feature extraction module extracts structural semantic phrases and key quota codes of the bill of quantities items based on the multidimensional semantic feature dataset of the components, calculates the item feature vectors using a word vector model, and constructs a multidimensional semantic feature set of items. The matching degree calculation module calls the multidimensional semantic feature set of the entries and the multidimensional semantic feature dataset of the components, calculates the matching degree based on cosine similarity, filters the component entry pairs whose matching degree exceeds the matching threshold, marks abnormal matching pairs, and generates a mapping relationship table. The mapping relationship correction module calls the mapping relationship table, calculates the cluster center coordinates based on the component functional classification number, geometric scale and material code, calculates the entry vector center coordinates, quantifies the matching deviation, performs offset vector analysis and adjusts the mapping, and generates a corrected mapping relationship. The cost prediction module, based on the corrected mapping relationship, inputs the mapping data into the cost calculation model, combines the number of components and material prices, performs a cost summary calculation, and generates the project cost prediction result.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by quantifying the semantic features of component attributes and bill of quantities entries, the relationship between components and entries can be established through calculation rather than experience, ensuring the accuracy of information matching. Multidimensional constraints using geometric proportions and material codes improve matching accuracy and correct for matching problems involving semantically similar but different attributes. Deviation quantification control reduces repeated manual verification, improving calculation consistency and stability. This solution addresses the problems of existing technologies in matching accuracy, data consistency, and prediction stability, enhancing the reliability of prediction results. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a flowchart illustrating the process of obtaining the multidimensional semantic feature dataset of components in this invention. Figure 3This is a flowchart illustrating the process of obtaining the multidimensional semantic feature set of entries in this invention. Figure 4 This is a flowchart illustrating the process of obtaining the mapping relationship table in this invention; Figure 5 This is a flowchart illustrating the process of obtaining the corrected mapping relationship in this invention; Figure 6 This is a flowchart illustrating the process of obtaining the engineering cost prediction results in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] Example 1: Please refer to Figure 1 This invention provides a technical solution: a BIM-driven method for predicting engineering costs, comprising the following steps: S1: Obtain the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, use BERT to perform text encoding to extract multi-dimensional feature vectors of component text, and generate a multi-dimensional semantic feature dataset of components. S2: Based on the component multidimensional semantic feature dataset, extract the structural semantic phrases and key quota codes of the bill of quantities items, use the word vector model to calculate the item feature vectors, and construct the item multidimensional semantic feature set; S3: Call the multidimensional semantic feature set of entries and the multidimensional semantic feature dataset of components, calculate the matching degree based on cosine similarity, filter component entry pairs whose matching degree exceeds the matching threshold, mark abnormal matching pairs, and generate a mapping relationship table; S4: Call the mapping relationship table, calculate the cluster center coordinates based on the component functional classification number, geometric scale and material code, calculate the entry vector center coordinates, quantify the matching deviation, perform offset vector analysis and adjust the mapping, and generate the correction mapping relationship; S5: Based on the correction mapping relationship, input the mapping data into the cost calculation model, combine the component quantity and material price, perform cost summary calculation, and generate the project cost prediction result.

[0020] The component multidimensional semantic feature dataset includes component functional semantic categories, construction attribute feature labels, material semantic attribution identifiers, and geometric morphological feature parameters. The item multidimensional semantic feature dataset includes list item semantic types, quota code semantic orientations, and pricing attribute feature identifiers. The mapping relationship table includes component item correspondence, matching validity identifiers, and similarity level distinction results. The corrected mapping relationship includes deviation correction coefficients, consistency test results, and optimized correspondence indexes. The engineering cost prediction results include sub-item cost results, total overall engineering cost, and cost composition ratio indicators.

[0021] Please see Figure 2 The specific steps for obtaining the multidimensional semantic feature dataset of components are as follows: S111: Obtain the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, perform word segmentation, perform symbol mapping on the name field, map the material code to a value and normalize it with the component geometric feature ratio, and generate a component text structure code set. A connection channel is established with the SQL Server storage based on the underlying communication protocol of the BIM database. The "three-story above-ground core tube shear wall" component instance in the target project is located. The original field data of this component is extracted via the ODBC interface. This includes extracting the functional description text "load-bearing, seismic, fireproof partition," the construction method text "C35 reinforced concrete pouring, double-sided plywood formwork support," the main material name field "C35 commercial concrete," the component's geometric feature ratio value of 14.0, and the material code "MAT-CON-35." A Chinese word segmentation engine is then used to perform a bidirectional maximum matching algorithm on the functional description and construction method texts, segmenting the text stream into independent words such as "load-bearing," "seismic," "C35," "reinforced concrete," and "pouring." A pre-set engineering symbol mapping table is called to map the non-standard character "C35" in the main material name to the strength grade identifier 35.0. Unit symbols such as "mm" are removed and converted to standard floating-point format. Range normalization is performed on the geometric feature ratio, and the preset value range for the height-to-thickness ratio of this type of component is set to [value missing]. Read the current component's height-to-thickness ratio value of 14.0 and perform the calculation. The normalized geometric feature value of 0.6842 is obtained. At the same time, the material code "MAT-CON-35" is subjected to Hash discretization and mapped to the integer index value 1024. The processed word segmentation sequence, the mapped name identifier, the normalized geometric feature value and the material code index are encapsulated in a predetermined byte order to construct a heterogeneous data package containing text modalities and numerical modalities. This completes the structured reorganization of the basic information of the component and generates a component text structure encoding set.

