A BIM-based construction project cost evaluation management method and system

Abstract

The application discloses a kind of based on BIM's construction engineering project cost evaluation management method and system, by obtaining construction engineering characteristic data, establish multi-scale graph time sequence fusion prediction model, spatial relationship channel uses HGAT hierarchical graph attention network to layer-by-layer aggregation spatial feature, time evolution channel utilizes TCN time sequence convolution network to capture the time feature of cost accumulation, by cross-modal attention mechanism adjustment spatial feature and time feature weight;The construction engineering characteristic data is input into multi-scale graph time sequence fusion prediction model;Based on the engineering cost prediction value is judged, when engineering cost actual cost is greater than engineering cost prediction value, trigger attribution mechanism based on SHAP value and counterfactual reasoning, generate quantifiable optimization scheme, replace traditional artificial accounting and experience judgment, greatly reduce manpower investment, shorten cost evaluation cycle.

Description

Technical Field

[0001] This invention relates to the field of construction project management technology, and in particular to a BIM-based method and system for construction project cost assessment and management. Background Technology

[0002] Cost assessment and management for construction projects are central to project cost control, and their accuracy directly impacts investment decisions and profits. Traditional cost assessment methods heavily rely on historical quotas and manual experience, resulting in drawbacks such as lagging data updates, insufficient consideration of individual differences, and weak dynamic adjustment capabilities. Existing BIM-based cost management methods, based on historical data, cannot reflect real-time fluctuations in market prices for labor, materials, and machinery, leading to a disconnect between estimates and actual market conditions. Furthermore, the difficulty in quantifying the impact of non-linear, high-dimensional factors such as geological conditions, construction environment, climate, and supply chain risks on costs further hinders the efficiency of construction project cost assessment and management. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a BIM-based method and system for cost assessment and management of building construction projects.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned BIM-based construction project cost assessment and management method: The BIM model data is acquired, and the 3D mesh of the components in the BIM model data is converted into a point cloud sequence with normal vectors. Feature extraction is performed through the PointNet++ network, and the component type, material and construction technology are hierarchically encoded to obtain the building engineering feature data. A multi-scale graph temporal fusion prediction model is established. The spatial relationship channel uses the HGAT hierarchical graph attention network to aggregate spatial features layer by layer, while the temporal evolution channel uses the TCN temporal convolutional network to capture the temporal features of cost accumulation. The weights of spatial and temporal features are adjusted through a cross-modal attention mechanism. The construction project feature data is input into the multi-scale map time series fusion prediction model, and the predicted project cost is output. Based on the predicted project cost, when the actual project cost exceeds the predicted cost, an attribution mechanism based on SHAP value and counterfactual reasoning is triggered to generate a quantifiable optimization solution.

[0005] Furthermore, in the aforementioned BIM-based construction project cost assessment and management method, the step of acquiring BIM model data involves converting the 3D mesh of the BIM model data into a point cloud sequence with normal vectors, extracting features using a PointNet++ network, and hierarchically encoding the component type, materials, and construction techniques to obtain construction project feature data, including: Obtain BIM model data, including at least the 3D mesh of components, component type, materials, and construction process. Then, filter, clean, and standardize the collected BIM model data in sequence to obtain preprocessed BIM model data. A mesh segmentation algorithm is used to divide the 3D mesh of components in the standardized preprocessed BIM model data, dividing the 3D entity of each component into a uniform triangular mesh to obtain mesh component data; The triangular mesh of each component in the mesh component data is uniformly sampled to generate the initial point cloud data of the component; the least squares method is used to fit the plane where the sampling points are located, and the normal vector of each sampling point is calculated to obtain the point cloud sequence with normal vector.

[0006] Furthermore, in the aforementioned BIM-based construction project cost assessment and management method, the step of acquiring BIM model data involves converting the 3D mesh of the BIM model data into a point cloud sequence with normal vectors, extracting features using a PointNet++ network, and hierarchically encoding the component type, materials, and construction techniques to obtain construction project feature data, including: A PointNet++ network was built with an input layer dimension of 4, a sampling layer using the FPS farthest point sampling algorithm, a feature extraction layer divided into 3 scales, each scale containing 2 convolutional layers, an activation function using the ReLU function, and an output layer dimension of 128. The point cloud sequence with normal vectors is input into the PointNet++ network. Local geometric features of the component surface are extracted through local scale convolutional layers, and local features are aggregated into regional features through regional scale convolutional layers. The global-scale convolutional layer fuses regional features and outputs a global feature vector of the component.

