Design construction drawing output efficiency optimization method based on BIM

Abstract

The invention provides a BIM-based design construction drawing output efficiency optimization method, and belongs to the technical field of construction drawing output efficiency optimizing.According to the BIM-based design construction drawing output efficiency optimization method, component visibility judgment is accelerated by establishing a spatial octree index structure, and the rendering performance is optimized by adopting sight distance grading and a heterogeneous computing architecture; a component semantic recognition network is used for automatically recognizing component categories, a two-dimensional construction drawing conforming to specifications is generated in combination with a drawing rule base, and an improved shortest path pipeline optimization algorithm is applied to coordinate multi-objective optimization of path length and cross conflicts through a double-layer game model. Building information model data are updated according to the pipeline comprehensive arrangement result, iterative execution is triggered until the construction drawing meets the specification requirement, full-process automatic processing from a three-dimensional model to a two-dimensional drawing to pipeline comprehensive arrangement is achieved, and the technical problem that automatic conversion from the model to the drawing cannot be achieved when the BIM generates the construction drawing is solved.

Description

Technical Field

[0001] This invention belongs to the field of construction drawing output efficiency optimization technology, and specifically, it relates to a BIM-based method for optimizing the output efficiency of design and construction drawings. Background Technology

[0002] In the field of Building Information Modeling (BIM) applications, traditional methods for generating construction drawings primarily rely on manual operation of computer-aided design software to create 2D drawings, or on projecting 3D models into 2D drawings using simple model view export functions. This approach has significant limitations in practical engineering applications. Current technologies lack intelligent component semantic recognition methods, making it impossible to automatically identify component categories and match drafting standards during the model-to-drawing conversion process. Designers must manually set layer line types and other parameters for each component. Furthermore, existing component recognition methods mainly rely on manually written rules to match component attributes. When dealing with complex and diverse component types, the rule base maintenance cost is high, and the recognition accuracy is low, failing to meet the component classification needs of different projects. In addition, traditional methods often employ trial-and-error manual adjustments when handling integrated pipeline layouts, making it difficult to ensure the economy of pipeline paths while avoiding pipeline intersections and conflicts. This results in a heavy reliance on manual intervention in the conversion process from 3D models to compliant construction drawings, leading to extremely low conversion efficiency and a high risk of omissions and errors. In other words, existing technologies suffer from the technical problem of being unable to automate the model-to-drawing conversion process when generating construction drawings using BIM. Summary of the Invention

[0003] In view of this, the present invention provides a method for optimizing the efficiency of BIM-based design and construction drawing output, which can solve the technical problem in the prior art that the automatic conversion from model to drawing cannot be achieved when generating construction drawings using BIM.

[0004] This invention is implemented as follows: It provides a method for optimizing the efficiency of BIM-based design and construction drawing output. This includes collecting building information model data and establishing a spatial octree index structure; obtaining viewpoint positions and view frustum parameters and determining component visibility; performing distance-level processing on retained components; using a component feature classification model to classify component processing tasks and assign them to a graphics processor or central processing unit; performing instantiation and rendering processing on components assigned to the graphics processor; establishing a component semantic recognition network to identify component categories in the 3D model and output semantic category labels and expression priority weights; querying the drafting rule base based on the semantic category labels to obtain layer configuration parameters and line style parameters; using a depth buffer sorting algorithm to process component occlusion relationships and determine the drawing order; performing orthogonal projection transformation on the 3D model according to the projection direction and sequentially drawing the 2D outlines of the components to complete the construction drawing generation; establishing a pipeline path node network and using an improved shortest path pipeline optimization algorithm for pipeline integration and layout; updating the building information model data based on the pipeline integration and layout results and triggering re-execution until the construction drawings meet the specifications.

[0005] The building information model data includes the geometric attributes of the components and the spatial location attributes of the components. The geometric attributes include the vertex coordinates of the components, the triangular facets of the components, and the normal vectors of the components. The spatial location attributes include the center point coordinates of the components and the bounding box boundary coordinates of the components.

[0006] The spatial octree index structure recursively divides the cubic space into eight sub-cubes. Each octree node stores a list of component identifiers that fall within its spatial range. The number of components contained in a leaf node does not exceed a component number threshold, which is ∈ [50, 200].

[0007] The parameters of the view frustum include viewpoint coordinates, line-of-sight vector, view angle, near clipping plane distance, and far clipping plane distance. The method for determining the intersection relationship between the view frustum and the octree node is to extract the coordinates of the eight vertices of the octree node's cube bounding box and project the coordinates of the eight vertices onto the six clipping planes of the view frustum.

[0008] The calculation method for the multi-level detail level coefficient is to calculate the Euclidean distance from the center point of the component to the viewpoint, divide the Euclidean distance by the distance standard value to obtain the normalized distance ratio, and determine the multi-level detail level coefficient based on the interval in which the normalized distance ratio is located. The distance standard value is 100m.

[0009] The component feature classification model uses a classification decision function for determination. When the product of the number of triangular faces of a component divided by the standard value of the number of triangular faces and the number of vertices of a component divided by the standard value of the number of vertices is greater than the number of attribute fields of a component divided by the standard value of the number of attribute fields, the processor allocation identifier is output as the graphics processor identifier; otherwise, the processor allocation identifier is output as the central processing unit identifier.

[0010] The parallel rendering process of the graphics processor adopts a thread block partitioning strategy. The required number of thread blocks is calculated based on the total number of geometrically dense components allocated to the graphics processor and the number of threads in a single thread block. The number of threads in a single thread block is 256.

[0011] The shared geometric data for instantiated rendering processing includes an array of vertex positions, an array of normal vectors, and an array of texture coordinates. The transformation matrix is ​​a composite result of a translation matrix, a rotation matrix, and a scaling matrix, and the transformation matrix has a dimension of 4 rows and 4 columns.

[0012] The component semantic recognition network consists of an input layer that receives point cloud data of components, three convolutional layers that extract local geometric features, and 64, 128 and 256 convolutional kernels for each layer, with a kernel size of 3x3. After the convolutional layers, a max pooling layer is connected to reduce the feature dimension, with a pooling window size of 2x2.

[0013] Among them, the component semantic recognition network adopts the principle of the egg-throwing problem to optimize the network depth search strategy. The number of network layers is regarded as the height of the building, and the network performance threshold is regarded as the critical point at which the egg breaks. The optimal layer exploration sequence is determined by dynamic programming to find the shallowest network structure that meets the network performance threshold with the fewest training times.

[0014] The expression priority weight ranges from [0, 1] and is determined based on the semantic category label. The expression priority weight of load-bearing structural components is greater than that of non-load-bearing components, and the expression priority weight of components in the main functional space is greater than that of components in the auxiliary space.

[0015] The cartographic rule base uses a key-value pair data structure for storage. The key is a semantic category label, and the value is a structure containing the layer name, line type code, line width value, and color code. The line type code includes solid line, dashed line, dotted line, and double dotted line, and the line width value is in mm.

[0016] The depth buffer sorting algorithm uses a depth buffer to store the current minimum depth value of each pixel. The depth value is the vertical distance from the surface point of the component to the projection plane. The drawing order is determined by sorting the components from largest to smallest depth value. For components with similar depth values, a second sort is performed based on the expression priority weight.

[0017] In the orthogonal projection transformation, a projection matrix is ​​used to map three-dimensional coordinates to two-dimensional coordinates. The projection matrix is ​​determined according to the projection direction, which includes the front view direction, the top view direction, and the side view direction.

[0018] The method for establishing the pipeline path node network is to divide the building space into uniform grids with a grid size of 0.5m. The intersections of the uniform grids are used as path nodes, and edge connections are established between adjacent path nodes. The initial weight of the edge connection is the Euclidean distance between the path nodes.