[0022] S112: Based on the component text structure encoding set, perform word frequency statistics and position calculation on the word encoding sequence, map the word to semantic vector, weight the structural parameter vector, calculate the cosine similarity, filter samples with a consistency benchmark value, align the text semantic vector and the structural parameter vector and concatenate them to generate a component multidimensional semantic vector matrix. Based on the generated component text structure encoding set, each word in the segmented sequence is traversed, and the frequency of high-frequency words such as "concrete" and "steel bar" in the descriptive text is counted. The relative position index of the word in the text stream is calculated. For example, the word "concrete" is located at the 4th position in the sequence, with a total length of 15, and its relative position value is calculated. The pre-trained Word2Vec engineering-specific embedding model is invoked. Using the word text as an index, the corresponding 128-dimensional word vector is found. The positional information is encoded into a positional vector of the same dimension and added bitwise with the word vector to generate a position-aware semantic vector. Subsequently, the structural parameter vector is weighted. The normalized value of the geometric features is read as 0.6842. The weighting coefficient for the influence of geometric features on cost is set to 0.35, and the weighting coefficient for the influence of material codes on cost is set to 0.65. A weighted operation is then performed. The algorithm combines the embedded vector of the material code with feature fusion, followed by semantic and structural consistency judgment. It calculates the cosine similarity between the semantic vector generated for "C35 commercial concrete" and the material code vector for "MAT-CON-35". If the calculated result (0.92) is greater than the preset consistency judgment benchmark of 0.85, the data is deemed valid; otherwise, an anomaly marker is triggered. After passing the judgment, the text semantic vector sequence and the weighted structural parameter vector are concatenated along the feature channel dimension to construct a dimension... The mixed feature matrix, where For sequence length, Using structural parameters as the dimension, physical alignment of heterogeneous features is completed, generating a multidimensional semantic vector matrix of components.

[0023] S113: Call the component multidimensional semantic vector matrix, perform dimension normalization and numerical normalization on the component vectors, perform index mapping based on the component identifier, arrange and verify the vector set, and generate a component multidimensional semantic feature dataset. The component's multidimensional semantic vector matrix is ​​invoked, and Z-Score normalization is performed on each column of the feature vectors in the matrix to calculate the arithmetic mean of the current batch of vectors. with standard deviation For specific values ​​in the matrix Perform operation The calibrated value was 1.083, eliminating the numerical scale differences between features of different dimensions. Dimensional normalization was then performed, and principal component analysis (PCA) was used to reduce the dimension of the vector matrix, preserving the cumulative variance contribution rate. The former Each principal component is used to compress the original sparse matrix into a dense feature matrix. Based on the component's globally unique identifier (GUID) "8F3E2A…", the corresponding storage address pointer is found in the global index table using the B+ tree algorithm. The compressed feature matrix is ​​written into the corresponding memory address block, and the corresponding cost label data is arranged as shown in Table 1. The integrity of the generated feature vector set is checked to see if there are any NaN values ​​or infinity values. After confirming that there are no errors, the component's feature matrix is ​​stacked row by row and encapsulated into a tensor file in HDF5 format to generate a component multidimensional semantic feature dataset.

[0024] Table 1: Example of component feature vector processing and label mapping: ; As shown in Table 1, Table 1 lists the key feature parameters of different components and the cost labels of the final mapping, showing the feature mapping relationship from the original physical parameters to the normalized values, and then to the feature mapping relationship after combining the weight coefficients. The normalized feature values ​​are the key numerical basis for the input of the subsequent neural network model.