[0007] Furthermore, in the aforementioned BIM-based construction project cost assessment and management method, the establishment of a multi-scale graph-temporal fusion prediction model, the spatial relationship channel uses an HGAT hierarchical graph attention network to aggregate spatial features layer by layer, and the temporal evolution channel uses a TCN temporal convolutional network to capture the temporal features of cost accumulation. The weights of spatial and temporal features are adjusted through a cross-modal attention mechanism, including: Using the components in the architectural engineering feature data as nodes and the spatial relationships between components as edges, a spatial relationship graph of architectural engineering is constructed. The node features are architectural engineering feature data, and the edge weights are set according to the strength of the spatial relationship between components. A hierarchical graph attention network (HGAT) was constructed, consisting of three layers: an attention mechanism layer, a convolutional layer, and an activation layer. The attention mechanism layer employed multi-head attention, and the convolutional layers used 3×3 kernels and LeakyReLU activation functions. The spatial relationship graph of the building project was input into the HGAT network, and the spatial features were output.

[0008] Furthermore, in the aforementioned BIM-based construction project cost assessment and management method, the establishment of a multi-scale graph-temporal fusion prediction model, the spatial relationship channel uses an HGAT hierarchical graph attention network to aggregate spatial features layer by layer, and the temporal evolution channel uses a TCN temporal convolutional network to capture the temporal features of cost accumulation. The weights of spatial and temporal features are adjusted through a cross-modal attention mechanism, including: A TCN temporal convolutional network was constructed with a 4-layer structure. The convolutional layers used dilated convolution with a kernel size of 3×3, the pooling layers used max pooling, and the activation function was ReLU. The construction engineering feature data is input into the TCN temporal convolutional network, and the cost accumulation features at different time nodes are captured through dilated convolutional layers. The captured temporal features are reduced in dimensionality through a pooling layer to remove redundant temporal information; finally, the temporal features are output through an output layer.

[0009] Furthermore, in the aforementioned BIM-based construction project cost assessment and management method, the establishment of a multi-scale graph-temporal fusion prediction model, the spatial relationship channel uses an HGAT hierarchical graph attention network to aggregate spatial features layer by layer, and the temporal evolution channel uses a TCN temporal convolutional network to capture the temporal features of cost accumulation. The weights of spatial and temporal features are adjusted through a cross-modal attention mechanism, including: Design a cross-modal attention mechanism, with the input being the global spatial feature vector output by the HGAT network and the global temporal feature vector output by the TCN network, construct an attention weight calculation function, and calculate the weight coefficients of spatial features and temporal features based on feature similarity; Based on the calculated weight coefficients, the global spatial feature vector and the global temporal feature vector are weighted and fused. When the feature weight is lower than the threshold, the weight coefficients are adjusted.

[0010] Furthermore, in the aforementioned BIM-based construction project cost assessment and management method, the judgment based on the project cost forecast value, when the actual project cost exceeds the forecast value, triggers an attribution mechanism based on SHAP value and counterfactual reasoning to generate a quantifiable optimization solution, including: Input the architectural engineering feature data of the current project into the trained multi-scale graph time series fusion prediction model, and calculate the SHAP value of each feature; We selected the top 10 features with positive SHAP values ​​and large absolute values, and analyzed the specific reasons for cost overruns caused by each core feature in light of the actual project situation.

[0011] Furthermore, in a BIM-based construction project cost assessment and management system, the construction project cost assessment and management system includes the following modules: The BIM data acquisition module is used to acquire BIM model data, convert the 3D mesh of the components in the BIM model data into a point cloud sequence with normal vectors, extract features through the PointNet++ network, and hierarchically encode the component type, material and construction technology to obtain building engineering feature data. The prediction model building module is used to build a multi-scale graph temporal fusion prediction model. The spatial relationship channel uses the HGAT hierarchical graph attention network to aggregate spatial features layer by layer, while the temporal evolution channel uses the TCN temporal convolutional network to capture the time features of cost accumulation. The weights of spatial and temporal features are adjusted through a cross-modal attention mechanism. The engineering cost assessment module is used to input the architectural engineering characteristic data into the multi-scale map time series fusion prediction model and output the predicted engineering cost value. The cost assessment and management module is used to make judgments based on the predicted project cost. When the actual project cost is greater than the predicted project cost, it triggers an attribution mechanism based on SHAP value and counterfactual reasoning to generate a quantifiable optimization solution.

[0012] Furthermore, in a BIM-based construction project cost assessment and management system, the prediction model building module includes the following sub-modules: A submodule is constructed to design a cross-modal attention mechanism. The input is the global spatial feature vector output by the HGAT network and the global temporal feature vector output by the TCN network. An attention weight calculation function is constructed to calculate the weight coefficients of spatial features and temporal features based on feature similarity. The fusion submodule is used to perform weighted fusion of the global spatial feature vector and the global temporal feature vector based on the calculated weight coefficients, and adjust the weight coefficients when the feature weights are lower than the threshold.