[0019] Among them, the improved shortest path pipeline optimization algorithm adopts a two-layer game model for path optimization. The two-layer game model includes an upper-layer model with the goal of minimizing the total path length and a lower-layer model with the goal of minimizing the number of pipeline intersections. The coupling term of the two-layer game model is the occupancy penalty coefficient of the path node.

[0020] This invention automatically identifies component categories by constructing a component semantic recognition network and generates compliant 2D drawings by combining it with a drafting rule base. It employs deep learning to extract local geometric features from component point clouds and outputs semantic category probability distributions. The feature learning capabilities of convolutional and fully connected layers replace manual rule design, improving recognition accuracy and adaptability in complex scenarios. Simultaneously, a drafting rule base is established to store layer line type parameters for different component categories in key-value pairs, allowing semantic recognition results to be directly mapped to drafting parameters. A deep buffer sorting algorithm handles component occlusion relationships and determines the drawing order based on expression priority weights. An improved shortest path pipeline optimization algorithm coordinates path length optimization and intersection conflict optimization through a two-layer game model. Heuristic distance estimation and dynamic weight adjustment mechanisms are introduced to automatically generate reasonable pipeline layout schemes. Based on the pipeline layout results, the building information model data is updated, triggering re-execution until the construction drawings meet the specifications. This achieves fully automated processing from 3D model rendering to 2D drawing generation and pipeline optimization, solving the technical problem of the inability to automatically convert models to drawings when generating construction drawings from BIM. Attached Figure Description

[0021] Figure 1 A 3D visualization of the spatial octree index structure.

[0022] Figure 2 A comparison chart showing the number of path nodes searched between the improved algorithm and the traditional algorithm.

[0023] Figure 3 The graph shows the iterative curves of the optimization process for the two-layer game model. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] This invention provides a method for optimizing the efficiency of BIM-based design and construction drawing output, comprising the following steps:

[0026] A method for optimizing the efficiency of design and construction drawing output based on BIM includes the following steps:

[0027] S01. Collect building information model data and establish a spatial octree index structure. Recursively divide the three-dimensional space into eight subspaces until the number of components is lower than the component number threshold. Record the component identifier contained in each octree node.

[0028] S02. Obtain the viewpoint position and view frustum parameters, determine the visibility of components based on the viewpoint position and view frustum parameters, calculate the intersection relationship between each octree node and the view frustum, and remove completely invisible octree nodes and their contained component identifiers.

[0029] S03. Perform view distance grading processing on the components corresponding to the retained component identifiers, and calculate the multi-level detail level coefficient based on the Euclidean distance from the component center point to the view point position. The multi-level detail level coefficient is used to determine the component rendering accuracy level.

[0030] S04. Use a component feature classification model to classify the components corresponding to the component rendering accuracy level into processing tasks. Based on the classification results, assign geometrically dense components to the graphics processor for parallel rendering and assign attribute query components to the central processing unit for serial processing.

[0031] S05. Instantiate and render geometrically dense components allocated to the graphics processor, extract shared geometric data of similar components and store them in the video memory buffer, generate component instances in batches through transformation matrix and complete the 3D scene rendering.

[0032] S06. Establish a component semantic recognition network to identify component categories of the 3D model after the 3D scene rendering is completed. The component semantic recognition network outputs the semantic category label of each component and its expression priority weight in the 2D projection plane.

[0033] S07. Query the drafting rule base according to the semantic category label to obtain the corresponding layer configuration parameters and line style parameters. The drafting rule base stores the standard expression of different component categories in construction drawings.

[0034] S08. The depth buffer sorting algorithm is used to process the occlusion relationship of the components. The depth value of each component pixel is calculated and sorted from far to near according to the depth value. The drawing order of the components is determined according to the sorting result and the expression priority weight.

[0035] S09. Perform orthogonal projection transformation on the three-dimensional model according to the set projection direction, draw the two-dimensional outlines of the components in sequence according to the drawing order, and apply the line style parameters to complete the construction drawing generation.

[0036] S10. Establish a pipeline path node network and use an improved shortest path pipeline optimization algorithm for comprehensive pipeline layout. The improved shortest path pipeline optimization algorithm optimizes pipeline paths through a two-layer game model.

[0037] S11. Update the building information model data based on the integrated pipeline layout results, write the optimized pipeline path coordinates into the geometric attributes of the corresponding components, and trigger step S01 to be re-executed until the construction drawings meet the specifications.

[0038] The building information model data includes the geometric attributes of the components and the spatial location attributes of the components. The geometric attributes include the vertex coordinate data, the triangular facet data, and the normal vector data of the components. The spatial location attributes include the center point coordinates and the bounding box boundary coordinates of the components.

[0039] The spatial octree index structure is a three-dimensional spatial data organization method that recursively divides a cubic space into eight sub-cubes. Each octree node stores a list of component identifiers falling within its spatial range. The number of components contained in a leaf node does not exceed a component quantity threshold, which is ∈ [50, 200]. The bounding box of the root node of the spatial octree index structure is determined by the minimum bounding cube of the building information model data, where the side length of the minimum bounding cube is the maximum span of the model along the three coordinate axes.

[0040] The parameters of the view frustum include viewpoint coordinates, line-of-sight vector, viewing angle, near clipping plane distance, and far clipping plane distance. The method for determining the intersection relationship between the view frustum and the octree node is as follows: extract the coordinates of the eight vertices of the bounding box of the octree node, project these eight vertex coordinates onto the six clipping planes of the view frustum. If all vertex coordinates are located outside any clipping plane, the octree node is completely invisible; if some vertex coordinates are located inside the clipping plane, the octree node is partially visible; if all vertex coordinates are located inside all clipping planes, the octree node is completely visible.

[0041] The calculation method for the multi-level detail level coefficient is as follows: First, calculate the Euclidean distance from the center point of the component to the viewpoint position, where the Euclidean distance is in meters. Then, divide the Euclidean distance by the standard distance value to obtain the normalized distance ratio, where the standard distance value is 100 meters. Determine the multi-level detail level coefficient based on the interval of the normalized distance ratio. When the normalized distance ratio ∈ [0, 0.2), the multi-level detail level coefficient is 3, corresponding to a high-precision model. When the normalized distance ratio ∈ [0.2, 0.6], the multi-level detail level coefficient is 2, corresponding to a medium-precision model. When the normalized distance ratio ∈ (0.6, 1), the multi-level detail level coefficient is 1, corresponding to a low-precision model.

[0042] The component feature classification model is a binary classification model based on the component's geometric complexity and attribute data volume. The input parameters of the model include the number of triangles, the number of vertices, and the number of attribute fields. The output is a processor allocation identifier. The model uses a classification decision function, which is described as follows: when the product of the number of triangles divided by the standard value of the number of triangles and the number of vertices divided by the standard value of the number of vertices is greater than the product of the number of attribute fields divided by the standard value of the number of attribute fields, the processor allocation identifier outputs a graphics processor identifier; otherwise, the processor allocation identifier outputs a central processing unit identifier. The standard value for the number of triangles is 1000, the standard value for the number of vertices is 500, and the standard value for the number of attribute fields is 20.

[0043] The parallel rendering process of the graphics processor adopts a thread block partitioning strategy. The required number of thread blocks is calculated based on the total number of geometrically dense components allocated to the graphics processor and the number of threads in a single thread block. The number of threads in a single thread block is 256. Each thread is responsible for processing the transformation matrix calculation and vertex shading of one component instance.

[0044] The shared geometric data of the instantiation rendering process includes a vertex position array, a normal vector array, and a texture coordinate array. The transformation matrix includes a composite result of a translation matrix, a rotation matrix, and a scaling matrix. The transformation matrix has a dimension of 4 rows and 4 columns. The vertex positions of the shared geometric data are transformed to the world coordinate positions of each component instance through matrix multiplication.