[0025] Please see Figure 3 The specific steps for obtaining the multidimensional semantic feature set of an item are as follows: S211: Based on the component multidimensional semantic feature dataset, call the term encoding sequence and structural parameters in the component text structure encoding set, perform encoding index mapping and positional relationship calculation, and perform vector mapping according to the geometric feature ratio and material code normalization result to generate an item structure semantic vector set; The generated multidimensional semantic feature dataset of components is read, and the target component data in the storage block is locked through the HDF5 file system interface. The row index data corresponding to the component "above-ground three-story core tube shear wall" is extracted to obtain the term encoding sequence contained therein. The component's geometric feature ratio normalization value is 0.6842, and the material code hash value is 1024. The term encoding sequence is traversed, and the index position of the first term "load-bearing" in the text stream is identified as 0. The maximum text sequence length truncation value is set to 20, and position normalization calculation is performed. The last term "concrete" was identified as having an index of 4, and a calculation was performed. The process involves constructing term position vectors containing relative location information, calling a pre-defined inverted index table, retrieving the corresponding material attribute feature vector using the material code hash value 1024 as the key, expanding the geometric feature vector by a scale of 0.6842 to a scalar vector of the same dimension as the term vector, performing a vector space mapping operation, and adding the position vector, material attribute vector, and expanded geometric feature vector element-wise for each time step in the sequence. eigenvectors With geometric characteristic scalar Perform fusion operation Set the fusion coefficient If it is 0.7, The value of a certain dimension is Then calculate The feature reorganization of a single component is completed. Following this logic, 1500 component entries in the dataset are traversed to generate an entry structure semantic vector set that integrates geometric and material properties.

[0026] S212: Based on the semantic vector set of the item structure, semantic alignment calculation is performed on the item vector in multiple dimensions according to the semantic identifier of the key quota encoding. Weights are assigned according to the similarity distribution between dimensions. Weight consistency verification is performed and the vector structure is reconstructed to generate the semantic feature vector set of the item. Read the item structure semantic vector set, load the standard quota database in the field of engineering cost, extract the benchmark semantic vector of the key quota code "A4-5-12" (corresponding to cast-in-place reinforced concrete wall), place the item structure semantic vector of the current component in the same vector space as the benchmark vector, and perform Manhattan distance calculation between dimensions. Calculate the absolute value of the difference based on each feature dimension. For example, if the value of the first dimension of a component vector is 0.55526, and the corresponding value of the baseline vector is 0.52, then the difference is... An inverse proportional weight allocation function is constructed based on the difference values, and a smoothing factor is set. Calculate the original weights ,get After calculating the weights for all 64 feature dimensions, Softmax normalization is performed to map the dimension weights to... Given an interval with a sum of 1, and setting the variance threshold for consistency verification to 0.05, calculate the variance of the normalized weight distribution. If the variance is 0.02 and less than the threshold of 0.05, the weight distribution is considered uniform and effective. Subsequently, the Hadamard Product operation is performed on the original entry vector using the normalized weight vector, i.e., element-wise multiplication. The feature distribution intensity of the reconstructed vector is determined, and the reconstructed vectors that pass the verification are classified and stored according to the quota encoding to generate the semantic feature vector set of the item.

[0027] S213: Based on the semantic feature vector set of the entries, perform pairwise similarity calculation on the entry vectors, perform clustering and sequential arrangement based on the similarity results, perform structural consistency verification, and generate a multidimensional semantic feature set of the entries; Call the semantic feature vector set of the entries to construct A similarity matrix container, where Select any two entry vectors to represent the number of entries in the current batch. and Perform cosine similarity calculation, and calculate the vector similarity respectively. and Length of the module and Calculate the dot product of the two. Perform division operation If vector The corresponding numerical sequence is ,vector for The dot product is The module lengths are respectively and The similarity results are The similarity threshold for clustering determination is set as follows: The similarity result is greater than The entries are divided into the same group and assigned a unique cluster ID. The cluster groups are then sorted in descending order based on the number of members within each cluster, as shown in Table 2. A structural consistency check is performed on each sorted cluster, calculating the cluster center vector and verifying that the Euclidean distance between each member vector and the center vector is less than the maximum discrete radius. Outliers exceeding the radius are removed, and the verified vector set is written into a contiguous memory block in order of cluster ID to generate a multidimensional semantic feature set for each entry.

[0028] Table 2: Example table of item clustering and consistency verification results: ; As shown in Table 2, Table 2 lists the clusters assigned to different entry IDs after similarity calculation and their distance offset from the cluster center. Among them, "TM-2023-104" was marked as a verification failure because its offset of 0.15 exceeded the set discrete radius of 0.1. This result indicates that although the entry was classified into the CL-05 cluster in the initial classification, its internal feature structure is significantly different from the mainstream features in the cluster and needs to be removed or reprocessed. The entries with the status of Pass are confirmed to meet the structural consistency requirements and will be retained in the final feature set.