[0013] Furthermore, in a BIM-based construction project cost assessment management system, the cost assessment management module includes the following sub-modules: The calculation submodule is used to input the architectural engineering feature data of the current project into the trained multi-scale graph time series fusion prediction model and calculate the SHAP value of each feature. The filtering submodule is used to filter out the top 10 features with positive SHAP values ​​and large absolute values, and analyze the specific reasons for cost overruns caused by each core feature in conjunction with the actual situation of the project.

[0014] Its beneficial effects lie in the precise extraction of BIM component features through the PointNet++ network, combined with the HGAT and TCN dual-channel fusion model, taking into account both spatial correlation and temporal cumulative features. Weights are optimized through a cross-modal attention mechanism, significantly reducing prediction errors and ensuring that the deviation rate between predicted values ​​and actual costs is controlled within 5%. It captures cost changes in real time, quickly triggering an attribution mechanism when cost overruns occur, avoiding cost waste caused by lagging management. The optimization solutions are scientifically sound and implementable. Based on SHAP values, it accurately locates the core causes of cost overruns, and verifies the accuracy of attribution through counterfactual reasoning. The generated optimization solutions clearly quantify objectives, responsible parties, and implementation paths, directly guiding on-site construction and management. It improves work efficiency by standardizing BIM data processing and automating algorithm operation, replacing traditional manual calculations and experience-based judgments, significantly reducing manpower input, shortening cost assessment cycles, and standardizing operating procedures. This provides standardized and replicable technical support for project cost management, helping projects reduce costs, increase efficiency, and improve management levels. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of a BIM-based construction project cost assessment and management method according to the present invention; Figure 2 This is a schematic diagram of a second embodiment of a BIM-based construction project cost assessment and management method according to the present invention; Figure 3 This is a schematic diagram of the first embodiment of a BIM-based building construction project cost assessment and management system according to the present 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] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a BIM-based method for cost assessment and management of building construction projects includes the following steps: Step 101: Obtain BIM model data, convert the 3D mesh of the components in the BIM model data into a point cloud sequence with normal vectors, extract features through PointNet++ network, and hierarchically encode the component type, material and construction process to obtain building engineering feature data. Specifically, in this embodiment, the BIM model data obtained includes at least the three-dimensional mesh of the components, component type, materials and construction technology. The collected BIM model data is sequentially screened, cleaned and standardized to obtain preprocessed BIM model data. A mesh segmentation algorithm is used to partition the 3D mesh of components in the standardized preprocessed BIM model data, dividing the 3D entity of each component into a uniform triangular mesh to obtain mesh component data. Uniform sampling is performed on the triangular mesh of each component in the mesh component data to generate initial point cloud data for the component. The least squares method is used to fit the plane where the sampling points are located, and the normal vector of each sampling point is calculated to obtain a point cloud sequence with normal vectors. A PointNet++ network is established with an input layer dimension of 4. The sampling layer uses the FPS (Farthest Point Sampling) algorithm, and the feature extraction layer is divided into 3 scales, each containing 2 convolutional layers. The ReLU activation function is used, and the output layer dimension is 128. The point cloud sequence with normal vectors is input into the PointNet++ network. Local geometric features of the component surface are extracted through local scale convolutional layers, and local features are aggregated into regional features through regional scale convolutional layers. Finally, regional features are fused through global scale convolutional layers to output the global feature vector of the component.

[0020] This process involves batch exporting complete BIM model data from BIM modeling software used in the design phase of architectural engineering projects. The data collection covers all disciplines, including architecture, structure, MEP (Mechanical, Electrical, and Plumbing), and decoration, ensuring no professional omissions or missing components. Redundant and invalid data are removed from the BIM model; key component attributes are supplemented and improved, including but not limited to component name, specifications, material type, installation location, construction batch, and design parameters, ensuring the completeness and traceability of each component's attributes; and abnormal data is corrected. The exported BIM model data is then uniformly converted into a processable, universal format.

[0021] A mesh partitioning algorithm, such as region growing, is used to perform 3D meshing on the standardized BIM model components. Each component's 3D entity is divided into uniform triangular or quadrilateral meshes. The mesh density is dynamically adjusted according to the component's complexity. For complex components such as irregular beams and columns, the mesh side length is set to 5-10 cm; for simple components such as walls and floor slabs, the mesh side length is set to 10-20 cm. This ensures that the mesh fully preserves the component's geometric features while avoiding data redundancy caused by excessive meshes. Uniform sampling is performed on the 3D mesh of each component, with a sampling density controlled at 1-2 sampling points per square centimeter, generating initial point cloud data for the component. The least squares method is used to fit the plane containing the sampling points, calculating the normal vector for each sampling point, including the normal vector direction and magnitude, clarifying the spatial orientation characteristics of the component's surface. The coordinates of the sampling points are bound to the corresponding normal vectors, forming a point cloud sequence with normal vectors. Each point cloud sequence corresponds to one BIM component. Outlier removal algorithms, such as statistical filtering algorithms, are used to remove outliers in the point cloud sequence caused by sampling errors. A confidence interval of 95% is set, and sampling points outside the interval are identified as outliers. For sampling points with missing normal vectors, neighboring point normal vector interpolation is used to supplement them, ensuring the integrity and accuracy of each point cloud sequence.