[0045] The specific structure of the component semantic recognition network is as follows: The input layer receives point cloud data of the component, where the point cloud data dimension is the number of component vertices multiplied by 3. Local geometric features are extracted through three convolutional layers, with 64, 128, and 256 convolutional kernels in each layer, and a kernel size of 3x3. After the convolutional layers, a max-pooling layer is connected to reduce the feature dimension, with a pooling window size of 2x2. Then, feature fusion is performed through two fully connected layers, with 512 and 128 neurons in each layer. Finally, the output layer generates a component semantic category probability distribution, with the number of neurons in the output layer equal to the predefined total number of component categories, which is 50. The component semantic recognition network uses an adaptive attention mechanism to adjust feature weights. The weight parameters of the adaptive attention mechanism are determined based on the number of triangular faces and the surface area of ​​the component. When the number of triangular faces divided by the standard value of the number of triangular faces is greater than the surface area divided by the standard value of the surface area, the geometric feature weights are increased; otherwise, the topological feature weights are increased. The standard value of the surface area is 10. .

[0046] The steps for establishing the training dataset of the component semantic recognition network specifically include: extracting component samples with labeled semantic categories from the building information model database; normalizing the vertex coordinates of each component sample to make them lie within a unit cube; uniformly sampling the surface of the component sample using the Poisson disk sampling method to generate the point cloud data, wherein the number of sampling points in the point cloud data is 2048; forming training sample pairs with the point cloud data and their corresponding semantic category labels; and dividing the training set and validation set according to an 8:2 ratio.

[0047] The specific steps of training the component semantic recognition network include: initializing the network weight parameters as random numbers following a normal distribution with a mean of 0 and a standard deviation of 0.01; using the cross-entropy loss function to measure the difference between the predicted category and the true category; updating the network weight parameters using the momentum gradient descent optimization algorithm, wherein the initial learning rate of the momentum gradient descent optimization algorithm is 0.001, the momentum coefficient is 0.9, the training batch size is 32, the number of training rounds is 100, the network performance is evaluated on the validation set every 10 rounds, and training is terminated and the network weight parameters with the highest accuracy are saved when the accuracy of the validation set does not improve for 5 consecutive rounds.

[0048] The component semantic recognition network employs the egg-throwing problem principle to optimize its network depth search strategy. The implementation of this principle in the network is as follows: the number of network layers is considered as the height of a building, and the network performance threshold is considered as the critical point at which the egg breaks. A dynamic programming method is used to determine the optimal layer exploration sequence to find the shallowest network structure that satisfies the network performance threshold with the fewest training iterations. Specifically, a state transition matrix is ​​established to record the optimal exploration strategy under different network layers and remaining attempts. The first row of the state transition matrix is ​​initialized so that only linear search can be used in a single attempt. For multiple attempts, a recursive formula is used to calculate the optimal depth search strategy for the current state. The optimal decision-making process is based on the recursive formula: the optimal number of trial layers in the current state is equal to the number of layers that minimizes the maximum number of attempts for both the broken and unbroken egg scenarios. The optimal exploration path from the initial layer to the target layer is obtained by filling the state transition matrix from bottom to top. During the training of the component semantic recognition network, the layer configurations in the optimal exploration path are tried sequentially. If the validation set accuracy of the current layer configuration is lower than the network performance threshold, the number of network layers is increased; if the validation set accuracy of the current layer configuration is higher than the network performance threshold, the number of network layers is decreased. This process is repeated until the shallowest network layer that satisfies the network performance threshold is found, where the network performance threshold is set to 90%. The principle behind the egg-throwing problem is to establish an optimal substructure using dynamic programming, decomposing the network depth search problem into multiple subproblems. Each subproblem corresponds to a combination of network layer number and remaining attempts. The row index of the state transition matrix represents the remaining attempts, the column index represents the network layer to be searched, and the matrix element value represents the minimum number of attempts required to start from the current state. The derivation of the recursive formula is based on the following logic: after trying a certain layer, two results will occur. If the validation set accuracy is lower than the network performance threshold, the search needs to continue at a deeper layer and the remaining attempts will decrease by one. If the validation set accuracy is higher than the network performance threshold, the search needs to continue at a shallower layer and the remaining attempts will decrease by one. To minimize the number of attempts in the worst-case scenario, the optimal number of attempts needs to be selected, which minimizes the maximum number of attempts for both outcomes. By filling the state transition matrix, the optimal strategy in the initial state is obtained by progressively calculating from the bottom subproblems to the top-level problem. The technical effect brought by the principle of the egg-throwing problem is to avoid the computational overhead of exhaustively searching all network layer configurations. By intelligently selecting the order of attempts, the number of attempts is significantly reduced while ensuring that the shallowest network structure that meets the network performance threshold is found. Compared with the linear search method, the average number of attempts can be reduced by about 60%, shortening the network structure optimization time, so that the network depth optimization can still be completed within an acceptable time when the network layer search range is large.

[0049] The expression priority weight ranges from [0, 1]. The larger the expression priority weight, the higher the importance of the component in the two-dimensional drawing. The expression priority weight is determined according to the semantic category label. The expression priority weight of the load-bearing structural component is greater than that of the non-load-bearing component. The expression priority weight of the component in the main functional space is greater than that of the component in the auxiliary space.

[0050] The mapping rule base is stored using a key-value pair data structure. The key is the semantic category label, and the value is a structure containing the layer name, line type code, line width value, and color code. The line type code includes solid line, dashed line, dotted line, and double dotted line. The line width value is in mm and is ∈ [0.1, 1.0]. The color code is represented by three channels: red, green, and blue, with each channel having a value range of [0, 255].

[0051] The depth buffer sorting algorithm uses a depth buffer to store the current minimum depth value of each pixel. The depth value is the vertical distance from the component surface point to the projection plane, and the unit of the depth value is meters (m). When traversing the projected pixels of all components, the depth value of the current pixel is compared with the value stored in the depth buffer. If the depth value of the current pixel is less than the value stored in the depth buffer, the depth buffer is updated and the pixel is recorded as belonging to the current component; otherwise, the pixel is occluded by the component in front. The method for determining the drawing order is as follows: first, the components are sorted from largest to smallest according to the depth value. For components with similar depth values, a second sort is performed according to the expression priority weight. The component with the larger expression priority weight is drawn first. The threshold for judging similar depth values ​​is 0.01m.

[0052] The orthogonal projection transformation uses a projection matrix to map three-dimensional coordinates to two-dimensional coordinates. The projection matrix is ​​determined according to the projection direction, which includes the front view direction, the top view direction, and the side view direction. For the front view direction, the projection matrix retains the horizontal and vertical coordinates of the component but discards the depth coordinates. For the top view direction, the projection matrix retains the horizontal and depth coordinates of the component but discards the vertical coordinates. For the side view direction, the projection matrix retains the depth and vertical coordinates of the component but discards the horizontal coordinates.

[0053] The method for establishing the pipeline path node network is as follows: the building space is divided into uniform grids according to the grid size, the grid size is 0.5m, the intersection of the uniform grids is used as path nodes, and edge connections are established between adjacent path nodes. The initial weight of the edge connection is the Euclidean distance between the path nodes. If the edge connection passes through a wall or other obstacle, the weight of the edge connection is set to infinity to indicate that it is impassable.

[0054] The improved shortest path pipeline optimization algorithm uses a two-layer game model for path optimization. The two-layer game model includes an upper-layer model that aims to minimize the total path length and a lower-layer model that aims to minimize the number of pipeline intersections.

[0055] The objective function of the upper-level model is expressed as follows: the normalized total path length equals the length of each edge connection in all pipeline paths divided by the sum of the standard edge length values ​​divided by the number of pipelines. The standard edge length value is 1m. The constraints of the upper-level model include that each pipeline path must continuously reach the end point from the starting point without passing through the edge connections with infinite weights, and that the pipeline paths must not exceed the building space boundaries. The objective function of the upper-level model measures the compactness of the pipeline layout. The input of the objective function of the upper-level model includes the set of edge connections for all pipeline paths and the length of each edge connection. The output of the objective function of the upper-level model is the normalized total path length value.