[0029] Please see Figure 4 The specific steps for obtaining the mapping table are as follows: S311: Call the multidimensional semantic feature set of entries and the multidimensional semantic feature set of components, extract the dimensional components of the entry vector and the component vector, perform consistency check on the number of dimensions of the two types of vectors, perform dimension alignment for vectors with inconsistent number of dimensions, and generate a vector dimension alignment result set. Retrieve the multidimensional semantic feature set of entries and the multidimensional semantic feature dataset of components from the memory, and read the dimension attribute values ​​of the entry feature matrix. Dimensional attribute values ​​of the component feature matrix Execute numerical comparison logic The system identified a discrepancy in the number of dimensions and identified the index in the entry vector as... to Redundant feature channels are slicing to cut the dimension range of the entry vector to a smaller value. , keep before Each principal feature component is read again, and the truncated dimension value is confirmed. The process involves completing the physical alignment of the dimensional space, then traversing the aligned vector set to check for any zero vectors or abnormal data with a magnitude of NaN. The cleaned vector matrix is ​​then repackaged according to the row index to generate a vector dimension alignment result set.

[0030] S312: Based on the vector dimension alignment result set, extract the dimension component values ​​of the entry vector and component vector, and perform pairwise cosine similarity calculation on the entry vector and component vector, using the formula: ; Calculate the similarity index of item components, filter the pairings of item components that exceed the matching threshold, and generate a candidate matching pairing set; in, Representing the The first entry and the first Similarity index between components Representing the The entry vector at the th entry vector in the th th The numerical values ​​of the components of the dimension, Representing the The component vector in the th... The numerical values ​​of the components of the dimension, Represents the total number of vector dimensions. Representing the The first entry and the first Semantic offset between components Represents the semantic offset attenuation coefficient; Read the numerical matrix from the vector dimension-aligned result set, for the formula: ; The summation is an accumulation operation across the vector dimensions, used to aggregate the feature correlations in each dimension, and the dot product operation in the numerator is also included. Used to quantize the overlap of projections of vectors in the feature space, the denominator part and These are the Euclidean norms (modulus) of the entry vector and the component vector, respectively, used to eliminate the interference of vector scale differences on similarity determination. It is an exponential decay function based on the natural constant, and its logical purpose is to introduce a semantic offset penalty mechanism, when the semantic offset... When it increases, through the attenuation coefficient Control similarity index The coefficient of variation decreases rapidly and nonlinearly, thus filtering out pseudo-matches that are similar in direction but have large deviations in semantic connotation. The advantage of this formula is that by introducing an exponential decay term, it adds a sensitivity constraint on semantic distance to the traditional cosine similarity, effectively solving the distance failure problem caused by the "curse of dimensionality" in high-dimensional space, and significantly improving the accuracy of matching engineering cost items.

[0031] In the specific calculation process, the first... Entry vectors for (To simplify the explanation, three main dimensions are selected.) The first dimension is... component vectors for The semantic offset decay coefficient is set through regression analysis of historical project data. for The semantic offset between the two is calculated through subspace projection. for First, calculate the vector dot product: ; Next, calculate the vector magnitude: ; ; Calculate the fundamental cosine similarity: ; Calculate the semantic decay factor: ; Finally, calculate the similarity index of the item components. : .

[0032] This result indicates that although entries and components are extremely close in the vector direction (basic similarity as high as 0.9981), due to the existence of... The semantic offset, after being weighted by a decay coefficient, is finally corrected to the following: This reflects the subtle semantic differences between the two. (Regarding the sample set...) Perform the above calculation on the vectors, and the statistical similarity distribution interval is: Set the matching threshold to the upper quartile of the distribution. Calculation results With threshold Compare and determine The pairing is retained, as shown in Table 3. The above process is repeated until all combinations are traversed, generating a candidate matching pairing set.

[0033] Table 3: Calculation and Screening Results of Item and Component Similarity: ; As shown in Table 3, Table 3 lists the detailed calculation parameters and screening status of different pairing combinations after being substituted into the formula. Among them, the P-002 combination was eliminated because its semantic offset was large, resulting in a final similarity below the threshold.

[0034] S313: Based on the candidate matching pairing set, extract the semantic offset values ​​of the entry vector and component vector in the pairing combination, compare the semantic offset values ​​with the offset threshold, perform an anomaly marking operation on the pairing combination whose semantic offset values ​​exceed the offset threshold, and establish a mapping relationship table. Read the data stream from the candidate matching pair set and extract the semantic offset value for each pair. Before selection Discreteness statistics were performed on each paired sample, and the arithmetic mean of the semantic offsets was calculated. and standard deviation According to the normal distribution The upper limit of the fluctuation range is set in principle as follows: This value Establish the offset threshold, traverse the current pairing set, and read the semantic offset value of a pair. Perform numerical comparison If the pairing is deemed normal, read the semantic offset value of the other pair. Perform comparison The system determines that the pairing has a semantic drift anomaly and sets the "abnormal status" field of the pairing to zero in the database. It records the associated entry ID and component ID, establishes index key-value pairs containing exception markers, and generates a mapping table.