[0022] A PointNet++ network model was built, with the following network parameters: input layer dimension of 4 (3 coordinate dimensions + 1 normal vector dimension), sampling layer using FPS (Farthest Point Sampling), feature extraction layer divided into 3 scales, each scale containing 2 convolutional layers, activation function using ReLU, and output layer dimension of 128. The optimized point cloud sequence was input into the PointNet++ network. First, local scale convolutional layers extracted local geometric features of the component surface; then, region scale convolutional layers aggregated local features into region features; finally, global scale convolutional layers fused the region features, outputting a global feature vector of the component, achieving comprehensive extraction of the component's geometric and spatial features. L1 regularization was used to filter the extracted global feature vectors, removing redundant and invalid features and retaining core features related to cost. The filtered feature vectors were then normalized to avoid the impact of differences in the magnitude of features across different dimensions on subsequent encoding performance.

[0023] A three-tiered hierarchical coding system was established: Level 1 codes corresponded to component types; Level 2 codes corresponded to component materials; and Level 3 codes corresponded to construction techniques. For each component, Level 1, Level 2, and Level 3 codes were assigned sequentially based on its attribute information, ensuring that each component's code was unique and traceable. The codes were converted into One-Hot vectors and concatenated with normalized feature vectors extracted by the PointNet++ network to form the final building engineering feature data, providing standardized input data for subsequent cost prediction. A coding verification mechanism was established to verify each assigned code, avoiding code duplication and errors. If component attributes changed in the BIM model, the corresponding codes and One-Hot vectors were updated promptly to ensure consistency between the building engineering feature data and the BIM model data.

[0024] Step 102: Establish a multi-scale graph temporal fusion prediction model. The spatial relationship channel uses the HGAT hierarchical graph attention network to aggregate spatial features layer by layer. The temporal evolution channel uses the TCN temporal convolutional network to capture the time features of cost accumulation. The weights of spatial features and temporal features are adjusted through a cross-modal attention mechanism. Specifically, in this embodiment, components in the building engineering feature data are used as nodes, and the spatial relationships between components are used as edges to construct a building engineering spatial relationship graph. The node features are the building engineering feature data, and the edge weights are set according to the strength of the spatial relationship between components. An HGAT hierarchical graph attention network is built, with three layers. Each layer contains an attention mechanism layer, a convolutional layer, and an activation layer. The attention mechanism layer uses multi-head attention, the convolutional layer has a 3×3 kernel size, and the activation function is LeakyReLU. The building engineering spatial relationship graph is input into the HGAT network, and the spatial features are output. A TCN temporal convolutional network is built, with four layers. The convolutional layers use dilated convolution with a 3×3 kernel size, the pooling layers use max pooling, and the activation function is ReLU. The building engineering feature data is input into the TCN temporal convolutional network. The dilated convolutional layer captures the cost accumulation features at different time nodes. The pooling layer reduces the dimensionality of the captured temporal features and removes redundant temporal information. Finally, the output layer outputs the temporal features. A cross-modal attention mechanism is designed, with the input being the global spatial feature vector output by the HGAT network and the global temporal feature vector output by the TCN network. An attention weight calculation function is constructed, and the weight coefficients of spatial features and temporal features are calculated based on feature similarity. According to the calculated weight coefficients, the global spatial feature vector and the global temporal feature vector are weighted and fused. When the feature weight is lower than the threshold, the weight coefficients are adjusted.

[0025] Using components from the building engineering feature data as nodes and the spatial relationships between components, such as connection relationships, adjacency relationships, and force transmission relationships, as edges, a spatial relationship graph of the building engineering is constructed. The node features are the building engineering feature data output in step 1, and the edge weights are set according to the strength of the spatial relationship between components. For example, the edge weight of directly connected components is 0.8-1.0, the edge weight of adjacent but not connected components is 0.3-0.7, and the edge weight of unrelated components is 0. An HGAT hierarchical graph attention network is constructed, with 3 layers corresponding to the local space, regional space, and overall space of the components. Each layer contains an attention mechanism layer, a convolutional layer, and an activation layer. The attention mechanism layer adopts multi-head attention with 8 heads, the convolutional layer has a 3×3 kernel size, and the activation function is the LeakyReLU function. The output layer dimension is consistent with the dimension of the building engineering feature data, which is 136 dimensions. The constructed spatial relationship graph is input into the HGAT network. The first layer, the local spatial layer, focuses on the spatial relationship between a single component and its adjacent components. It assigns weights to adjacent components through an attention mechanism, aggregates local spatial features, and outputs the local spatial feature vector of the component. The second layer, the regional spatial layer, groups the local spatial feature vectors according to building regions, such as basement, ground floor, and standard floor, aggregates the spatial features of all components within the region, and outputs the regional spatial feature vector. The third layer, the global spatial layer, performs global aggregation of the spatial feature vectors of all regions, captures the spatial layout features and component relationship features of the entire building project, and outputs the global spatial feature vector, thus completing the hierarchical extraction and aggregation of spatial features.

