Building engineering cost intelligent prediction method based on BIM and machine learning

Abstract

The invention relates to the technical field of building cost prediction, and discloses a building engineering cost intelligent prediction method based on BIM and machine learning, which comprises the steps of extracting a component parameter table from a plurality of BIM model files of heterogeneous versions based on a preset unique identifier, and further, performing multi-scale discrete barrel separation processing on component attribute data to obtain a component parameter table; generating a bucket coding matrix by combining disturbance feedback analysis, performing joint coding mapping processing on the bucket coding matrix to generate an attribute structure vector with a stable structure and unified semantics, constructing an edge index matrix by combining component connection data, and performing node attribute reconstruction processing on the attribute structure vector to obtain an edge index matrix; therefore, a component diagram structure with a topological structure and attribute information fusion capability is constructed, automatic prediction of the building cost can be realized on the premise that a manual mapping rule is not needed, and the adaptation stability of a cost estimation model in a multi-version BIM environment and the consistency of prediction results are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of construction cost prediction. More specifically, the present invention relates to an intelligent prediction method for construction project costs based on BIM and machine learning. Background Art

[0002] In construction project management, methods for predicting project costs based on Building Information Modeling (BIM) have been widely studied and applied. Existing technologies usually extract component parameter data (such as component type, size, material, construction stage, etc.) from the BIM model as input features, and use regression models, neural network models or rule inference methods to construct cost estimation models, so as to achieve intelligent prediction of component-level or stage-level project costs. Some methods further introduce construction progress information, construction process paths or project resource allocation data to improve the dynamics and real-time nature of cost estimation.

[0003] For example, existing technologies usually adopt a method of parsing and cost calculation for BIM components based on standard field rule matching, specifically including: extracting a component attribute table from the BIM model, and performing field-level matching with a preset quota library or project bill of quantities entries, and then completing the calculation of component costs and the estimation of the total project cost. Such methods often construct a mapping relationship between component attributes and project cost elements, convert BIM attributes into computable data items using static rules or templates, and achieve project-level cost prediction by summarizing the cost data of each component.

[0004] However, such cost modeling methods that rely on field matching and template rules have significant deficiencies in dealing with multi-version and multi-source heterogeneous problems of actual BIM models. On the one hand, there are non-uniform phenomena in the naming, classification methods and even unit systems of component attribute fields between different version models, resulting in the difficulty of generalization of standard fields and frequent failure of field mapping. On the other hand, existing methods mostly ignore the connection relationships and semantic structures between components, and only perform matching calculations starting from isolated attributes, lacking the ability to model component structural semantics and context relationships. In addition, the cost estimation results highly depend on manually configured rule libraries or quota mappings, and it is difficult to automatically adapt to new version models or complex model change scenarios, thus affecting the generalization ability and deployment stability of the cost prediction system.

[0005] In view of this, the present invention proposes an intelligent prediction method for construction project costs based on BIM and machine learning to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent prediction method for construction project costs based on BIM and machine learning.

[0007] To achieve the above object, the present invention provides the following technical solutions: In a first aspect, an intelligent prediction method for construction project costs based on BIM and machine learning is provided, including: Obtain S heterogeneous versions of BIM model files, and extract a component parameter table from the S heterogeneous versions of BIM model files based on a preset unique identifier. The component parameter table includes component attribute data and component connection data, where S is an integer greater than 1; Perform multi-scale discrete bucketing processing on the component attribute data to generate a bucketing coding matrix, and perform joint coding mapping processing on the bucketing coding matrix to generate an attribute structure vector; Perform connection pair extraction processing on the component connection data to generate an edge index matrix, perform node attribute reconstruction processing on the attribute structure vector to generate a node feature matrix, use the edge index matrix as an edge set, and use the node feature matrix as a node set to obtain a component graph structure; Input the component graph structure into a trained neural network model to obtain the construction cost.

[0008] In some embodiments, the method for extracting a component parameter table from the S heterogeneous versions of BIM model files based on a preset unique identifier includes: Extract the numerical fields and categorical fields of the components from the S heterogeneous versions of BIM model files; Perform structure separation processing on the numerical fields and categorical fields respectively to obtain a numerical field matrix and a categorical field coding set; Perform multi-scale normalization processing on the numerical field matrix, perform positional embedding coding processing on the categorical field coding set, and merge the normalized numerical field matrix and the coded categorical field coding set to obtain a standard field feature matrix; Construct a structured feature string based on the standard field feature matrix, perform a hash function operation on the structured feature string to generate a unique identifier, and establish a mapping relationship table from the unique identifier to a preset component index; Locate the corresponding components from the BIM model files based on the mapping relationship table, and extract the component attribute data and component connection data to obtain a component parameter table.

[0009] In some embodiments, the method for performing multi-scale normalization processing on the numerical field matrix includes: Preset a corresponding normalization scale factor for each field in the numerical field matrix; Use the field value in each field as the numerator and the corresponding normalization scale factor of each field as the denominator to perform proportional scaling processing on the field value; Align the proportionally scaled field values in the original arrangement order of each field, and splice the aligned field values horizontally to form an extended field matrix; Perform min-max normalization on each column of the extended field matrix, and use the normalized extended field matrix as the numerical field matrix after normalization processing.

[0010] In some embodiments, the method for performing multi-scale discrete bucketing processing on component attribute data to generate a bucketing coding matrix includes: Set multiple scale factors for each field value in the component attribute data, where each scale factor corresponds to a segmentation granularity and a margin value; Based on the segmentation granularity and the margin value, perform proportional reduction processing on each field value to obtain a set of reduction results, and perform interval landing point positioning operation on the set of reduction results to obtain a corresponding set of segmentation numbers; Concatenate the set of segmentation numbers of each component under all field values and all scales in the original arrangement order of the field values in the component attribute data to generate the number sequence of the component, and summarize the number sequences of all components to form the original bucketing matrix; Perform perturbation feedback analysis on the original bucketing matrix in the field dimension direction to obtain the jump rate of the corresponding numbers of each field value among all components, and rearrange the arrangement order of the field dimension according to the jump rate to generate the bucketing coding matrix.

[0011] In some embodiments, the method for performing proportional reduction processing on each field value based on the segmentation granularity and the margin value to obtain a set of reduction results includes: Based on the margin value corresponding to each scale factor, set a floating boundary buffer for the segmentation interval where the field value is located, and perform dynamic boundary correction on the field value based on the upper boundary of the buffer to obtain the corrected field value; Perform division calculation on the corrected field value and the segmentation granularity corresponding to each scale factor to obtain a set of ratio values; Perform consistency test on the ratio values in the set of ratio values at adjacent scales, and perform linear interpolation processing on the deviated ratio values according to the consistency test results to obtain a set of reduction results.

[0012] In some embodiments, the method for performing interval landing point positioning operation on the set of reduction results to obtain a corresponding set of segmentation numbers includes: Obtain each ratio value in the set of reduction results, and construct an equidistant number interval based on the scale factor corresponding to each ratio value; Perform floor operation on each ratio value to determine the segmentation number of each ratio value in the equidistant number interval; Combine the segmentation numbers obtained by each field under all scale factors in the original arrangement order of the field in the component attribute data to obtain a set of segmentation numbers.