[0035] Please see Figure 5 The specific steps for obtaining the corrected mapping relationship are as follows: S411: Call the mapping table, extract the component function classification number, geometric ratio value and material code in the paired combination, perform mean calculation on the geometric ratio value under the same function classification number, perform encoding mapping conversion on the material code and calculate the weighted mean, and obtain the cluster center coordinate set; The generated mapping table is invoked, and the database retrieval engine filters out valid paired data marked as "abnormal state" as 0. The set of records with the component function classification number "FC-Wall-Shear-03" is then identified. This set covers 5 above-ground core tube shear wall component samples, and the geometric ratio numerical sequence (height-to-thickness ratio) of the components is extracted. Perform arithmetic mean calculation Obtain the geometric feature cluster center components under this functional classification, and simultaneously extract the material code sequence. The system calls a pre-defined material property mapping table to convert the code into a standard compressive strength value encoding sequence. The weighting coefficient is calculated based on the volume ratio of the components in the BIM model. Perform weighted average calculation Geometric mean With material mean Combine and construct the two-dimensional cluster center coordinate vector under this functional classification. traverse the project in this way Each functional classification number generates a set of cluster center coordinates containing the feature distribution of the entire domain components.

[0036] S412: Based on the cluster center coordinate set and mapping table, extract the numerical values ​​of the dimensional components of the entry vector, and calculate the category center vector for the dimensional components of the entry vectors under the same functional classification number, using the following formula: ; Calculate the center deviation of the category entries and perform a difference operation with the cluster center coordinate values ​​to generate a set of quantified matching deviation values; in, Representing the Central deviation of category entries under each functional classification number Representing the Under the first functional category number The component values ​​extracted from each entry vector within the same analytical dimension as the cluster center coordinate vector. Representing the The cluster center coordinate vectors under each functional classification number are used to calculate the component values ​​of the corresponding analysis dimension in the current bias calculation. Representing the Normalized weight coefficients of each entry vector. Representing the Number of entry vectors under each functional classification number Representing the The mean of the similarity index of paired combinations under each functional classification number; Based on the cluster center coordinate set, regarding the formula: ; Perform logical analysis and operations, where Indicates the first The absolute deviation between each entry vector component and its corresponding component in the cluster center coordinate vector under the same analysis dimension. These are normalized weighting coefficients used to adjust the confidence levels of different data sources for each entry. The mean similarity of paired combinations under this category is used as a smoothing term in the correction factor. The advantage of this formula is that it uses the mean similarity to inversely scale the deviation. When the matching similarity is high, it automatically reduces the estimated value of systematic deviation, thereby avoiding over-correction of high-confidence data and ensuring the stability of cost prediction. In specific implementation, the functional classification number "FC-Wall-Shear-03" is selected, and the number of paired entries under this category is known. Cluster center geometric components Extract the corresponding geometric feature components of the three matching entries. for Set normalized weight coefficients for The mean similarity of the pairings was obtained from the results. Calculate the cumulative term in the numerator: ; ; ; Summation result: ; Calculate the denominator: ; Calculate the center deviation of category entries : ; Then, the difference operation is performed to calculate... and relative ratio deviation As shown in Table 4, this result indicates that the entry database data has approximately [a certain degree of geometrical variation] compared to the actual components. The systematic drift generates a set of quantized matching deviation values.

[0037] Table 4: Calculation and Correction Parameters for Functional Classification Center Deviation ; As shown in Table 4, Table 4 lists the component characteristic parameters and the calculated deviations under different functional classification numbers, including the deviations under the FC-Beam-12 classification. A value of "extremely small" indicates that the entry matches the component characteristics well and does not require significant subsequent modifications.

[0038] S413: Based on the matching deviation quantization value set, extract the category item center deviation and cluster center coordinate values, perform weighted average calculation on the numerical differences, obtain the offset vector component values, perform reverse correction operation on the item vector coordinates, and establish a correction mapping relationship; Based on the generated set of matching deviation quantification values, for the "FC-Wall-Shear-03" classification, the center deviation of the category entries is extracted. Cluster center coordinates A bias description vector is constructed in the feature space. Based on the bias description, a weighted average correction strategy is adopted, and the correction step size coefficient is set. The actual correction amount for the entry feature is calculated as follows: Read the component values ​​of the original entry feature vector under the corresponding analysis dimension. Perform reverse correction operation: Through a linear correction process, the entry features are pulled back to the cluster center distribution area in the opposite direction of the deviation, thereby eliminating the systematic distribution error introduced by different data sources. The corrected feature component values ​​are overwritten to the feature database, and a correction log is recorded synchronously to establish a corrected mapping relationship between entry features and cluster centers.