[0056] The objective function of the lower-level model is expressed as follows: the normalized value of the number of pipeline intersections equals the sum of the number of shared path nodes of any two pipeline paths divided by the number of standard node numbers divided by the number of pipelines. The standard node number is 100. The constraints of the lower-level model include that the minimum distance between pipelines must not be less than a safety distance threshold. The safety distance threshold is determined according to the pipeline type; the safety distance threshold for electrical pipelines and water supply / drainage pipelines is 0.3m, and the safety distance threshold for pipelines of the same type is 0.1m. The objective function of the lower-level model measures the degree of conflict in pipeline layout. The input of the objective function includes the set of path nodes for all pipeline paths and the coordinates of the path nodes. The output of the objective function is the normalized value of the number of pipeline intersections.

[0057] In this two-layer game model, the coupling term is the occupancy penalty coefficient of a path node. When the pipeline path selected by the upper-layer model passes through a path node, the occupancy penalty coefficient of that path node increases. When calculating the number of pipeline intersections, the lower-layer model applies extra weight to path nodes with high occupancy penalty coefficients, prompting subsequent pipeline paths to avoid already occupied path nodes. The update method for the occupancy penalty coefficient is as follows: the initial value is 1, and each time a pipeline path passes through the path node, the occupancy penalty coefficient is multiplied by a penalty growth factor, where the penalty growth factor is set to 1.5.

[0058] The improved shortest path pipeline optimization algorithm introduces heuristic distance estimation and dynamic weight adjustment mechanisms on the basis of traditional shortest path algorithms. The heuristic distance estimation uses Manhattan distance as the estimated remaining distance from the current path node to the target path node. The Manhattan distance is the sum of the absolute values ​​of the coordinate differences between the current path node and the target path node along the three coordinate axes. This heuristic distance estimation guides the algorithm to prioritize exploring path nodes closer to the target, thereby reducing the search space. The dynamic weight adjustment mechanism adjusts the weights of edge connections based on pipeline type and path history information. For edge connections traversed by multiple pipeline paths, their weights are increased to reduce their selection priority. The weight adjustment formula for edge connections is: the adjusted weight equals the initial weight multiplied by 1 plus the product of the number of pipelines traversed by the edge connection divided by a standard value of the number of pipelines, where the standard value of the number of pipelines is 5. The improved shortest path pipeline optimization algorithm uses a priority queue to maintain the path nodes to be visited. The priority queue is sorted according to the sum of the actual distance of the path node and the heuristic distance estimate. Each time, the path node with the smallest total distance is taken from the priority queue for expansion. The new path distance is calculated for all adjacent path nodes of the path node. If the new path distance is less than the known shortest distance of the adjacent path node, the shortest distance of the adjacent path node is updated and the adjacent path node is added to the priority queue. This process is repeated until the target path node is visited or the priority queue is empty.

[0059] The improved shortest path pipeline optimization algorithm offers the following technical benefits: It significantly reduces the number of path nodes searched through heuristic distance estimation, lowering the average number of searched path nodes by approximately 60% compared to traditional shortest path algorithms. This shortens path planning computation time, enabling single pipeline path planning to be completed within seconds even in large-scale building spaces containing tens of thousands of path nodes. The dynamic weight adjustment mechanism effectively avoids congestion caused by multiple pipeline paths concentrating on a few paths. By increasing the weight of high-throughput edges, it guides subsequent pipeline paths to choose other paths, achieving a uniform distribution of pipeline paths in space and reducing the probability of pipeline path intersections. The introduction of the two-layer game model allows path length optimization and intersection conflict optimization to proceed collaboratively. The upper-layer model pursues shortest paths while the lower-layer model avoids intersections. The upper and lower-layer models are coupled through the occupancy penalty coefficient to form a dynamic balance, ensuring both the economy of the pipeline path and meeting the construction specifications' requirements for pipeline spacing. The improved shortest path pipeline optimization algorithm generates a comprehensive pipeline layout scheme that reduces the number of pipeline path intersections by about 70% while maintaining a shorter total path length. This significantly improves the rationality and constructability of the comprehensive pipeline layout and reduces rework issues caused by pipeline path conflicts during later construction.

[0060] The combined application of the component feature classification model, the component semantic recognition network, and the improved shortest path pipeline optimization algorithm automates the entire process from 3D model rendering to 2D drawing generation and pipeline optimization, significantly improving the efficiency of design and construction drawing output based on Building Information Modeling (BIM). The component feature classification model intelligently allocates processor resources based on the geometric complexity and attribute data volume of the components, allowing geometrically dense components to fully utilize the parallel computing capabilities of the graphics processor, while attribute-query-type components are processed serially on the central processing unit, avoiding unnecessary data transmission overhead. This heterogeneous computing architecture improves overall rendering performance by approximately three times compared to a single-processor solution, resolving the performance bottleneck of real-time rendering of massive BIM data. The component semantic recognition network automatically identifies component categories using deep learning methods and generates compliant 2D drawings based on the drafting rule base. Compared to traditional rule-based methods, it improves recognition accuracy by approximately 25% in complex scenarios. Furthermore, the component semantic recognition network can continuously optimize its recognition performance through training to adapt to new component categories, significantly reducing the workload of manual annotation and rule maintenance. The improved shortest path pipeline optimization algorithm generates a comprehensive pipeline layout scheme that ensures both the economy of pipeline paths and compliance with construction specifications, reducing design changes and construction rework caused by pipeline path conflicts and shortening the overall project cycle. The synergistic effect of the component feature classification model, the component semantic recognition network, and the improved shortest path pipeline optimization algorithm makes the conversion process from building information model to construction drawings more efficient and accurate, providing reliable technical support for architectural design and construction.

[0061] Furthermore, in step S04, the component feature classification model adopts a collaborative processing architecture of a central processing unit (CPU) and a graphics processing unit (GPU). The data flow of the collaborative processing architecture is as follows: the CPU is responsible for reading the component's attribute data and calculating the classification decision function. Based on the output of the classification decision function, the geometric data of the geometrically dense components that need to be processed by the GPU is transmitted to the GPU's video memory via a high-speed bus. After receiving the geometric data, the GPU allocates computing resources according to the thread block partitioning strategy. Each thread block processes a batch of rendering tasks for geometrically dense components. After rendering is completed, the result frame buffer data is transmitted back to the CPU's system memory. The CPU integrates the GPU's rendering results and its own processed attribute query results to form complete 3D scene data.

[0062] Furthermore, in step S06, when the confidence level of the semantic category output by the component semantic recognition network is lower than the confidence level threshold, a manual review process is triggered. The confidence level threshold is 0.8. The manual review process includes displaying the point cloud data and predicted category of the low-confidence component to the reviewer. After the reviewer confirms or corrects the component category, the correction result is added to the training dataset, and the component semantic recognition network is retrained periodically to improve the recognition accuracy.

[0063] Furthermore, in step S10, when there are pipeline pairs in the pipeline layout scheme that cannot meet the safety distance threshold constraint, the pipeline paths with lower priority in the pipeline pairs are automatically adjusted. The priority is determined according to the pipeline type, with fire protection pipelines having the highest priority, followed by water supply and drainage pipelines, then electrical pipelines, and finally ventilation pipelines having the lowest priority. The pipeline path adjustment method is to offset the start or end point of the pipeline path with lower priority along the vertical direction by the safety distance threshold and then re-execute the improved shortest path pipeline optimization algorithm.