[0013] In some embodiments, the method for performing perturbation feedback analysis on the original bucket matrix in the field dimension direction to obtain the jump rate of the corresponding numbers of each field value among all components includes: Based on the field arrangement order in the component attribute data, extract the number sequence corresponding to the field in the original bucket matrix column by column to obtain a set of component number sequences; For each column in the set of component number sequences, perform a sliding window traversal process based on a preset window length, extract adjacent component number pairs within each window, and calculate their number differences to generate a number difference sequence; Compare each difference in the number difference sequence with a perturbation threshold. If the difference is greater than or equal to the preset perturbation threshold, mark it as a jump event, count the total number of jump events, and divide the number of jump events by the total number of component samples to obtain the jump rate of the field value.

[0014] In some embodiments, the method for performing joint coding mapping processing on the bucket coding matrix to generate an attribute structure vector includes: Perform embedding coding operations on the field values of each column in the bucket coding matrix, and map the segment numbers to fixed-length embedding vectors based on a preset embedding dictionary; Aggregate the set of embedding vectors by component dimension to obtain an embedding vector sequence corresponding to each component; Concatenate the embedding vector sequences corresponding to each component in the order of fields to obtain an initial structure matrix; Perform unified mapping on the embedding vectors in the initial structure matrix in the component dimension to obtain an attribute structure vector.

[0015] In some embodiments, the method for performing connection pair extraction processing on component connection data to generate an edge index matrix includes: Obtain a set of component index pairs with connection relationships in the component connection data, where each component index pair consists of two connection fields with an associated relationship, and the connection fields are used to identify two interconnected components; Based on each component index pair, in accordance with the preset field parsing order, respectively use the component index corresponding to the first connection field as the starting node in the edge index matrix, and use the component index corresponding to the second connection field as the target node; Arrange all component index pairs in the order of field parsing and write them into a two-dimensional matrix structure to obtain an edge index matrix.

[0016] In some embodiments, the method for performing node attribute reconstruction processing on the attribute structure vector to generate a node feature matrix includes: Write the attribute structure vector corresponding to each component into the corresponding row of the matrix structure in the arrangement order of the components in the bucket coding matrix to construct an initial attribute matrix; Perform min-max normalization processing on each column field in the initial attribute matrix to obtain a normalized attribute matrix; Based on the normalized attribute matrix, calculate the Pearson correlation coefficient between each field and other fields, and construct a field correlation matrix; Calculate the correlation coefficient based on the field correlation matrix, and filter out the field pairs with the absolute value of the correlation coefficient greater than the preset correlation threshold to form a field combination set; Perform feature fusion processing on each field group in the field combination set to generate a fused field vector set; Concatenate the fused field vector set with the field vector set in the normalized attribute matrix that has not participated in the fusion to construct a node feature matrix.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention extracts the component parameter table from multiple heterogeneous versions of BIM model files based on a preset unique identifier, solving the problem of inconsistent identification of components due to differences in field naming methods or data structures between models; further, by performing multi-scale discrete bucketing processing on component attribute data, combining perturbation feedback analysis to generate a bucketing coding matrix, and then performing joint coding mapping processing on it to generate an attribute structure vector with stable structure and unified semantics, effectively enhancing the compatibility of component attribute expressions between different versions; on this basis, combining component connection data to construct an edge index matrix, and performing node attribute reconstruction processing on the attribute structure vector to generate a complete node feature matrix, thereby constructing a component graph structure with the ability to fuse topological structure and attribute information; finally, inputting the component graph structure into the trained neural network model can automatically predict the construction cost without the need for manual mapping rules, significantly improving the adaptation stability of the cost estimation model in a multi-version BIM environment and the consistency of prediction results. Brief Description of the Drawings

[0018] Figure 1 It is a schematic flowchart of the intelligent construction cost prediction method based on BIM and machine learning in the present invention; Figure 2 It is a schematic diagram of a computer-readable storage medium in the present invention. Detailed Embodiments

[0019] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the following further describes the present invention in detail with reference to specific embodiments and the accompanying drawings. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the described exemplary embodiments. However, it will be apparent to those skilled in the art that some or all of these specific details may be practiced without these specific details. In other exemplary embodiments, well-known structures are not described in detail to avoid unnecessarily obscuring the concepts of the present disclosure. It should be understood that the specific embodiments described herein are merely used to explain the present invention and are not used to limit the present invention. At the same time, various aspects described in the embodiments can be arbitrarily combined without conflict.

[0020] Embodiment 1 Please refer to Figure 1 As shown, this embodiment discloses and provides an intelligent prediction method for construction project costs based on BIM and machine learning, including: S10: Obtain S heterogeneous versions of BIM model files, and extract a component parameter table from the S heterogeneous versions of BIM model files based on a preset unique identifier. The component parameter table includes component attribute data and component connection data, where S is an integer greater than 1; In this embodiment, the heterogeneous versions of BIM model files refer to multiple versions of BIM model files generated in actual construction projects due to different factors such as the design stage, modeling unit, modeling tool, modeling specification, or data format. The heterogeneous versions of the models differ in the naming method of component attribute fields, the number of fields, data types, geometric hierarchical structure, or the organization method of component connection relationships, resulting in the component parameter tables being not completely consistent in structure or semantics. Common heterogeneous versions include preliminary design models, construction drawing models, as-built models, or model versions established by different units using different software platforms (such as Revit, Tekla, ArchiCAD). The method for obtaining heterogeneous versions of BIM model files can export multiple historical modeling versions through the model version management system in the building collaboration platform, or perform structured extraction of multi-version models through BIM modeling tools (such as Revit) combined with modeling stage identifiers.

[0021] The method for extracting a component parameter table from the S heterogeneous versions of BIM model files based on a preset unique identifier includes: Extract the numerical fields and categorical fields of the components from the S heterogeneous versions of BIM model files; Perform structural separation processing on the numerical fields and categorical fields respectively to obtain a numerical field matrix and a categorical field coding set; Perform multi-scale normalization processing on the numerical field matrix, perform position embedding encoding processing on the category field encoding set, and merge the normalized numerical field matrix with the encoded category field encoding set to obtain a standard field feature matrix; Construct a structured feature string based on the standard field feature matrix, perform a hash function operation on the structured feature string to generate a unique identifier, and establish a mapping relationship table from the unique identifier to the preset component index; Locate the corresponding component from the BIM model file based on the mapping relationship table, and extract the component attribute data and component connection data to obtain a component parameter table.

[0022] It can be understood that in the data structure of the BIM model, each component contains multiple attribute fields, such as the length, width, height, center point coordinates, material type, component type name, etc. of the component. Although these fields are uniformly represented inside the model (for example, exported in JSON or table form), they can be clearly distinguished from the perspective of this field: Fields such as "Length = 3.2 meters", "Area = 12.5 square meters", and "X coordinate = 1.6" have numerical values that can participate in mathematical calculations and are typical numerical fields; While fields such as "Material = reinforced concrete", "Type = shear wall", and "Component = beam" are text labels with limited options and belong to category fields.