[0026] Collect time-series cost data for the entire lifecycle of a construction project, covering the cumulative cost data of the design, construction preparation, construction, and completion acceptance stages. Divide the time series by time nodes, such as monthly or quarterly. The time series step size is set according to the project duration: 1 month for a duration ≤ 1 year, and 3 months for a duration > 1 year. Normalize the time series data and map it to the [0,1] interval. Associate it with the construction project feature data output in step 1 to form time series-feature association data, which serves as the input data for the TCN network. A TCN temporal convolutional network was constructed with a four-layer structure: an input layer, a convolutional layer, a pooling layer, and an output layer. The input layer has a dimension of 137: the dimension of the architectural engineering feature data plus the dimension of the temporal data. The convolutional layer uses dilated convolution with dilation coefficients of 1, 2, 4, and 8, and the kernel size is 3×3. The pooling layer uses max pooling with a kernel size of 2×2. The activation function is ReLU. The output layer has a dimension of 136, consistent with the dimension of the global spatial feature vector output by the HGAT network. The time-series-feature correlation data is input into the TCN network. Through dilated convolutional layers, the local receptive field limitation of traditional convolution is overcome, capturing cost accumulation features at different time points, such as material cost growth, labor cost accumulation, and machinery cost changes during the construction phase. Pooling layers are used to reduce the dimensionality of the captured time-series features and remove redundant time-series information. Finally, the output layer outputs a global time-series feature vector, achieving accurate capture of the time evolution law of cost accumulation and time-series correlation features. The Dropout algorithm is adopted with a dropout rate of 0.3 to avoid network overfitting and improve the stability of time-series feature extraction.

[0027] Design a cross-modal attention mechanism module, with the input being the global spatial feature vector output by the HGAT network and the global temporal feature vector output by the TCN network; construct an attention weight calculation function, which calculates the weight coefficients of spatial features and temporal features based on feature similarity, with the sum of the weight coefficients being 1, and the formula is: ω1=cos(S,T) / (cos(S,T)+cos(T,S)), ω2=1-ω1, where ω1 is the spatial feature weight, ω2 is the temporal feature weight, S is the global spatial feature vector, T is the global temporal feature vector, and cos(·) is the cosine similarity function. Based on the calculated weight coefficients, the global spatial feature vector and the global temporal feature vector are weighted and fused using the formula: F = ω1 × S + ω2 × T, where F is the fused global feature vector with a dimension of 136. A dynamic adjustment threshold is set: 0.2 ≤ ω1 ≤ 0.8, 0.2 ≤ ω2 ≤ 0.8. When the weight of a certain feature is lower than the threshold, the weight coefficient is automatically adjusted to ensure a balanced fusion of spatial and temporal features and avoid a single feature dominating the prediction result. The fused global feature vector serves as the core feature input of the multi-scale graph-temporal fusion prediction model and is used for subsequent cost prediction.

[0028] Step 103: Input the building engineering feature data into the multi-scale map time series fusion prediction model and output the predicted engineering cost; Specifically, in this embodiment, the preprocessed construction project feature data is batch-input into the trained multi-scale graph time-series fusion prediction model. Prediction parameters are set with a batch size of 32 and the number of prediction iterations set to 10 to ensure the stability of the prediction process. Internally, the model re-aggregates the spatial features of the current project through the HGAT network and captures the cumulative time-series cost features of the current project through the TCN network. After adjusting the weights via a cross-modal attention mechanism, the fused global feature vector is used for prediction calculation, outputting the prediction results of 10 iterations, i.e., 10 predicted construction cost values.

[0029] Step 104: Based on the predicted project cost, when the actual project cost is greater than the predicted project cost, an attribution mechanism based on SHAP value and counterfactual reasoning is triggered to generate a quantifiable optimization solution.

[0030] Specifically, in this embodiment, the architectural engineering feature data of the current project is input into the trained multi-scale graph time series fusion prediction model to calculate the SHAP value of each feature; the top 10 features with positive SHAP values ​​and large absolute values ​​are selected, and the specific reasons for cost overruns caused by each core feature are analyzed in conjunction with the actual situation of the project.