[0064] Furthermore, this method also includes step S12, which automatically verifies the generated construction drawings. The automatic verification includes checking whether there are unclosed outlines in the drawings, checking whether the dimension annotations are complete, and checking whether the layer settings comply with the provisions of the drafting rule library. When the automatic verification finds non-compliance items, it automatically records the location and type of the non-compliance items and generates a verification report. After correcting the building information model data or drafting parameters according to the verification report, steps S01 to S11 are re-executed.

[0065] Furthermore, the component semantic recognition network employs a transfer learning method to improve training efficiency. The steps of the transfer learning method are as follows: first, the convolutional layers of the component semantic recognition network are pre-trained on a large-scale general-purpose 3D model dataset to obtain general geometric feature extraction capabilities; then, the weight parameters of the convolutional layers are fixed, and only the fully connected layers and the output layer are trained to adapt to the category distribution of building components. The large-scale general-purpose 3D model dataset contains ≥100,000 labeled samples. Through the transfer learning method, the training time of the component semantic recognition network can be shortened by about 50%, and a high recognition accuracy can still be maintained even with small samples.

[0066] Furthermore, in step S03, when the multi-level detail level coefficient of the same component changes under different viewing distances, a transition animation is used to smoothly switch between models of different precision. The duration of the transition animation is ∈ [0.2, 0.5], and the unit of the duration of the transition animation is s. The transition animation gradually replaces the vertex position of the low precision model with the vertex position of the high precision model through a linear interpolation method, thus avoiding the visual jump caused by the sudden change in model precision.

[0067] Furthermore, in step S02, the view frustum parameters are provided by user interaction input or by a preset view configuration file. The user interaction input includes mouse drag operation to change the viewpoint position and the line-of-sight vector. The preset view configuration file stores commonly used viewpoint position and view frustum parameter combinations for quick view switching.

[0068] Furthermore, in step S05, the transformation matrix is ​​calculated based on the spatial position attributes of the component instance, the translation matrix is ​​determined by the center point coordinates of the component instance, the rotation matrix is ​​determined by the orientation angle of the component instance, and the scaling matrix is ​​determined by the size ratio of the component instance.

[0069] Furthermore, in step S07, the layer configuration parameters include layer display order and layer visibility identifier. The layer display order determines the superposition order of different layers in the construction drawing, and the layer visibility identifier is used to control the display or hiding state of the layer in the construction drawing.

[0070] Furthermore, in step S08, the resolution of the depth buffer is consistent with the pixel resolution of the two-dimensional projection plane, each pixel corresponds to a storage unit in the depth buffer, and the storage unit is initialized to the maximum depth value representing infinity.

[0071] Furthermore, in step S09, the two-dimensional contour line is drawn using a vector graphics format, which includes geometric descriptions of straight line segments and arc segments. The straight line segments are defined by the coordinates of the starting point and the coordinates of the ending point, and the arc segments are defined by the coordinates of the center, the radius, and the starting and ending angles.

[0072] Furthermore, in step S10, the pipeline types include water supply and drainage pipelines, electrical pipelines, ventilation pipelines, and fire protection pipelines. Different pipeline types correspond to different safety distance thresholds and priorities. The water supply and drainage pipelines include water supply pipelines and drainage pipelines, and the electrical pipelines include high-voltage pipelines and low-voltage pipelines.

[0073] Furthermore, in step S11, the pipeline path coordinates include the three-dimensional coordinates of all path nodes traversed by the pipeline path. After the pipeline path coordinates are written into the building information model data, the spatial position attributes and geometric attributes of the corresponding components are updated, so that the three-dimensional model can reflect the optimized pipeline layout status.

[0074] Furthermore, the specifications include construction drawing line type specifications, construction drawing dimensioning specifications, and construction drawing layer setting specifications. The construction drawing line type specifications define the line type codes corresponding to different component categories. The construction drawing dimensioning specifications define the accuracy and position of dimensioning. The construction drawing layer setting specifications define the layer names and layer display order.

[0075] The number of attribute fields for the component is derived from the component attribute information stored in the building information model data. This component attribute information includes component name, component material, component specifications, component cost, and component construction date. The number of vertices of the component is derived from the vertex coordinate data in the geometric attributes, obtained by counting the number of coordinate points in the vertex coordinate data. The number of triangular faces of the component is derived from the triangular face data in the geometric attributes, obtained by counting the number of faces in the triangular face data.

[0076] The component instance refers to multiple component copies generated in batches from a single component template through the instantiation rendering processing technology. The component instances share the same geometric data but have different spatial location attributes. The instantiation rendering processing technology can reduce video memory usage and improve rendering efficiency.

[0077] The geometric feature weights and topological feature weights are used to adjust the degree of attention the component semantic recognition network pays to different types of features. The larger the geometric feature weights, the more the component semantic recognition network pays attention to the surface shape information of the component. The larger the topological feature weights, the more the component semantic recognition network pays attention to the connection relationship information of the component. The sum of the geometric feature weights and the topological feature weights is always equal to 1.

[0078] The semantic category confidence represents the degree of certainty of the component semantic recognition network regarding the prediction result. The semantic category confidence is calculated by the softmax function of the output layer, which converts the output of the neurons in the output layer into a probability distribution. The semantic category confidence is equal to the probability value of the neuron corresponding to the predicted category.

[0079] The pipeline pair refers to two pipeline paths in the integrated pipeline layout scheme whose spatial distance is less than the safety distance threshold. The spatial distance of the pipeline pair is the Euclidean distance between the nearest point pairs on the two pipeline paths.

[0080] The drawing parameters include the projection direction, the line width value, the layer display order, and the drawing scale, wherein the drawing scale determines the scaling factor from the size of the three-dimensional model to the size of the two-dimensional drawing.

[0081] The heterogeneous computing architecture refers to a computing architecture that uses both the central processing unit and the graphics processing unit to allocate computing tasks. The heterogeneous computing architecture uses the component feature classification model to allocate computing tasks to the most suitable processor in order to improve overall computing efficiency.

[0082] Optionally, this method also provides a computer-based approach to form a design and construction drawing output efficiency optimization system based on Building Information Modeling (BIM). The computer is equipped with a readable storage medium storing program instructions. When these program instructions are executed on the computer, they perform the aforementioned BIM-based design and construction drawing output efficiency optimization method.

[0083] It should be noted that the present invention also solves the following technical problems: the structural design of component semantic recognition networks in the prior art lacks a systematic optimization method, and the selection of the number of network layers depends on manual experience or exhaustive search, which leads to the technical problem that network structure optimization is time-consuming and it is difficult to find the shallowest network configuration that meets the performance requirements. This invention optimizes the network depth search strategy using the principle of the egg-dropping problem. It treats the number of network layers as the height of a building and the network performance threshold as the critical point at which the egg breaks. A dynamic programming method is used to establish a state transition matrix to record the optimal exploration strategy under different network layers and remaining attempts. A recursive formula is used to calculate the optimal layer number that minimizes the maximum number of attempts for both broken and unbroken egg scenarios. The state transition matrix is ​​filled from bottom to top to obtain the optimal exploration path from the initial layer number to the target layer number. During the training of the component semantic recognition network, the layer configurations in the optimal exploration path are tried sequentially, and the network layers are dynamically adjusted based on the comparison between the validation set accuracy and the network performance threshold. By intelligently selecting the trial order, the computational overhead of exhaustively searching all network layer configurations is avoided. This significantly reduces the number of attempts and shortens the network structure optimization time while ensuring that the shallowest network structure that meets the network performance threshold is found. This allows for network depth optimization to be completed within an acceptable timeframe even when the network layer search range is large.