[0023] In engineering practice, distinguishing field types is a basic step in all BIM attribute processing or machine learning modeling. The unit or format of the field is usually explicitly reflected in the data table (for example, the "Material" field is represented by text, and the "Length" field is a floating point number). Even if it is not displayed, users can also judge its attributes through the form of its value.

[0024] After completing the division of field types, the data is naturally divided into two parts: One group is numerical fields, such as "Length = 5.0" and "X coordinate = 23.1"; One group is category fields, such as "Material = C30" and "Type = shear wall".

[0025] Numerical field matrix: Arranging the numerical fields of all components in order will form a two-dimensional table. For example, the first column is "Length", the second column is "Height", and the third column is "X coordinate", and each row represents a component.

[0026] Category field encoding set: Since category fields are text labels and cannot directly participate in numerical operations, they are usually converted into digital encodings. For example, if the "Material" field may contain "C20", "C30", "C40", it can be encoded as [1, 2, 3], or a "coding dictionary" can be constructed for management.

[0027] A method for performing multi-scale normalization processing on a numerical field matrix includes: Presetting a corresponding normalization scale factor for each field in the numerical field matrix; Taking the field value in each field as the numerator and the corresponding normalization scale factor of each field as the denominator, and performing ratio scaling processing on the field value; Aligning the ratio-scaled field values in the original arrangement order of each field, and splicing the aligned field values horizontally to form an extended field matrix; Performing min-max normalization processing on each column in the extended field matrix, and using the normalized extended field matrix as the numerical field matrix after normalization processing.

[0028] The process of performing multi-scale normalization processing on a numerical field matrix includes the following steps. First, for each field in the numerical field matrix, according to the physical meaning and typical value range of the field in the actual construction project, preset one or more normalization scale factors. For example, the length field can be set to 5 meters, 10 meters, the area field can be set to 100 square meters, etc., to represent the feature expression ability at different scales. Second, taking the field value in each field as the numerator and the corresponding normalization scale factor as the denominator, perform ratio scaling processing on the field value to obtain a set of field values at multiple scales. For example, if the length of a component is 4 meters and the scale factors are 5 meters and 10 meters, then the two ratio values 0.8 and 0.4 can be obtained after scaling.

[0029] Next, for each set of field values obtained under each normalization scale factor, it is necessary to keep its arrangement order in the component dimension consistent with the original numerical field matrix, that is, each row still corresponds to the same component unit, to ensure that there will be no disorder in the component order during the subsequent splicing process. After completing the order alignment, perform horizontal splicing processing on the field values under each normalization scale in the field dimension direction, that is, connect the ratio scaling processing results of the same field at different scales in sequence along the horizontal direction to construct an extended field matrix. This extended field matrix can be used to simultaneously retain the feature expression information of each field under multi-scale conditions. For example, if a component obtains normalized values 0.8 and 0.4 respectively after being processed by two scale factors in the "length" field and obtains 0.6 and 0.3 in the "area" field, then a row in its extended field matrix is represented as [0.8, 0.4, 0.6, 0.3].

[0030] Finally, perform normalization processing on each column of field values in the extended field matrix. Specifically, calculate the minimum and maximum values of each column of field values, and use the minimum value as the starting point and the difference between the maximum value and the minimum value as the denominator to perform linear proportional scaling on all field values in this column. The above normalization processing can uniformly map the field values of each column to a standard numerical interval, such as [0, 1], while enhancing the comparability of different scale expressions and maintaining the numerical differences between components, thus finally obtaining a normalized numerical field matrix, providing a consistent and stable numerical input structure for constructing the standard field feature matrix in the subsequent steps.

[0031] The method for performing positional embedding encoding processing on the category field encoding set includes: Perform a mapping operation on each field value in the category field encoding set to convert it into a corresponding integer category index; Obtain the category field sequence corresponding to the field value in the category field encoding set, and sequentially splice the integer category indexes in order according to the category field sequence to construct a category index matrix; Assign unique field position information to each column of field values in the category index matrix, and perform fusion processing on the category index matrix and its corresponding field position information based on a preset embedding function to generate an embedding representation matrix; Use the embedding representation matrix as the category field encoding set after encoding processing.

[0032] To achieve the embedded expression of the category field, the process of performing positional embedding encoding processing on the category field encoding set in this embodiment includes the following steps. First, perform a mapping operation on each field value in the category field encoding set to convert the original text category labels (such as "shear wall", "C30 concrete", etc.) into corresponding integer category indexes. For example, by constructing a static encoding dictionary, map "shear wall" to 0, "frame beam" to 1, etc., so that the category field can participate in numerical representation and subsequent calculations. Subsequently, extract the field sequence information in the category field encoding set, that is, determine the arrangement order of each field in the original input data, and sequentially splice all integer category indexes according to this field sequence to construct a category index matrix; where each row in this matrix corresponds to a sample record, and each column corresponds to a specific category field. Then, assign unique field position information to each column of the category index matrix. For example, number them in order from left to right (such as 0, 1, 2...) to identify the original position of each field in the overall category structure and retain the structural order characteristics between fields. Finally, perform fusion mapping processing on the category index matrix and its corresponding field position information based on a preset embedding function (such as table lookup embedding, positional encoding function, or fusion projection network) to generate an embedding representation matrix. This embedding representation matrix, as the category field encoding set after encoding processing, can be used for feature splicing with the numerical field matrix to form a unified input feature structure.

[0033] It should be added that the method of merging the normalized numerical field matrix and the encoded category field encoding set to obtain the standard field feature matrix can be to respectively determine the row consistency of the normalized numerical field matrix and the embedding representation matrix (i.e., the encoded category field encoding set), ensure that the two are in one-to-one correspondence by row in the dimension of the number of samples, and then, on the premise of keeping the original component order unchanged, perform a feature-level concatenation operation on the numerical field vector and the category field embedding vector corresponding to each component. Specifically, the two vectors can be concatenated end to end in the column direction to form a unified high-dimensional feature vector. By performing the above concatenation operation on all component samples in turn, the standard field feature matrix can be constructed.

[0034] In this embodiment, a component alignment mechanism based on a unique identifier is introduced in the processing logic, and a construction path for the standard field feature matrix is designed. By performing multi-scale normalization processing on the numerical field and position embedding encoding processing on the category field, it can effectively solve the problem of inconsistent semantic of component attributes caused by differences in field naming methods, data types, or arrangement structures, and realize the automatic normalized expression and structural unified reconstruction of component attribute data in multiple version models without relying on artificial field mapping rules.

[0035] Even in a heterogeneous BIM model environment with inconsistent component field naming and format drift in the data structure, it can still stably generate a component parameter table with strong consistency and available for cost prediction modeling; and through the construction method of the standard field feature matrix, the stability and comparability of subsequent model inputs are enhanced, and the adaptation ability and deployment robustness of multi-version models in component-level data processing tasks are significantly improved.