[0039] Please see Figure 6 The specific steps for obtaining the project cost prediction results are as follows: S511: Based on the correction mapping relationship, the mapping data is matched with the component coding rules item by item, the component quantity parameter associated with each mapping item is retrieved, and a weighted operation is performed on the component quantity parameter and the correction coefficient in the mapping data to generate a component correction quantity set; The established calibration mapping table is invoked, the component parameter correction processing engine is initialized, and the component coding rule library in the engineering quantity calculation standard "GB-50854-2013" is loaded. Each record in the mapping table is traversed, and the mapping item with index ID "Map-Link-092" is locked. This item is associated with a set of components categorized as "three-story above-ground core tube shear wall". The unique identifier "GUID-Wall-C35-001" of the component is extracted through the mapping relationship. The original engineering quantity parameters of the component are retrieved by calling the BIM model database interface, and the net concrete volume is obtained. and template contact area Simultaneously read the geometric feature deviation recorded in the mapping data. And the correction direction indicator, based on the deviation amount, calls the correction coefficient calculation logic, and sets the deviation-coefficient conversion function. Calculate the volume correction factor Perform weighted calculations to determine the corrected concrete volume. Similarly, for the template area, the corresponding surface area deviation coefficient is read. Perform calculation Perform a reasonableness boundary check on the corrected data and set the maximum allowable increase threshold as follows: ,verify If the correction is deemed valid, and the verification fails, an exception is logged and the original value is retained. This logic is then applied to the remaining entries in the list. Each component mapping item structurally binds the calculated corrected engineering quantity value with the corresponding component code to generate a component correction quantity set.

[0040] S512: Based on the component correction quantity set, collect the material price data of the corresponding component type, perform a weighted calculation on the material price value and the component correction quantity item by item, and perform numerical aggregation on the weighted results of the same type of components to obtain the total component cost value. Read the component correction quantity set, connect it to the "December 2025 Construction Engineering Material Market Price Database", and extract the correction quantity of "C35 ready-mixed concrete" for the "core tube shear wall" component type. Query material information prices for this region at this point in time. Yuan / Set the material loss rate to Perform weighted cost calculation including losses. Yuan, extract the corrected quantity for "18mm laminated plywood template" Check the comprehensive unit price (including labor costs). Yuan / Perform calculation Yuan, extract the corrected quantity of the associated "HRB400E rebar". Tons, check the market price of steel bars Yuan / ton, calculate the cost of steel bars After completing the individual calculations, the numerical aggregation operation for similar components is performed, adding up the concrete cost, formwork cost, and steel reinforcement cost to calculate the total value of the shear wall section. As shown in Table 5, the total cost of the "cast-in-place frame beam" section is calculated using the same process, including the cost of concrete. Yuan and steel reinforcement cost Yuan, to obtain the total value of the beam division. The total cost of components is calculated by iterating through the sub-projects until the entire project's component pricing is completed.

[0041] Table 5: Detailed Table of Component Cost Calculation and Summary ; As shown in Table 5, Table 5 details the specific material items under the component sub-section, the revised quantities, the corresponding market unit prices, and the calculated total costs for each item and sub-section. The total sub-section value for the "core tube shear wall" is shown in Table 5. The amount will serve as the basis for subsequent cost calculations.

[0042] S513: For the total component cost, input the component cost values ​​into the summary rules set in the cost calculation model, perform sequential integration and pricing caliber alignment processing on the cost value sequence, and generate the project cost prediction results. For the total component cost, the total cost of each sub-project (such as shear walls, frame beams, slabs, etc.) is added up to obtain the total cost of each sub-item. The value is entered into the "2025 Construction Engineering Cost Calculation Model". The model's preset "General Tax Calculation Method" aggregation rule is called. First, the cost of the measures items is integrated sequentially, and the cost of each sub-item of the project is multiplied by the safety and civilized construction fee rate. ,calculate Yuan, multiplied by the rate of other measures fees ,calculate The pre-tax project cost base is obtained by adding the itemized costs and the provisional costs together. Next, indirect costs are calculated, applying the enterprise management fee rate. ,calculate Yuan, application profit margin ,calculate Yuan, calculate the regulatory fee The total cost, excluding tax, is calculated in yuan. The final amount is calculated using value-added tax, applying the standard tax rate for the construction industry. Calculate taxes Yuan, performing final numerical integration The result is then aligned to two decimal places to generate the project cost prediction result.