[0084] It should be noted that this invention also solves the following technical problems: Existing pipeline path planning algorithms, using traditional shortest path algorithms, only consider the single objective of minimizing path length. They fail to predict and avoid pipeline intersection conflicts during the path planning process, resulting in pipeline layout schemes with shorter paths but a large number of pipeline intersections, requiring extensive manual adjustments to meet construction specifications. This invention employs an improved shortest path pipeline optimization algorithm, introducing heuristic distance estimation and a dynamic weight adjustment mechanism. The heuristic distance estimation uses Manhattan distance as the estimated remaining distance from the current path node to the target path node, guiding the algorithm to prioritize exploring path nodes closer to the target, thus reducing the search space. The dynamic weight adjustment mechanism adjusts the weights of edge connections based on pipeline type and path history information. For edge connections traversed by multiple existing pipeline paths, the weight is increased to reduce the selection priority of that edge connection, guiding subsequent pipeline paths to choose other paths to achieve a uniform distribution of pipeline paths in space. Simultaneously, a two-layer game model is introduced, including a strategy with the objective of minimizing the total path length. The upper-level model and the lower-level model, which aims to minimize the number of pipeline intersections, are coupled through the occupancy penalty coefficient of path nodes. When the pipeline path selected by the upper-level model passes through a certain path node, the occupancy penalty coefficient of that path node increases. When calculating the number of pipeline intersections, the lower-level model applies extra weight to path nodes with high occupancy penalty coefficients, prompting subsequent pipeline paths to avoid already occupied path nodes. This allows path length optimization and intersection conflict optimization to be carried out in synergy. The resulting integrated pipeline layout scheme reduces the number of pipeline intersections while maintaining a shorter total path length, thus improving the rationality and constructability of the integrated pipeline layout.

[0085] Specifically, the principle of this invention is as follows: The invention solves this technical problem by employing deep learning technology to achieve automatic semantic recognition of components and establish a rule-driven automatic drawing generation mechanism. The component semantic recognition network extracts local geometric features from the component point cloud through convolutional layers. These geometric features contain the surface shape and spatial structure information of the components. After feature fusion through fully connected layers, a semantic category probability distribution is output. The network learns geometric feature patterns of different component categories by training on a large number of labeled component samples, thereby accurately classifying unknown components. The automatic feature extraction capability of deep learning avoids the complexity and limitations of manually designed recognition rules, enabling the component recognition process to be completed automatically without human intervention. The drafting rule base establishes a mapping relationship between semantic category labels and layer configuration parameters and line style parameters, storing the standard representation of different component categories in construction drawings. When the component semantic recognition network outputs the component category, the system automatically queries the drafting rule base to obtain the corresponding drafting parameters, and draws the two-dimensional outline of the component according to the drawing order and expression priority weight calculated by the depth buffer sorting algorithm. The line style parameters are then applied to complete the construction drawing generation. The entire process, from component recognition to parameter matching to drawing, is executed automatically by the program, eliminating the need for manual setting of layer line types for each component. The improved shortest path pipeline optimization algorithm guides the search direction through heuristic distance estimation, reducing the number of path nodes searched. The dynamic weight adjustment mechanism automatically adjusts the edge weights based on the path node occupancy, guiding subsequent pipeline paths to select other paths. The two-layer game model couples the upper-layer path length optimization with the lower-layer cross-conflict optimization through the occupancy penalty coefficient, enabling pipeline planning to automatically avoid occupied paths and meet safety distance constraints while pursuing the shortest path. The generated integrated pipeline layout scheme automatically updates the building information model data and triggers a re-execution of the construction drawing generation process until the construction drawing meets the specifications. The entire iterative optimization process is completed automatically without manual intervention. This invention achieves automatic semantic recognition through deep learning, automatic drawing generation driven by a rule base, automatic pipeline optimization through two-layer game coordination, and automatic compliance with specifications through an iterative mechanism. The various technical links support each other to form a complete automated processing chain, thus solving the technical problem of the inability to achieve automatic conversion from model to drawing when generating construction drawings using BIM.

[0086] The following provides a specific embodiment 1 of the present invention. In this embodiment, the specific implementation of S07, S09 and S11 in response to S12 is the same as described above, and will not be repeated here. The specific implementation of other steps is described in detail below.

[0087] The specific implementation of step S01 involves collecting building information model data and establishing a spatial octree index structure. The three-dimensional space is recursively divided into eight subspaces until the number of components falls below a component count threshold. The component identifiers contained within each octree node are recorded. The bounding box side length of the root node of the spatial octree index structure is also specified. The calculation formula is expressed as follows:

[0088] ;

[0089] In the formula, The bounding box side length of the root node, in meters; For Building Information Modeling data in The span value along the axial direction, in meters (m). For Building Information Modeling data in The span value along the axial direction, in meters (m). For Building Information Modeling data in The span value along the axial direction, in meters (m). This represents the function that takes the maximum value.

[0090] The specific implementation of step S02 is to obtain the viewpoint position and the view frustum parameters, determine the visibility of components based on the viewpoint position and the view frustum parameters, calculate the intersection relationship between each octree node and the view frustum, and remove completely invisible octree nodes and their contained component identifiers.

[0091] The specific implementation of step S03 involves performing view distance grading processing on the components corresponding to the retained component identifiers. A multi-level detail level coefficient is calculated based on the Euclidean distance from the component's center point to the viewpoint. This multi-level detail level coefficient is used to determine the component's rendering precision level. The Euclidean distance... The calculation formula is expressed as follows:

[0092] ;

[0093] In the formula, This is the Euclidean distance from the center point of the component to the viewpoint, in meters (m). The center point of the component Coordinates, in meters; The center point of the component Coordinates, in meters; The center point of the component Coordinates, in meters; For viewpoint position Coordinates, in meters; For viewpoint position Coordinates, in meters; For viewpoint position Coordinates, in meters. Normalized distance ratio. The calculation formula is expressed as follows:

[0094] ;

[0095] In the formula, This is the normalized distance ratio, dimensionless. This is the Euclidean distance from the center point of the component to the viewpoint, in meters (m). The distance standard value is 100m, with an empirical value of 100m. The multi-level detail coefficient... The formula for determining it is expressed as follows:

[0096] ;

[0097] In the formula, These are multi-level detail level coefficients, dimensionless, with values ​​of 1, 2, or 3, corresponding to low-precision model, medium-precision model, and high-precision model, respectively. This is the normalized distance ratio, which is dimensionless.

[0098] The specific implementation of step S04 involves using a component feature classification model to classify the components corresponding to the component rendering accuracy level into processing tasks. Based on the classification results, geometrically dense components are assigned to the graphics processor for parallel rendering, while attribute query components are assigned to the central processing unit for serial processing. The classification decision function... The statement is as follows:

[0099] ;

[0100] In the formula, The value of the classification decision function is dimensionless. The number of triangular facets of the component is dimensionless. This is the standard value for the number of triangular facets, with an empirical value of 1000. The number of vertices of the component is dimensionless. The standard value for the number of vertices is 500; The number of attribute fields for a component, dimensionless; The standard value for the number of attribute fields is 20, based on experience. When the processor is assigned an identifier, it outputs the graphics processor identifier; otherwise, it outputs the central processing unit identifier. The number of thread blocks... The calculation formula is expressed as follows:

[0101] ;

[0102] In the formula, The number of thread blocks required, dimensionless; The total number of geometry-dense components allocated to the graphics processor, dimensionless; The number of threads in a single thread block, an empirical value is 256; This represents the floor function. The design principle of this formula is to determine the computational intensity of a component by the difference between the product of geometric complexity and the amount of attribute data. Components with high geometric complexity are more suitable for parallel processing by graphics processors, while components with large amounts of attribute data are more suitable for serial queries by central processing units. This decision function achieves a unified comparison of data of different magnitudes through dimensionless normalization parameters, making the classification boundary clear and consistent with the actual hardware characteristics.