[0036] It should be noted that the standard field feature matrix is constructed by feature-level concatenation after the normalization processing of the numerical field and the encoding processing of the category field, and is used to uniformly express the standardized attribute features of multiple components under different fields; while the attribute structure vector is a component structure representation vector generated after the combined encoding mapping processing is performed after the construction of the bucket encoding matrix. The two respectively correspond to the two processing stages of component attribute normalization expression and component structure modeling. The standard field feature matrix serves as the standard input form of the original component attributes and serves the subsequent multi-scale discrete bucketing processing process.

[0037] S20: Perform multi-scale discrete bucketing processing on the component attribute data to generate a bucket encoding matrix, and perform combined encoding mapping processing on the bucket encoding matrix to generate an attribute structure vector; In this embodiment, component attribute data may refer to a set of fields extracted from a component parameter table that has the ability to express specific process or structural attributes, typically including but not limited to component size parameters (such as length, width, height), position parameters (such as the X, Y, and Z coordinates of the component center point), material type fields (such as "C30 concrete" and "steel structure"), construction type fields (such as "shear wall" and "frame column"), component levels (such as the floor on which they are located), component identification information (such as component type code and component number), etc. These fields are usually stored in the form of attribute pairs (key-value pairs) in multiple versions of BIM models and are typical components of component attribute data. In the actual processing process, the component attribute data, after unified field normalization, screening, and encoding conversion, can be used as input to perform multi-scale discrete bucketing processing and joint encoding mapping processing, thereby generating an attribute structure vector reflecting the component structural semantics.

[0038] The method of performing multi-scale discrete bucketing processing on component attribute data to generate a bucket coding matrix includes: Setting multiple scale factors for each field value in the component attribute data, wherein each scale factor corresponds to a segment granularity and a margin value; Based on the segment granularity and margin value, a proportional reduction process is performed on each field value to obtain a reduction result set. An interval location operation is performed on the reduction result set to obtain a corresponding segment number set. The segment number sets of each component under all field values and all scales are spliced according to the original arrangement order of the field values in the component attribute data to generate the component number sequence. The number sequences of all components are summarized to form the original bucket matrix. Perturbation feedback analysis is performed on the original bucket matrix in the field dimension direction to obtain the jump rate of each field value in the corresponding number of all components, and the arrangement order of the field dimension is rearranged according to the jump rate to generate a bucket coding matrix.

[0039] In this embodiment, multiple scale factors are set for each field value in the component attribute data. Among them, each scale factor corresponds to a segmentation granularity and a margin value. The scale factor is used to control the segmentation precision of the field value, while the margin value is used to buffer and expand the segmentation boundary to accommodate the boundary value fluctuations caused by modeling errors or unit differences between versions. Taking the "length" field as an example, when the field value range is distributed between 1 meter and 9 meters, the scale factors can be set to 5 meters, 2 meters, and 1 meter, corresponding to coarse-grained, medium-grained, and fine-grained segmentations respectively. For the field values at each scale, proportional reduction processing is performed based on the corresponding granularity and margin parameters, and interval landing point positioning is performed according to the reduced values to obtain the segmentation numbers of each field value at each scale. Through multi-scale setting, the adaptability of the field in the heterogeneous version BIM model and the redundant expression ability of feature segmentation can be enhanced, and the subsequent component numbering stability and cross-version compatibility can be improved.

[0040] The method for performing proportional reduction processing on each field value based on the segmentation granularity and the margin value to obtain a set of reduction results includes: Based on the margin value corresponding to each scale factor, a floating boundary buffer zone is set for the segmentation interval where the field value is located, and dynamic boundary correction is performed on the field value based on the upper boundary of the buffer zone to obtain the corrected field value; The corrected field value is divided by the segmentation granularity corresponding to each scale factor to obtain a set of ratio values; Perform a consistency check on the ratio values in the set of ratio values at adjacent scales, and perform linear interpolation processing on the deviated ratio values according to the consistency check results to obtain a set of reduction results.

[0041] It should be noted that the set of ratio values refers to the floating-point ratio values obtained by dividing the field value by the segmentation granularities corresponding to multiple scale factors after dynamic boundary correction. This set of ratio values retains the segmentation trend information of the component attributes at multiple scales and is the basic data for subsequent consistency checks and interpolation corrections; during the proportional reduction processing, each item in the set of ratio values will be used in the subsequent interval landing point positioning step to complete the segmentation number attribution of the component at each scale and ensure the numbering stability.

[0042] During the execution of multi-scale discrete bucketing processing, for each field value in the component attribute data, first, based on the segmentation granularity and margin value corresponding to each scale factor, floating boundary buffers for the interval where the field value is located are set, that is, a certain redundant interval is reserved on the basis of the regular segmentation boundary to absorb the boundary errors generated during the modeling accuracy or unit conversion of the field value. This buffer can be formed by expanding the upper boundary value of each segmentation interval by a margin value Δ. For example, when the scale factor is 5 meters and the margin value is 0.5 meters, the original boundaries are [0, 5) and [5, 10), and after expansion, they become [0, 5.5) and [5.5, 10.5). A sample with a field value of 5.3 meters is thus included in the first segmentation instead of being misjudged as the second segmentation.

[0043] Subsequently, for each field value, after the buffer is set and the boundary is corrected, the corrected field value is divided by the segmentation granularity in its corresponding multiple scale factors respectively to obtain multiple ratio values, forming a set of ratio values. For example, if a field value is 4.7 meters, and the corresponding scale factors are 5 meters and 2 meters, and their segmentation granularities are 5 and 2, the calculation results are 0.94 and 2.35, forming the set of ratio values [0.94, 2.35].

[0044] Furthermore, to ensure the consistency and robustness of cross-scale representation, a consistency check of the ratio values in adjacent scales is performed on the set of ratio values. The check rule can be set as follows: if the difference between the ratio values in any two adjacent scales exceeds a preset tolerance threshold (such as 0.5), it is regarded as a deviation in ratio expression, and linear interpolation processing is performed on the deviated ratio value to make it approach the main value in the adjacent scale. For example, when the set of ratio values is [0.94, 2.35], 2.35 can be adjusted to (0.94 × 2 = 1.88), or the mean value processing can be used to obtain 1.65. Finally, the set of ratio values after consistency interpolation correction is obtained as the reduction result set of this field value at multiple scales, providing a basis input structure with balanced accuracy for subsequent segmentation number positioning.

[0045] It should be noted that the set of ratio values refers to the floating-point ratio values obtained by dividing the field value by the segmentation granularity corresponding to multiple scale factors after dynamic boundary correction. This set of ratio values retains the segmentation trend information of the component attributes at multiple scales and is the basic data for subsequent consistency check and interpolation correction; during the execution of ratio reduction processing, each item in the set of ratio values will be used in the subsequent interval landing point positioning step to complete the segmentation number attribution of the component at each scale and ensure the stability of the numbering.