[0031] Real-time collection of actual cost data for current construction projects, covering all cost items including material costs, labor costs, machinery costs, management costs, regulatory fees, and taxes. Data is aggregated according to time nodes, consistent with the time-series data step size, ensuring the completeness and accuracy of the actual cost data. Supporting documents such as purchase invoices for material costs, payroll records for labor costs, and lease agreements for machinery costs are required. The aggregated actual cost data is compared with the project cost forecast output in step 3 to calculate the cost deviation rate using the formula: Deviation Rate = (Actual Cost - Forecast Value) / Forecast Value × 100%. A judgment threshold is set: when the deviation rate is ≤ 5%, i.e., actual cost ≤ forecast value × 1.05, the cost control is deemed satisfactory, and no optimization mechanism is triggered; the actual cost change is continuously monitored. When the deviation rate is > 5%, i.e., actual cost > forecast value × 1.05, the cost is deemed excessive, and an attribution mechanism based on SHAP value and counterfactual reasoning is immediately triggered.

[0032] The current project's architectural feature data is input into the trained multi-scale graph time-series fusion prediction model. The SHAP value is calculated for each feature, including component type, material, construction technology, spatial features, and temporal features. A positive SHAP value indicates that the feature leads to increased costs, while a negative SHAP value indicates that the feature leads to decreased costs. The larger the absolute value of the SHAP value, the stronger the influence of the feature on the cost. All features are ranked, and the top 10 features with positive and large absolute SHAP values ​​are selected as the core features leading to cost overruns. Based on the actual project situation, the specific reasons for cost overruns caused by each core feature are analyzed.

[0033] For the core cost overrun characteristics identified, counterfactual scenarios are constructed. For example, assuming the high SHAP value characteristic of cast-in-place concrete walls in the current project is replaced with precast concrete walls, the cost prediction value under the counterfactual scenario is calculated by inputting it into the prediction model. The difference between the original prediction value and the counterfactual scenario prediction value is compared to verify the degree of influence of this characteristic on cost overruns and ensure the accuracy of the attribution results. Based on the counterfactual reasoning results, the optimization direction for each core cost overrun characteristic is clarified. The optimization direction must be relevant to the actual project and be operable. For example, for material-related cost overruns, the optimization direction is to replace materials with more cost-effective ones and reduce material wastage rate; for construction process-related cost overruns, the optimization direction is to optimize construction processes and improve construction efficiency; for spatial layout-related cost overruns, the optimization direction is to adjust the component layout and reduce construction difficulty. Based on the optimization direction, quantifiable optimization plans are formulated. Each plan must clearly define the optimization measures, responsible parties, implementation time, and quantifiable objectives.

[0034] Its beneficial effects lie in the precise extraction of BIM component features through the PointNet++ network, combined with the HGAT and TCN dual-channel fusion model, taking into account both spatial correlation and temporal cumulative features. Weights are optimized through a cross-modal attention mechanism, significantly reducing prediction errors and ensuring that the deviation rate between predicted values ​​and actual costs is controlled within 5%. It captures cost changes in real time, quickly triggering an attribution mechanism when cost overruns occur, avoiding cost waste caused by lagging management. The optimization solutions are scientifically sound and implementable. Based on SHAP values, it accurately locates the core causes of cost overruns, and verifies the accuracy of attribution through counterfactual reasoning. The generated optimization solutions clearly quantify objectives, responsible parties, and implementation paths, directly guiding on-site construction and management. It improves work efficiency by standardizing BIM data processing and automating algorithm operation, replacing traditional manual calculations and experience-based judgments, significantly reducing manpower input, shortening cost assessment cycles, and standardizing operating procedures. This provides standardized and replicable technical support for project cost management, helping projects reduce costs, increase efficiency, and improve management levels.

[0035] Please see Figure 2In a BIM-based method for cost assessment and management of building construction projects, the following steps are taken to obtain building construction feature data: acquiring BIM model data, converting the 3D mesh of components in the BIM model data into a point cloud sequence with normal vectors, extracting features using a PointNet++ network, and hierarchically encoding component type, material, and construction process. Step 201: Obtain BIM model data, including at least the 3D mesh of components, component type, materials and construction technology. The collected BIM model data is then filtered, cleaned and standardized in sequence to obtain preprocessed BIM model data. Step 202: Use a mesh segmentation algorithm to divide the 3D mesh of the components in the standardized preprocessed BIM model data into uniform triangular meshes to obtain mesh component data. Step 203: Uniformly sample the triangular mesh of each component in the mesh component data to generate the initial point cloud data of the component; use the least squares method to fit the plane where the sampling points are located, calculate the normal vector of each sampling point, and obtain the point cloud sequence with normal vector.