[0103] The specific implementation of step S05 involves instantiating and rendering geometrically dense components allocated to the graphics processor, extracting shared geometric data of similar components and storing it in the video memory buffer, batch generating component instances through a transformation matrix, and completing the 3D scene rendering. The transformation matrix... The calculation formula is expressed as follows:

[0104] ;

[0105] In the formula, The transformation matrix has dimensions of 4 rows and 4 columns; It is a scaling matrix with dimensions of 4 rows and 4 columns; It is a rotation matrix with dimensions of 4 rows and 4 columns; The translation matrix has dimensions of 4 rows and 4 columns; matrix multiplication is performed on the vertex coordinates from right to left. The specific representation is as follows:

[0106] ;

[0107] In the formula, The center point of the component instance Coordinates, in meters; The center point of the component instance Coordinates, in meters; The center point of the component instance Coordinates, in meters. The scaling matrix. The specific representation is as follows:

[0108] ;

[0109] In the formula, For component instances in Dimensional proportions in the axial direction, dimensionless; For component instances in Dimensional proportions in the axial direction, dimensionless; For component instances in The dimensional proportions along the axial direction are dimensionless. The rotation matrix... The orientation angle of the component instance is determined, with a dimension of 4 rows and 4 columns.

[0110] The specific implementation of step S06 involves establishing a component semantic recognition network to identify component categories in the 3D model after the 3D scene rendering is completed. The component semantic recognition network outputs the semantic category label of each component and its representation priority weight on the 2D projection plane. The weight parameters of the adaptive attention mechanism... and The calculation formula is expressed as follows:

[0111] ;

[0112] ;

[0113] In the formula, The geometric feature weights are dimensionless and range from 0 to 1. The topological feature weights are dimensionless and range from 0 to 1. The attention modulation coefficient is dimensionless and has an empirical value of 2.0. The number of triangular facets of the component is dimensionless. This is the standard value for the number of triangular facets, with an empirical value of 1000. The surface area of ​​the component, in units of ; The standard value for surface area is 10, which is an empirical value. ; It is an exponential function. The semantic category confidence level. The calculation uses the softmax function, which is described in detail below:

[0114] ;

[0115] In the formula, For the first The confidence level of each semantic category is dimensionless and ranges from 0 to 1. The output layer of the component semantic recognition network The output value of each neuron is dimensionless. The total number of predefined component categories, empirically estimated to be 50; The index is for the semantic category, with values ​​ranging from 1 to... ; It is an exponential function. This formula uses the sigmoid function to achieve a smooth weight distribution, where the sigmoid function is... When the normalized difference between geometric complexity and surface area is positive, the weight of geometric features increases, reflecting that components with high grid density but low surface area need to pay more attention to geometric details. This mechanism enables the network to adaptively extract the most discriminative features based on the geometric topology of the components by dynamically adjusting the feature weights, thereby improving the accuracy and robustness of semantic recognition.

[0116] The specific implementation of step S08 involves using a depth buffer sorting algorithm to handle component occlusion relationships, calculating the depth value of each component pixel and sorting them from farthest to nearest according to the depth values, and determining the component rendering order based on the sorting result and the expression priority weight. The rendering priority... The calculation formula is expressed as follows:

[0117] ;

[0118] In the formula, To determine the priority level, use a dimensionless value ranging from 0 to 1. The perpendicular distance from a point on the component surface to the projection plane is expressed in meters (m). This represents the maximum depth value in the scene, in meters (m). To express priority weights, the values ​​are dimensionless and range from 0 to 1. A smaller drawing priority value indicates that the component is drawn more preferentially. This formula uses a depth normalization term. and priority weight items The weighted combination achieves a synergistic consideration of depth ranking and semantic importance, ensuring that important components can still be correctly represented even when occluded.

[0119] The specific implementation of step S10 involves establishing a pipeline path node network and using an improved shortest path pipeline optimization algorithm for integrated pipeline layout. The improved shortest path pipeline optimization algorithm optimizes pipeline paths through a two-layer game model, where the objective function of the upper-layer model is... The statement is as follows:

[0120] ;

[0121] In the formula, This is the normalized value of the total path length, which is dimensionless. The quantity of pipelines is dimensionless. This is the index of the pipeline, with a value range of 1 to... ; For the first The set of edges connected by a pipeline path; The index of the edge connection; Connect the edges The length, in meters (m); The standard value for the side length is 1m, which is empirically considered to be 1m. The objective function of the lower-level model is... The statement is as follows:

[0122] ;

[0123] In the formula, This is a normalized value for the number of pipeline intersections, and is dimensionless. The quantity of pipelines is dimensionless. This is the index of the first pipeline, with a value ranging from 1 to... ; This is the index of the second pipeline, with a value range of [value missing]. to ; For the first Pipeline route and the first The number of shared path nodes for each pipeline route, dimensionless; The standard value for the number of nodes is 100, based on experience. The occupancy penalty coefficient... The update formula is expressed as follows:

[0124] ;

[0125] In the formula, The updated occupancy penalty coefficient is dimensionless. The initial value is 1, representing the occupancy penalty coefficient before the update. The penalty growth factor is dimensionless and has an empirical value of 1.5. The number of pipelines that have passed through this path node is dimensionless. The weight adjustment formula for the edge connection is expressed as follows:

[0126] ;

[0127] In the formula, The adjusted weights are in meters (m). These are the initial weight values, in meters (m). Connect the edges Number of pipelines that have passed through, dimensionless; The standard value for the number of pipelines is 5, based on experience. The heuristic distance estimation uses Manhattan distance. The formula for calculating the remaining distance estimate is as follows:

[0128] ;

[0129] In the formula, Manhattan distance, in meters; For the current path node Coordinates, in meters; For the current path node Coordinates, in meters; For the current path node Coordinates, in meters; For the target path node Coordinates, in meters; For the target path node Coordinates, in meters; For the target path node Coordinates, in meters; This represents the absolute value function. The heuristic function of the improved shortest path algorithm. The calculation formula is expressed as follows:

[0130] ;

[0131] In the formula, Path node The heuristic function value is dimensionless; The index of the path node; From the origin to the path node The actual distance, in meters; From path node Manhattan distance to the target path node, in meters; The distance is a normalized standard value, with an empirical value of 1m. This two-layer game model achieves multi-objective collaborative optimization of pipeline layout through the coupling effect of path length optimization in the upper-layer model and cross-conflict optimization in the lower-layer model. The exponential growth mechanism of the occupancy penalty coefficient effectively avoids excessive concentration of path nodes, and the dynamic weight adjustment enables the algorithm to adaptively disperse the pipeline path distribution, significantly reducing the probability of pipeline cross-conflict and improving the rationality and constructability of the pipeline layout scheme.

[0132] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team undertook the electromechanical detailed design task for a comprehensive office building project with 15 floors above ground and 3 floors underground. The total building height is 68m, and the total building area is 32,000 square meters. The project includes 1280 pipelines across water supply and drainage, electrical, HVAC, and fire protection systems. The building information model (BIM) data contains a total of 45,600 components, including 12,800 structural components, 18,500 architectural components, and 14,300 mechanical and electrical components. Traditional construction drawing methods require manual, layer-by-layer, system-by-system 2D drawing, and the integrated pipeline layout requires repeated coordination and adjustments, resulting in an overall drawing cycle of up to 45 working days. The technical team decided to adopt the optimized method of this invention to improve drawing efficiency and reduce pipeline collisions and conflicts.

[0133] The technical team first collected building information model (BIM) data. The geometric attributes of the components included approximately 5.8 million vertex coordinates and approximately 3.2 million triangular facets, with a one-to-one correspondence between normal vector data and triangular facet data. When establishing the spatial octree index structure, the minimum side length of the building space's bounding cube was determined to be 75m, and the component quantity threshold was set to 100. The recursive partitioning process generated a total of 2048 octree nodes, including 1536 leaf nodes, with each leaf node containing an average of 30 component identifiers. For example... Figure 1 As shown, the octree index structure enables subsequent view frustum pruning to quickly locate visible components.