[0046] The methods for performing interval landing point positioning operations on the reduction result set to obtain the corresponding set of segmentation numbers include: Obtain each ratio value in the reduction result set and construct an equally spaced numbering interval based on the scale factor corresponding to each ratio value; Perform a floor operation on each ratio value to determine the segment number of each ratio value in the equally-spaced numbering interval; Combine the segment numbers obtained for each field under all scale factors in the original arrangement order of the fields in the component attribute data to obtain a set of segment numbers.

[0047] Specifically, first obtain each ratio value in the reduction result set, and construct an equally-spaced numbering interval based on the scale factor corresponding to this ratio value. This numbering interval starts from 0 and is divided according to the scale factor to form a sequence of segment numbers numbered 0, 1, 2... For example, when the scale factor is 2, the corresponding equally-spaced numbering interval is [0, 2), [2, 4), [4, 6).

[0048] Subsequently, perform a floor operation on the ratio value to determine its landing number position in the numbering interval. For example, if the ratio value is 4.7 and the corresponding scale factor is 2, then perform 4.7÷1 = 4.7, and take the floor to get 4, which can be located in the 4th numbering interval; if the ratio value is 1.9 and the corresponding scale factor is 1, it is classified into number 1.

[0049] For the segment numbers obtained for each field under multiple scales, perform a combination process according to the original arrangement order of this field in the component attribute data to generate a set of segment numbers. For example, the segment numbers corresponding to the field "length" under scale factors of 5 and 2 are [1, 2], and the segment numbers corresponding to the field "height" under scale factors of 5 and 2 are [0, 1], then the complete set of segment numbers for this component is [1, 2, 0, 1].

[0050] After obtaining the segment numbers of each field value in the component attribute data under each scale, splice the set of segment numbers of each component under all field values and all scales in the original arrangement order of the fields in the component attribute data to obtain the numbering sequence of this component. For example, for a certain component, when the two scale factors of the "length" field are 5 meters and 10 meters, they fall into segment numbers 2 and 1 respectively, and the "material type" field is encoded as category number 3, then the numbering sequence of this component can be [2, 1, 3]. On this basis, arrange the numbering sequences of all components in the order of component numbers in sequence to construct an original bucketing matrix. Each row in the original bucketing matrix corresponds to the segment encoding representation of a component, and each column corresponds to the numbering dimension of a certain field under a certain scale. This matrix serves as the basic structure input for subsequent perturbation analysis and structure encoding processing, laying a coding foundation for the unified structure representation of component attributes.

[0051] The method for obtaining the jump rate of the corresponding numbers of each field value in all components by performing perturbation feedback analysis on the original bucketing matrix in the field dimension direction includes: Based on the field arrangement order in the component attribute data, extract the subsequent number sequence corresponding to the fields in the original bucketing matrix column by column to obtain a set of component number sequences; For each column in the set of component number sequences, perform a sliding window traversal process based on a preset window length, extract adjacent component number pairs within each window, and calculate their number differences to generate a number difference sequence; Compare each difference in the number difference sequence with a perturbation threshold. If the difference is greater than or equal to the preset perturbation threshold, mark it as a jump event, count the total number of jump events, and divide the number of jump events by the total number of component samples to obtain the jump rate of the field value.

[0052] It should be noted that the perturbation feedback analysis is used to evaluate the fluctuation of each field in the original bucketing matrix in the entire component number sequence. By extracting component number pairs through the sliding window method and calculating the jump rate, the higher the jump rate value, the greater the number jump of the field among components; in this embodiment, the field arrangement order of the bucketing coding matrix is rearranged according to the jump rate of the fields, and the fields with lower jump rates are preferentially arranged to enhance the expression stability and continuity of the bucketing coding matrix in the component dimension and improve the structural consistency of the subsequent joint coding mapping process.

[0053] In this embodiment, the process of performing perturbation feedback analysis on the original bucketing matrix in the field dimension direction includes the following steps. First, based on the original arrangement order of the fields in the component attribute data, sequentially extract the number sequences of each column of fields in the original bucketing matrix under all component samples to form a set of component number sequences. Each column of number sequences represents the segmented numbering result of a certain field on all components. For example, if the segmented numbering of a certain field in the first to fifth components is [1, 1, 2, 2, 5], then the component number sequence of this field is the vector composed of these five numbers.

[0054] Subsequently, for each component number sequence, perform a sliding window traversal process based on the preset window length parameter, slide forward one component sample each time, and extract the adjacent component number pairs within each window. For example, under the condition that the window length is 3, the number sequence [1, 1, 2, 2, 5] can generate adjacent number pairs as (1, 1), (1, 2), (2, 2), (2, 5). Then, perform number difference calculation on each adjacent component number pair to obtain a number difference sequence, that is, the difference sequence is [0, 1, 0, 3].

[0055] Next, compare each difference in the above number difference sequence with a preset perturbation threshold (such as the threshold is set to 1). If the difference is greater than or equal to the perturbation threshold, mark it as a jump event. In the above difference sequence [0, 1, 0, 3], the corresponding jump events are two, that is, the number 1 → 2 and the number 2 → 5 are respectively identified as jump events.

[0056] A method for performing joint coding mapping processing on a bucket coding matrix to generate an attribute structure vector includes: Perform an embedding coding operation on each column field value in the bucket coding matrix, and map the segment number to a fixed-length embedding vector based on a preset embedding dictionary; Aggregate the embedding vector set according to the component dimension to obtain an embedding vector sequence corresponding to each component; Concatenate the embedding vector sequences corresponding to each component in the order of the fields to obtain an initial structure matrix; Perform unified mapping on each embedding vector in the initial structure matrix in the component dimension to obtain an attribute structure vector.

[0057] First, for each column field number in the bucket coding matrix, according to the field type and number dimension it represents, call the preset embedding dictionary to perform embedding coding mapping on the segment number, that is, map each integer number to a fixed-length vector representation. For example, the segment number 3 can be mapped to a vector of length 8 [0.1, 0.3, ..., 0.9]. Subsequently, for each component, extract the embedding vectors in sequence according to the field dimension in its number sequence to form an embedding vector sequence corresponding to the component. For example, if a component has numbers 1, 3, and 2 under the three fields of "length", "height", and "material type" respectively, and after embedding coding, three vectors of length d are obtained. Then, perform a concatenation operation on the above embedding vector sequence in the original order of the fields in the component attribute data, that is, connect the embedding vectors end to end in the dimension to form a high-dimensional vector, constituting the initial structure matrix. Each row in this structure matrix corresponds to a component, and each column corresponds to the dimension coordinate value of a certain field in the embedding space. Finally, perform unified mapping processing on each row embedding representation vector in the initial structure matrix, that is, perform normalization, principal component compression, or other projection processing on the row vector to obtain an attribute structure vector with stable structure and consistent scale.