[0036] The above describes an embodiment of the BIM-based construction project cost assessment and management method of the present invention. Please refer to [link / reference]. Figure 3 In a BIM-based construction project cost assessment management system, the system includes the following modules: The BIM data acquisition module is used to acquire BIM model data, convert the 3D mesh of the components in the BIM model data into a point cloud sequence with normal vectors, extract features through the PointNet++ network, and hierarchically encode the component type, material and construction technology to obtain building engineering feature data. The prediction model building module is used to build a multi-scale graph temporal fusion prediction model. The spatial relationship channel uses the HGAT hierarchical graph attention network to aggregate spatial features layer by layer, while the temporal evolution channel uses the TCN temporal convolutional network to capture the time features of cost accumulation. The weights of spatial and temporal features are adjusted through a cross-modal attention mechanism. The engineering cost assessment module is used to input the characteristic data of building engineering into the multi-scale map time series fusion prediction model and output the predicted value of engineering cost. The cost assessment and management module is used to make judgments based on the predicted cost of the project. When the actual cost of the project is greater than the predicted cost, it triggers an attribution mechanism based on SHAP value and counterfactual reasoning to generate a quantifiable optimization solution.

[0037] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and equivalents.

Claims

1. A BIM-based method for cost assessment and management of building construction projects, characterized in that, The aforementioned method for managing the cost assessment of construction projects includes the following steps: The BIM model data is acquired, and the 3D mesh of the components in the BIM model data is converted into a point cloud sequence with normal vectors. Feature extraction is performed through the PointNet++ network, and the component type, material and construction technology are hierarchically encoded to obtain the building engineering feature data. A multi-scale graph-temporal fusion prediction model is established. The spatial relationship channel uses an HGAT hierarchical graph attention network to aggregate spatial features layer by layer, while the temporal evolution channel utilizes a TCN temporal convolutional network to capture the temporal features of cost accumulation. A cross-modal attention mechanism is used to adjust the weights of spatial and temporal features. This includes constructing a spatial relationship graph of building engineering features, using components as nodes and spatial relationships between components as edges. Node features are the building engineering feature data, and edge weights are set according to the strength of spatial relationships between components. An HGAT hierarchical graph attention network is built, with three layers: an attention mechanism layer, a convolutional layer, and an activation layer. The attention mechanism layer uses multi-head attention, and the convolutional layer has a 3×3 kernel size and uses the LeakyReLU activation function. The building engineering spatial relationship graph is input into the HGAT network, and spatial features are output. A TCN temporal convolutional network is constructed with a 4-layer structure. The convolutional layers employ dilated convolutions with 3×3 kernels, and the pooling layers use max pooling with ReLU activation. Construction engineering feature data is input into the TCN temporal convolutional network. The dilated convolutional layers capture cost accumulation features at different time points. Pooling layers reduce the dimensionality of the captured temporal features, removing redundant temporal information. Finally, the output layer outputs the temporal features. A cross-modal attention mechanism is designed, taking the global spatial feature vector output from the HGAT network and the global temporal feature vector output from the TCN network as inputs. An attention weight calculation function is constructed, calculating the weight coefficients of spatial and temporal features based on feature similarity. Based on the calculated weight coefficients, the global spatial feature vector and the global temporal feature vector are weighted and fused. When the feature weight is below a threshold, the weight coefficients are adjusted. The construction project feature data is input into the multi-scale map time series fusion prediction model, and the predicted project cost is output. Based on the predicted project cost, when the actual project cost exceeds the predicted cost, an attribution mechanism based on SHAP value and counterfactual reasoning is triggered to generate a quantifiable optimization solution.

2. The BIM-based construction project cost assessment and management method as described in claim 1, characterized in that, The process of acquiring BIM model data involves converting the 3D mesh of the components in the BIM model data into a point cloud sequence with normal vectors, extracting features using a PointNet++ network, and hierarchically encoding the component type, material, and construction process to obtain architectural engineering feature data, including: Obtain BIM model data, including at least the 3D mesh of components, component type, materials, and construction process. Then, filter, clean, and standardize the collected BIM model data in sequence to obtain preprocessed BIM model data. A mesh segmentation algorithm is used to divide the 3D mesh of components in the standardized preprocessed BIM model data, dividing the 3D entity of each component into a uniform triangular mesh to obtain mesh component data; The triangular mesh of each component in the mesh component data is uniformly sampled to generate the initial point cloud data of the component; the least squares method is used to fit the plane where the sampling points are located, and the normal vector of each sampling point is calculated to obtain the point cloud sequence with normal vector.