[0043] The BIM-driven engineering cost prediction system is used to execute the aforementioned BIM-driven engineering cost prediction method. The system includes: The component feature extraction module obtains the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, and uses BERT to perform text encoding to extract multi-dimensional feature vectors of component text, generating a multi-dimensional semantic feature dataset of components. The item feature extraction module extracts structural semantic phrases and key quota codes of bill of quantities items based on the component multidimensional semantic feature dataset, calculates item feature vectors using word vector model, and constructs item multidimensional semantic feature set; The matching degree calculation module calls the multidimensional semantic feature set of entries and the multidimensional semantic feature dataset of components, calculates the matching degree based on cosine similarity, filters component entry pairs whose matching degree exceeds the matching threshold, marks abnormal matching pairs, and generates a mapping relationship table. The mapping relationship correction module calls the mapping relationship table, calculates the cluster center coordinates based on the component functional classification number, geometric scale and material code, calculates the entry vector center coordinates, quantifies the matching deviation, performs offset vector analysis and adjusts the mapping, and generates a correction mapping relationship. The cost prediction module, based on the correction mapping relationship, inputs the mapping data into the cost calculation model, combines the number of components and material prices, performs a cost summary calculation, and generates the project cost prediction result.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A BIM-driven method for predicting engineering costs, characterized in that, Includes the following steps: S1: Obtain the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, use BERT to perform text encoding to extract multi-dimensional feature vectors of component text, and generate a multi-dimensional semantic feature dataset of components. S2: Based on the multidimensional semantic feature dataset of the components, extract the structural semantic phrases and key quota codes of the bill of quantities items, use a word vector model to calculate the feature vectors of the items, and construct a multidimensional semantic feature set of the items; S3: Call the multidimensional semantic feature set of the entries and the multidimensional semantic feature dataset of the components, calculate the matching degree based on cosine similarity, filter the component entry pairs whose matching degree exceeds the matching threshold, mark abnormal matching pairs, and generate a mapping relationship table; S4: Call the mapping relationship table, calculate the cluster center coordinates according to the component functional classification number, geometric scale and material code, calculate the entry vector center coordinates, quantify the matching deviation, perform offset vector analysis and adjust the mapping, and generate a correction mapping relationship; S5: Based on the corrected mapping relationship, input the mapping data into the cost calculation model, combine the number of components and material prices, perform a cost summary calculation, and generate the project cost prediction result.

2. The BIM-driven engineering cost prediction method according to claim 1, characterized in that, The component multidimensional semantic feature dataset includes component functional semantic categories, construction attribute feature labels, material semantic attribution identifiers, and geometric morphological feature parameters. The item multidimensional semantic feature set includes list item semantic types, quota code semantic orientations, and pricing attribute feature identifiers. The mapping relationship table includes component item correspondence, matching validity identifiers, and similarity level distinction results. The corrected mapping relationship includes deviation correction coefficients, consistency test results, and optimized correspondence indexes. The engineering cost prediction results include sub-item cost results, total overall engineering cost, and cost composition ratio indicators.

3. The BIM-driven engineering cost prediction method according to claim 1, characterized in that, The steps for obtaining the component multidimensional semantic feature dataset are as follows: S111: Obtain the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, perform word segmentation, perform symbol mapping on the name field, map the material code to a value and normalize it with the component geometric feature ratio to generate a component text structure code set. S112: Based on the component text structure encoding set, perform word frequency statistics and position calculation on the word encoding sequence, map the word to semantic vector, weight the structural parameter vector, calculate the cosine similarity, filter samples with a consistency benchmark value, align the text semantic vector and the structural parameter vector and concatenate them to generate a component multidimensional semantic vector matrix. S113: Call the component multidimensional semantic vector matrix, perform dimensional normalization and numerical normalization on the component vectors, perform index mapping based on the component identifier, arrange and verify the vector set, and generate a component multidimensional semantic feature dataset.

4. The BIM-driven engineering cost prediction method according to claim 3, characterized in that, The steps for obtaining the multidimensional semantic feature set of the entries are as follows: S211: Based on the multidimensional semantic feature dataset of the component, call the term encoding sequence and structural parameters in the component text structure encoding set, perform encoding index mapping and positional relationship calculation, and perform vector mapping according to the geometric feature ratio and material code normalization result to generate an item structure semantic vector set; S212: Based on the item structure semantic vector set, according to the key quota coding semantic identifier, perform multi-dimensional semantic alignment calculation on the item vector, allocate weights according to the similarity distribution between dimensions, perform weight consistency verification and reconstruct the vector structure to generate an item semantic feature vector set. S213: Based on the set of semantic feature vectors of the items, perform pairwise similarity calculation on the item vectors, perform clustering and sequential arrangement according to the similarity results, perform structural consistency verification, and generate a multidimensional semantic feature set of the items.