[0058] In this embodiment, by introducing a multi-scale discrete bucketing processing and perturbation feedback analysis mechanism, not only is redundant segment coding based on multiple scale factors and margin values realized at the field level, but also the field order is rearranged based on the jump rate at the component dimension, ensuring the expression stability of the bucket coding matrix. Further, combined with joint coding mapping processing, a unified attribute structure vector is constructed, realizing the structural consistency modeling of heterogeneous BIM component attributes. Compared with the prior art, this processing logic does not need to rely on field naming mapping rules or manual classification operations, has stronger version adaptability and automatic feature expression capabilities, and can significantly improve the component semantic aggregation effect and the accuracy and interpretability of downstream learning tasks, especially in cross-platform and cross-stage modeling data.

[0059] S30: Perform connection pair extraction processing on the component connection data to generate an edge index matrix, perform node attribute reconstruction processing on the attribute structure vector to generate a node feature matrix, use the edge index matrix as the edge set, and use the node feature matrix as the node set to obtain the component graph structure; The method for performing connection pair extraction processing on the component connection data to generate an edge index matrix includes: Obtain a set of component index pairs with connection relationships in the component connection data, where each component index pair consists of two connection fields with an associated relationship, and the connection fields are used to identify two interconnected components; Based on each component index pair, according to the preset field parsing order, respectively use the component index corresponding to the first connection field as the starting node in the edge index matrix, and use the component index corresponding to the second connection field as the target node; Arrange all the component index pairs in sequence according to the field parsing order and write them into a two-dimensional matrix structure to obtain the edge index matrix.

[0060] First, extract a set of component index pairs with connection relationships from the component connection data in the component parameter table. Specifically, the component connection data usually contains several pieces of connection information stored in the form of field pairs, and each record contains two connection fields with an associated relationship. For example, "connected component A index" and "connected component B index" indicate a direct connection relationship between component A and component B. These connection fields can be automatically obtained through the connection table exported by the modeling software, the component topology structure field, or the spatial adjacency attribute, or can be obtained through the connection logic setting method in the structural modeling stage, and are used to identify the structural connection, node association, or construction assembly relationship between components in the actual engineering model.

[0061] Next, for each pair of component connection fields in the above set of component index pairs, determine their arrangement in the edge index matrix according to the preset field parsing order. Specifically, read the component index value corresponding to the first connection field in the component connection data record as the starting node number of the current row in the edge index matrix; read the component index value corresponding to the second connection field as the target node number of the current row. Through this method, the directivity of each connection relationship and the mapping relationship between the source component / target component can be clearly established.

[0062] Finally, write all the above component index pairs into the two-dimensional matrix structure in sequence according to the field parsing order. Each row represents a connection relationship pair, where the first column is the starting node index and the second column is the target node index, thus constructing the edge index matrix.

[0063] It should be noted that in the process of constructing the edge index matrix, if the component connection field itself has directionality, for example, it indicates an explicit connection order or dependency relationship between components, then this connection relationship is modeled as a directed edge pointing from the starting component to the target component; if the connection field has no directionality, for example, it only represents the spatial adjacency relationship between components, then each pair of component index pairs in the edge index matrix can also be represented as an undirected edge. In this embodiment, directed edges are defaultly used for modeling to clarify the connection direction relationship between components.

[0064] The method for performing node attribute reconstruction processing on the attribute structure vector to generate a node feature matrix includes: Write the attribute structure vector corresponding to each component into the corresponding row of the matrix structure according to the arrangement order of the components in the bucket encoding matrix to construct an initial attribute matrix; Perform maximum-minimum normalization processing on each column field in the initial attribute matrix to obtain a normalized attribute matrix; Based on the normalized attribute matrix, calculate the Pearson correlation coefficient between each field and other fields to construct a field correlation matrix; Calculate the correlation coefficients based on the field correlation matrix, and filter out the field pairs with the absolute value of the correlation coefficient greater than the preset correlation threshold to form a field combination set; Perform feature fusion processing on each field group in the field combination set to generate a set of fused field vectors; Concatenate the set of fused field vectors with the set of field vectors in the normalized attribute matrix that have not participated in the fusion to construct a node feature matrix.

[0065] It should be noted that when concatenating the set of fused field vectors with the set of field vectors that have not participated in the fusion, the component order remains the same as that of the initial attribute matrix to ensure that the node features corresponding to each component have a one-to-one correspondence in the graph structure; if some fields are not included in the field combination set, the corresponding field vectors will still be retained as valid inputs in the node feature matrix to achieve the joint expression structure of the fused features and the original features.

[0066] It should be noted that the specific implementation of the maximum - minimum normalization process is as follows: calculate the minimum and maximum values of each column of fields. Taking the minimum value as the origin and the difference between the maximum value and the minimum value as the denominator, use the linear interpolation formula to stretch or compress each field value, so that different fields have a unified reference system in terms of numerical scale, improving the stability and comparability of subsequent correlation analysis. The Pearson correlation coefficient is used to measure whether the numerical change trends of two fields on all components are linearly consistent, and its value range is [-1, 1], where 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no correlation. The field correlation matrix is a symmetric matrix, and its diagonal elements are always 1, and the remaining elements represent the Pearson correlation degree between any two fields.

[0067] The methods for calculating the correlation coefficient based on the field correlation matrix include: ; In the formula, is the correlation coefficient, represents the th value bucket after discretization of field in the normalized attribute matrix, represents the th value bucket after discretization of field in the normalized attribute matrix, represents the joint probability that in all component samples, the value of field is and the value of field is is , represents the marginal probability that the value of field is , is the natural logarithm function with base e.

[0068] It should be noted that before calculating the correlation coefficient, it is necessary to first discretize the field values in the normalized attribute matrix, divide the continuous numerical values into a fixed number of discrete value buckets, and the discrete value buckets are used as the basis for the ] and ] values, so that the joint probability and marginal probability involved in ] can be statistically obtained through the frequencies within the buckets in the component samples, ensuring that the mutual information calculation process has a stable probability basis and is applicable to the measurement of numerical correlation between field pairs.

[0069] ​To achieve the correlation evaluation among multiple fields in the component node attributes, this embodiment introduces a field relationship measurement method based on mutual information. The conceptual basis of this method is as follows: In the BIM data of construction projects, there may be non-explicit redundancy or repeated expressions among various field attributes (such as component materials, floor locations, construction methods). For example, "floor where located" and "construction time period" often show a highly consistent relationship. If these redundant fields are directly retained in the component diagram structure, it will not only introduce information noise but also may lead to a decrease in the learning efficiency of the neural network model. Therefore, a method that can automatically identify and compress such redundant field pairs needs to be constructed.

[0070] Common linear correlation indicators (such as the Pearson coefficient) can only capture the linear trend between fields. However, in this embodiment, the numerical types of component attributes are diverse, and some fields are discrete categorical variables or have a coding structure. The relationships between fields often do not have linear characteristics. For this reason, a correlation evaluation method for field pairs based on mutual information is proposed. By statistically analyzing the joint distribution of the values of each field in the component samples, the amount of information shared by two fields is quantified, so as to discover those field combinations where "the values are different but the changing trends are the same".