3. The method for cost assessment and management of building construction projects based on BIM as described in claim 1, characterized in that, The process of acquiring BIM model data involves converting the 3D mesh of the components in the BIM model data into a point cloud sequence with normal vectors, extracting features using a PointNet++ network, and hierarchically encoding the component type, material, and construction process to obtain architectural engineering feature data, including: A PointNet++ network was built with an input layer dimension of 4, a sampling layer using the FPS farthest point sampling algorithm, a feature extraction layer divided into 3 scales, each scale containing 2 convolutional layers, an activation function using the ReLU function, and an output layer dimension of 128. The point cloud sequence with normal vectors is input into the PointNet++ network. Local geometric features of the component surface are extracted through local scale convolutional layers, and local features are aggregated into regional features through regional scale convolutional layers. The global-scale convolutional layer fuses regional features and outputs a global feature vector of the component.

4. The BIM-based construction project cost assessment and management method as described in claim 1, characterized in that, The process involves determining the project cost based on the predicted cost. When the actual project cost exceeds the predicted cost, an attribution mechanism based on SHAP value and counterfactual reasoning is triggered to generate a quantifiable optimization solution, including: Input the architectural engineering feature data of the current project into the trained multi-scale graph time series fusion prediction model, and calculate the SHAP value of each feature; We selected the top 10 features with positive SHAP values ​​and large absolute values, and analyzed the specific reasons for cost overruns caused by each core feature in light of the actual project situation.

5. A BIM-based construction project cost assessment and management system, characterized in that, The construction project cost assessment and management system includes the following modules: The BIM data acquisition module is used to acquire BIM model data, convert the 3D mesh of the components in the BIM model data into a point cloud sequence with normal vectors, extract features through the PointNet++ network, and hierarchically encode the component type, material and construction technology to obtain building engineering feature data. The prediction model building module is used to build a multi-scale graph temporal fusion prediction model. The spatial relationship channel uses the HGAT hierarchical graph attention network to aggregate spatial features layer by layer, while the temporal evolution channel uses the TCN temporal convolutional network to capture the time features of cost accumulation. The weights of spatial and temporal features are adjusted through a cross-modal attention mechanism. This process includes: constructing a spatial relationship graph for building projects, using components from the building project feature data as nodes and the spatial relationships between components as edges. Node features are the building project feature data, and edge weights are set according to the strength of the spatial relationships between components. A hierarchical HGAT graph attention network is built, with three layers: each containing an attention mechanism layer, a convolutional layer, and an activation layer. The attention mechanism layer uses multi-head attention, the convolutional layers have 3×3 kernels, and the activation function is LeakyReLU. The building project spatial relationship graph is input into the HGAT network, and spatial features are output. A TCN temporal convolutional network is built, with four layers: dilated convolutions with 3×3 kernels, max pooling layers, and the activation function is ReLU. The building project feature data is input into the TCN temporal convolutional network, and the dilated convolutional layers capture cost accumulation features at different time points. The pooling layers reduce the dimensionality of the captured temporal features and remove redundant temporal information. Finally, the output layer outputs the temporal features. A cross-modal attention mechanism is designed, with the input being the global spatial feature vector output by the HGAT network and the global temporal feature vector output by the TCN network. An attention weight calculation function is constructed, and the weight coefficients of spatial features and temporal features are calculated based on feature similarity. According to the calculated weight coefficients, the global spatial feature vector and the global temporal feature vector are weighted and fused. When the feature weight is lower than the threshold, the weight coefficients are adjusted. The engineering cost assessment module is used to input the architectural engineering characteristic data into the multi-scale map time series fusion prediction model and output the predicted engineering cost value. The cost assessment and management module is used to make judgments based on the predicted project cost. When the actual project cost is greater than the predicted project cost, it triggers an attribution mechanism based on SHAP value and counterfactual reasoning to generate a quantifiable optimization solution.

6. The BIM-based construction project cost assessment and management system as described in claim 5, characterized in that, The prediction model building module includes the following sub-modules: A submodule is constructed to design a cross-modal attention mechanism. The input is the global spatial feature vector output by the HGAT network and the global temporal feature vector output by the TCN network. An attention weight calculation function is constructed to calculate the weight coefficients of spatial features and temporal features based on feature similarity. The fusion submodule is used to perform weighted fusion of the global spatial feature vector and the global temporal feature vector based on the calculated weight coefficients, and adjust the weight coefficients when the feature weights are lower than the threshold.

7. The BIM-based construction project cost assessment and management system as described in claim 5, characterized in that, The cost assessment management module includes the following sub-modules: The calculation submodule is used to input the architectural engineering feature data of the current project into the trained multi-scale graph time series fusion prediction model and calculate the SHAP value of each feature. The filtering submodule is used to filter out the top 10 features with positive SHAP values ​​and large absolute values, and analyze the specific reasons for cost overruns caused by each core feature in conjunction with the actual situation of the project.

Images