5. The BIM-driven engineering cost prediction method according to claim 4, characterized in that, The steps for obtaining the mapping table are as follows: S311: Call the item multidimensional semantic feature set and component multidimensional semantic feature dataset, extract the dimension components of the item vector and component vector, perform consistency check on the number of dimensions of the two types of vectors, perform dimension alignment for vectors with inconsistent number of dimensions, and generate a vector dimension alignment result set. S312: Based on the vector dimension alignment result set, extract the dimension component values ​​of the entry vector and component vector, and perform pairwise cosine similarity calculation on the entry vector and component vector using the formula: ； Calculate the similarity index of item components, filter the pairings of item components that exceed the matching threshold, and generate a candidate matching pairing set; in, Representing the The first entry and the first Similarity index between components Representing the The entry vector at the th entry vector in the th th The numerical values ​​of the components of the dimension, Representing the The component vector in the th... The numerical values ​​of the components of the dimension, Represents the total number of vector dimensions. Representing the The first entry and the first Semantic offset between components Represents the semantic offset attenuation coefficient; S313: Based on the candidate matching pairing set, extract the semantic offset values ​​of the entry vector and component vector in the pairing combination, compare the semantic offset values ​​with the offset threshold, perform an anomaly marking operation on the pairing combination whose semantic offset values ​​exceed the offset threshold, and establish a mapping relationship table.

6. The BIM-driven engineering cost prediction method according to claim 5, characterized in that, The matching threshold is obtained by performing interval statistics on the cosine similarity values ​​between the entry vector and the component vector in the original sample data. The interval statistics are determined based on the similarity distribution range under a preset number of samples. The offset threshold is obtained by performing discrete statistics on the semantic offset values ​​of the entry vector and component vector in the original sample data. The discrete statistics are determined based on the mean and fluctuation range of the semantic offset values, and the value corresponding to the upper limit of the fluctuation range is used as the offset threshold.

7. The BIM-driven engineering cost prediction method according to claim 5, characterized in that, The steps for obtaining the correction mapping relationship are as follows: S411: Call the mapping table, extract the component functional classification number, geometric ratio value and material code in the paired combination, perform mean calculation on the geometric ratio value under the same functional classification number, perform encoding mapping conversion on the material code and calculate the weighted mean, and obtain the cluster center coordinate set; S412: Based on the cluster center coordinate set and mapping relationship table, extract the numerical values ​​of the dimension components of the item vector, calculate the category center vector for the dimension components of the item vector under the same functional classification number, calculate the category item center deviation, and perform a difference operation with the cluster center coordinate values ​​to generate a set of matching deviation quantification values. S413: Based on the set of matching deviation quantization values, extract the deviation of the category entry center and the coordinate values ​​of the cluster center, perform a weighted average calculation on the numerical differences, obtain the offset vector component values, perform a reverse correction operation on the entry vector coordinates, and establish a correction mapping relationship.

8. The BIM-driven engineering cost prediction method according to claim 7, characterized in that, The steps for obtaining the project cost prediction results are as follows: S511: Based on the correction mapping relationship, the mapping data is matched item by item with the component coding rule, the component quantity parameter associated with each mapping item is retrieved, and a weighted operation is performed on the component quantity parameter and the correction coefficient in the mapping data to generate a component correction quantity set; S512: Based on the set of component correction quantities, collect the material price data of the corresponding component type, perform a weighted calculation on the material price values ​​and component correction quantities item by item, and perform numerical aggregation on the weighted results of similar components to obtain the total component cost value. S513: For the total component cost value, input the component cost value into the summary rules set by the cost calculation model, perform sequential integration and pricing caliber alignment processing on the cost value sequence, and generate the project cost prediction result.

9. A BIM-driven engineering cost prediction system, characterized in that, The system is used to implement the BIM-driven engineering cost prediction method according to any one of claims 1-8, and the system includes: The component feature extraction module obtains the functional description text, construction method text, main material name field, component geometric feature ratio and material code of building components from the BIM database, and uses BERT to perform text encoding to extract multi-dimensional feature vectors of component text, generating a multi-dimensional semantic feature dataset of components. The item feature extraction module extracts structural semantic phrases and key quota codes of the bill of quantities items based on the multidimensional semantic feature dataset of the components, calculates the item feature vectors using a word vector model, and constructs a multidimensional semantic feature set of items. The matching degree calculation module calls the multidimensional semantic feature set of the entries and the multidimensional semantic feature dataset of the components, calculates the matching degree based on cosine similarity, filters the component entry pairs whose matching degree exceeds the matching threshold, marks abnormal matching pairs, and generates a mapping relationship table. The mapping relationship correction module calls the mapping relationship table, calculates the cluster center coordinates based on the component functional classification number, geometric scale and material code, calculates the entry vector center coordinates, quantifies the matching deviation, performs offset vector analysis and adjusts the mapping, and generates a corrected mapping relationship. The cost prediction module, based on the corrected mapping relationship, inputs the mapping data into the cost calculation model, combines the number of components and material prices, performs a cost summary calculation, and generates the project cost prediction result.

Images