[0071] The core idea of mutual information is: If there is a stable correspondence between two fields, the difference between their joint distribution and marginal distribution will be significant, and thus their mutual information value will be high; conversely, if the values of two fields are independent, their mutual information value will approach zero. By calculating the mutual information values between multiple field pairs and screening out the field combinations with mutual information greater than a set threshold, the field sets with redundant expressions can be identified, providing a theoretical basis for subsequent PCA fusion and dimensionality reduction.

[0072] Suppose there are two fields: Field A represents "component area", and Field B represents "construction team", with a total of 4 component samples. The field values are as follows in the table:

[0073] The following distribution situations can be obtained: The value of Field A: A1 appears 2 times, A2 appears 2 times, ; The value of Field B: B1 appears 2 times, B2 appears 2 times ; Joint distribution: .

[0074] Substitute into the correlation coefficient calculation formula: ; The calculation is as follows: For (A1, B1): ; For (A2, B2): ; Therefore, the correlation coefficient is: .

[0075] It should be added that after the construction of the field combination set is completed, in order to further reduce feature redundancy, compress the input dimension and enhance the expression efficiency of component node features, it is necessary to perform feature fusion processing on each field group in the field combination set. Specifically, each field combination can be regarded as a sub-field set composed of several fields that are highly correlated in the normalized attribute matrix. There is a strong consistency in the numerical changes of each field in this set in the component dimension, which belongs to the potential redundant expression dimension. To achieve the purpose of structure compression and semantic condensation, in this embodiment, the principal component analysis (PCA, Principal Component Analysis) method is used to reduce the dimension and fuse the field values in this field combination.

[0076] For example, if a field combination includes field A, field B, and field C, and the Pearson correlation coefficients between any two of them in the field correlation matrix are all higher than 95%, it means that their value distributions on all components are highly consistent. On this basis, PCA processing can be performed on the sub-matrix composed of the values of fields A, B, and C, and the first K principal components (such as K = 1 or K = 2) are extracted as the fusion vector representation of this field combination. This not only retains most of the information in the original field combination, but also compresses the dimension from 3 to K, effectively reducing the overall dimension of the node feature matrix.

[0077] In this embodiment, the edge index matrix is used as the edge set, and the node feature matrix is used as the node set to obtain the component graph structure. Specifically, based on the edge index matrix and the node feature matrix that have been completed in construction, in accordance with the basic form of graph structure modeling, each component is regarded as a node in the graph, and its attribute information is represented by the feature vector of the component in the node feature matrix; the connection relationship between components identified in the component connection data is mapped into a directed or undirected edge in the graph according to the two-dimensional index pair composed of the starting component index and the target component index, thereby constructing the edge set of the graph.

[0078] Among them, the node set is composed of nodes corresponding to all components, and its attributes are provided by the node feature matrix. Each row in the matrix corresponds to a component node, and the row vector is the attribute vector of this node; the edge set is provided by the edge index matrix. Each row in the matrix represents a connection relationship, where the first column is the starting node number and the second column is the target node number, forming a directed edge from the starting node to the target node. If the connection relationship is non-directional, this graph can also be regarded as an undirected graph.

[0079] In the above manner, the constructed component graph structure not only accurately retains the connection relationships between components, but also retains the multi-dimensional attribute expressions of the components themselves, thereby providing a structured graph input basis for subsequent tasks such as aggregation learning, graph feature encoding, similarity analysis, or predictive modeling based on graph neural networks.

[0080] Taking the modeling of graph neural network (GNN) as an example, this component graph structure can be directly used as the input graph. The node feature matrix is input into the graph convolutional layer for node feature propagation and update, and the edge index matrix is used to guide the information transfer path between nodes, thereby realizing the joint modeling of the complex topological structure and semantic attributes between building components.

[0081] S40: Input the component graph structure into the trained neural network model to obtain the building cost; In this embodiment, the training process of the neural network model can be based on historical building project data. First, construct multiple sample component graph structures (including node feature matrices and edge index matrices) with known building costs. Use these component graph structures as inputs and the corresponding true building costs as output labels to form a training dataset. Then, use the graph neural network to extract and learn the features of the component graph structure, and use the mean squared error as the loss function. Use optimization algorithms such as gradient descent for iterative training. Finally, obtain a neural network model that can be used to predict the building cost of new projects. The training method is a prior art, and this embodiment will not elaborate on it too much.

[0082] It can be understood that in actual construction projects, the composition of building costs not only depends on the attribute characteristics of each component itself (such as size, material, construction method, etc.), but also is closely related to the connection relationships between components. The component connection relationships reflect the construction logic, structural dependencies, and process constraints between components. For example: Some high-strength beam components may require supporting columns, and the connection of some special material nodes may introduce additional construction processes, resulting in cost differences.

[0083] Therefore, in this embodiment, a technical path for predicting building costs based on the component graph structure is proposed. Specifically: Component attribute vectors as node features: Each component is represented as a numerical attribute structure vector through the node attribute reconstruction process. This vector comprehensively expresses its attribute characteristics closely related to cost, such as material, size, function, floor where it is located, etc. Component connection relationships as edge structures: By analyzing connection fields such as spatial adjacency, structural assembly, and construction processes in the BIM model, a set of directed edges between components is constructed to form a complete component graph structure. This graph structure not only retains the independent cost contribution characteristics of each component, but also encodes the influence relationships between components in terms of construction paths, structural support, or resource scheduling.

[0084] Embodiment 2 As Figure 2 shown, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the intelligent prediction method for construction project costs based on BIM and machine learning provided by the above-mentioned various methods.

[0085] The detailed description set forth above in connection with the accompanying drawings describes examples and does not represent all examples that may be implemented or fall within the scope of the claims. The terms "example" and "exemplary" when used in this specification mean "serving as an example, instance, or illustration" and do not mean "superior to or better than other examples".

[0086] The phrase "one embodiment" or "an embodiment" recited throughout this specification means that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of the present invention. Therefore, the use of these phrases may refer to more than just one embodiment. Additionally, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0087] It should also be noted that these embodiments may be described as processes depicted as flowcharts, structure diagrams, or block diagrams. Although a flowchart may describe the operations as a sequential process, many of these operations can be performed in parallel or concurrently. Additionally, the order of these operations may be rearranged.

Claims

1. An intelligent prediction method for construction project costs based on BIM and machine learning, characterized in that Including: Obtain S heterogeneous versions of BIM model files, and extract a component parameter table from the S heterogeneous versions of BIM model files based on a preset unique identifier. The component parameter table includes component attribute data and component connection data, where S is an integer greater than 1; Perform multi-scale discrete bucketing processing on the component attribute data to generate a bucketing coding matrix, and perform joint coding mapping processing on the bucketing coding matrix to generate an attribute structure vector; Perform connection pair extraction processing on the component connection data to generate an edge index matrix, perform node attribute reconstruction processing on the attribute structure vector to generate a node feature matrix, use the edge index matrix as the edge set, and use the node feature matrix as the node set to obtain a component graph structure; Input the component graph structure into the trained neural network model to obtain the building cost.

2. The intelligent prediction method for construction project costs based on BIM and machine learning according to claim 1, characterized in that The method for extracting a component parameter table from the S heterogeneous versions of BIM model files based on a preset unique identifier includes: Extract the numerical fields and categorical fields of the components from the S heterogeneous versions of BIM model files; Perform structure separation processing on the numerical fields and categorical fields respectively to obtain a numerical field matrix and a categorical field coding set; Perform multi-scale normalization processing on the numerical field matrix, perform positional embedding coding processing on the categorical field coding set, and merge the normalized numerical field matrix and the coded categorical field coding set to obtain a standard field feature matrix; Construct a structured feature string based on the standard field feature matrix, perform a hash function operation on the structured feature string to generate a unique identifier, and establish a mapping relationship table from the unique identifier to a preset component index; Locate the corresponding components from the BIM model files based on the mapping relationship table, and extract the component attribute data and component connection data to obtain a component parameter table.

3. The intelligent prediction method for construction project costs based on BIM and machine learning according to claim 2, wherein The method for performing multi-scale normalization processing on the numerical field matrix includes: Preset a corresponding normalization scale factor for each field in the numerical field matrix; Use the field value in each field as the numerator and the corresponding normalization scale factor of each field as the denominator to perform proportional scaling processing on the field value; Align the proportionally scaled field values in the original arrangement order of the fields in each field, and splice the aligned field values horizontally to form an extended field matrix; Perform maximum-minimum normalization processing on each column in the extended field matrix, and use the normalized extended field matrix as the normalized numerical field matrix.

4. The intelligent prediction method for construction project costs based on BIM and machine learning according to claim 1, characterized in that The method for performing multi-scale discrete bucketing processing on the component attribute data to generate a bucketing coding matrix includes: Set multiple scale factors for each field value in the component attribute data, where each scale factor corresponds to a segmentation granularity and a margin value; Based on the segmentation granularity and the margin value, perform proportional reduction processing on each field value to obtain a reduction result set, and perform interval landing point positioning operation on the reduction result set to obtain a corresponding segmentation number set; Splice the segmentation number sets of each component under all field values and all scales in the original arrangement order of the field values in the component attribute data to generate a number sequence of the component, and summarize the number sequences of all components to form an original bucketing matrix; Perform perturbation feedback analysis on the original bucket matrix in the field dimension direction to obtain the jump rate of the corresponding numbers of each field value in all components, and re-arrange the arrangement order of the field dimension according to the jump rate to generate a bucket coding matrix.

5. The intelligent prediction method for construction project cost based on BIM and machine learning according to claim 4, wherein The method for performing proportional reduction processing on each field value based on the segmentation granularity and the margin value to obtain a set of reduction results includes: Based on the margin value corresponding to each scale factor, set a floating boundary buffer for the segmentation interval where the field value is located, and perform dynamic boundary correction on the field value based on the upper boundary of the buffer to obtain the corrected field value; Perform division calculation on the corrected field value and the segmentation granularity corresponding to each scale factor to obtain a set of proportional values; Perform consistency check on the proportional values at adjacent scales for the set of proportional values, and perform linear interpolation processing on the deviated proportional values according to the consistency check results to obtain a set of reduction results.

6. The intelligent prediction method for construction project cost based on BIM and machine learning according to claim 5, characterized in that The method for performing interval landing point positioning operation on the set of reduction results to obtain a corresponding set of segmentation numbers includes: Obtain each proportional value in the set of reduction results, and construct an equidistant numbering interval based on the scale factor corresponding to each proportional value; Perform floor operation on each proportional value to determine the segmentation number of each proportional value in the equidistant numbering interval; Combine the segmentation numbers obtained by each field under all scale factors according to the original arrangement order of the fields in the component attribute data to obtain a set of segmentation numbers.

7. The intelligent prediction method for construction project cost based on BIM and machine learning according to claim 6, characterized in that The method for performing perturbation feedback analysis on the original bucket matrix in the field dimension direction to obtain the jump rate of the corresponding numbers of each field value in all components includes: Based on the field arrangement order in the component attribute data, extract the number sequence corresponding to the field in the original bucket matrix column by column to obtain a set of component number sequences; For each column in the set of component number sequences, perform sliding window traversal processing based on a preset window length, extract adjacent component number pairs within each window, and calculate their number differences to generate a sequence of number differences; Compare each difference in the sequence of number differences with the perturbation threshold. If the difference is greater than or equal to the preset perturbation threshold, mark it as a jump event, count the total number of jump events, and divide the number of jump events by the total number of component samples to obtain the jump rate of the field value.

8. The intelligent prediction method for construction project costs based on BIM and machine learning according to claim 7, characterized in that, The method for performing joint coding mapping processing on the bucket coding matrix to generate an attribute structure vector includes: Perform embedding coding operation on each column of field values in the bucket coding matrix, and map the segmentation number to a fixed-length embedding vector based on a preset embedding dictionary; Perform aggregation processing on the set of embedding vectors in the component dimension to obtain an embedding vector sequence corresponding to each component; Concatenate the embedding vector sequences corresponding to each component in the field order to obtain an initial structure matrix; Perform unified mapping on the embedding vectors in the initial structure matrix in the component dimension to obtain an attribute structure vector.

9. The intelligent prediction method for construction project costs based on BIM and machine learning according to claim 1, characterized in that The method for performing connection pair extraction processing on component connection data to generate an edge index matrix includes: Obtain a set of component index pairs with connection relationships in the component connection data, where each component index pair consists of two connection fields with an associated relationship, and the connection fields are used to identify two mutually connected components; Based on each component index pair, in accordance with the preset field parsing order, respectively take the component index corresponding to the first connection field as the starting node in the edge index matrix, and take the component index corresponding to the second connection field as the target node; Arrange all component index pairs in sequence according to the field parsing order and write them into a two-dimensional matrix structure to obtain an edge index matrix.

10. The intelligent prediction method for construction project costs based on BIM and machine learning according to claim 1, characterized in that, The method for performing node attribute reconstruction processing on the attribute structure vector to generate a node feature matrix includes: Write the attribute structure vector corresponding to each component into the corresponding row of the matrix structure according to the arrangement order of the components in the bucket encoding matrix to construct an initial attribute matrix; Perform maximum-minimum normalization processing on each column field in the initial attribute matrix to obtain a normalized attribute matrix; Based on the normalized attribute matrix, calculate the Pearson correlation coefficient between each field and other fields to construct a field correlation matrix; Calculate the correlation coefficient based on the field correlation matrix, and filter out the field pairs whose absolute value of the correlation coefficient is greater than the preset correlation threshold to form a field combination set; Perform feature fusion processing on each field group in the field combination set to generate a set of fused field vectors; Concatenate the set of fused field vectors with the set of field vectors in the normalized attribute matrix that did not participate in the fusion to construct a node feature matrix.

Images