[0134] The technical team set the viewpoint to be 20m east of the building's center and 35m upward. The viewing angle in the view frustum parameters was set to 60°, the near clipping plane distance to 0.5m, and the far clipping plane distance to 150m. Based on the intersection relationship between the view frustum and the octree nodes, a total of 856 completely invisible octree nodes were identified, resulting in the removal of 25,680 component identifiers and the retention of 19,920. The retained components underwent a view distance classification process, calculating the Euclidean distance from the component's center point to the viewpoint, ranging from 8m to 142m. Based on the normalized distance ratio, the number of components in the high-precision model was 3,580, the number in the medium-precision model was 9,650, and the number in the low-precision model was 6,690.

[0135] The technical team used a component feature classification model to classify components for processing tasks. The number of triangular faces, vertices, and attribute fields of each component were statistically analyzed, as shown in Table 1. The feature parameters of some typical components clearly exhibited either geometrically dense or attribute-query-oriented characteristics. Based on the classification decision function calculation results, a total of 12,850 geometrically dense components were allocated to the graphics processor for parallel rendering, while 7,070 attribute-query-oriented components were allocated to the central processing unit (CPU) for serial processing. The graphics processor used was an NVIDIA Tesla V100 with 5,120 CUDA cores, and the number of threads per thread block was set to 256, requiring 51 thread blocks for computation. The CPU used was an Intel Xeon E5-2680 with a clock speed of 2.8 GHz, responsible for processing material information, cost information, and construction date information queries for attribute-query-oriented components.

[0136] Table 1 Characteristic parameters of typical components

[0137] Component type Number of triangular facets Number of vertices Number of attribute fields Classification results reinforced concrete column 2840 1560 12 Geometric dense steel beam 3250 1820 15 Geometric dense air duct 580 320 8 Geometric dense water supply pipes 240 135 18 Attribute query type Cable trays 1680 920 22 Geometric dense fire sprinkler heads 95 58 25 Attribute query type Lighting fixtures 120 75 28 Attribute query type

[0138] The technical team instantiated and rendered the geometrically dense components allocated to the graphics processor. Sixty-eight similar component types were identified, including 360 standard column instances, 520 standard beam instances, and 840 standard duct instances. Shared geometric data of similar components was extracted and stored in a video memory buffer, occupying 1.8GB of video memory. Component instances were generated in batches using transformation matrices; each instance's transformation matrix was 4 rows and 4 columns, containing a composite result of translation, rotation, and scaling matrices. The 3D scene rendering achieved a frame rate of 60 frames per second, meeting the requirements for real-time interaction.

[0139] The technical team established a component semantic recognition network to identify component categories in the 3D model after 3D scene rendering. The input layer of the component semantic recognition network receives point cloud data of the components, with 2048 sampling points and a point cloud dimension of 2048 x 3. Local geometric features are extracted through three convolutional layers with 64, 128, and 256 kernels, each 3x3 in size. After the convolutional layers, a max-pooling layer with a 2x2 pooling window is connected. The two fully connected layers have 512 and 128 neurons, respectively. The output layer has 50 neurons, equal to the predefined total number of component categories. The training dataset contains 68,000 component samples with labeled semantic categories, divided into a training set of 54,400 and a validation set of 13,600 in an 8:2 ratio. The cross-entropy loss function and momentum gradient descent optimization algorithm are used, with an initial learning rate of 0.001, a momentum coefficient of 0.9, a batch size of 32, and 100 training epochs. The validation set accuracy reached a peak of 92.5% in round 78 and then remained unchanged for five consecutive rounds, at which point training terminated and the network weight parameters were saved. The component semantic recognition network outputs a semantic category label for each component, with a priority weight of 0.95 for load-bearing structural components, 0.65 for non-load-bearing components, 0.88 for components in the main functional space, and 0.52 for components in the auxiliary space.

[0140] The technical team retrieved the corresponding layer configuration parameters and line style parameters from the drafting rule library based on semantic category tags. The drafting rule library uses a key-value pair data structure and contains drafting rules corresponding to 50 semantic category tags. As shown in Table 2, different component categories correspond to different layer names, line type codes, line width values, and color codes. Structural components are represented by solid lines with a line width of 0.7mm and a color code of red (255, 0, 0). Building components are represented by solid lines with a line width of 0.5mm and a color code of black (0, 0, 0). In the electromechanical components, water supply and drainage pipelines are represented by dashed lines with a line width of 0.3mm and a color code of blue (0, 0, 255). Electrical conduits are represented by dotted-dash lines with a line width of 0.3mm and a color code of green (0, 255, 0).

[0141] Table 2. Drafting Rules for Typical Components

[0142]

[0143] The technical team employed a depth buffer sorting algorithm to handle component occlusion relationships. The depth buffer has a resolution of 3840 x 2160 pixels, with each pixel corresponding to a storage unit initialized to a maximum depth of 150m. The algorithm iterates through the projected pixels of all components, calculating the depth value for each pixel as the vertical distance from the component's surface to the projection plane. After sorting the depth values ​​from farthest to nearest, the rendering order of the components is determined by combining these depth values ​​with expression priority weights. A threshold of 0.01m is set for determining components with similar depth values; components with similar depth values ​​are then further sorted according to their expression priority weights, with components having higher expression priority weights being rendered first.

[0144] The technical team performed orthogonal projection transformation on the 3D model according to the set projection directions, including the front view direction, top view direction, and side view direction. For the front view direction, the projection matrix retained the horizontal and vertical coordinates of the components but discarded the depth coordinates. The 2D outlines of the components were drawn sequentially according to the drawing order, and line style parameters were applied to generate the construction drawings. The 2D outlines were drawn using vector graphics format, with straight line segments defined by the start and end coordinates, and arc segments defined by the center coordinates, radius, and start and end angles. The generated construction drawings include 15 floor plans, 3 basement floor plans, 4 elevation views, and 6 section views, totaling 28 construction drawings.

[0145] The technical team established a pipeline path node network for comprehensive pipeline layout. The building space was divided into uniform grids with a grid size of 0.5m, and grid intersections served as path nodes, totaling 186,000 path nodes. Edge connections were established between adjacent path nodes, with the initial weight of each edge connection being the Euclidean distance between the path nodes. Edge connections were checked for walls or other obstacles; edges crossing obstacles had their weights set to infinity to indicate impassable paths, with 38,500 such impassable edges. An improved shortest path pipeline optimization algorithm was used for comprehensive pipeline layout, incorporating heuristic distance estimation and a dynamic weight adjustment mechanism. Heuristic distance estimation used the Manhattan distance as the estimated remaining distance from the current path node to the target path node. The dynamic weight adjustment mechanism adjusted the weights of edge connections based on pipeline type and path history information; the weight adjustment factor for each edge connection was calculated by dividing the number of pipelines by a standard value of 5. Figure 2 As shown, the improved shortest path pipeline optimization algorithm significantly reduces the number of path nodes searched compared to the traditional shortest path algorithm.

[0146] The technical team employed a two-layer game theory model to optimize pipeline paths. The objective function of the upper-layer model is to minimize the normalized total path length, which is equal to the sum of the lengths of all connected edges in all pipeline paths divided by the standard edge length of 1m, divided by the number of pipelines (1280). The objective function of the lower-layer model is to minimize the normalized number of pipeline intersections, which is equal to the sum of the number of shared path nodes between any two pipeline paths divided by the standard node number of 100, divided by the number of pipelines (1280). The coupling term of the two-layer game theory model is the occupancy penalty coefficient for path nodes, initially set to 1. Each time a pipeline path passes through a path node, the penalty coefficient is multiplied by a penalty growth factor of 1.5. Figure 3 As shown, the two-layer game model enables path length optimization and intersection conflict optimization to be carried out in tandem. The resulting integrated pipeline layout scheme reduces the number of pipeline path intersections by about 70% while maintaining a shorter total path length.

[0147] The technical team conducted a safety distance check on the integrated pipeline layout scheme. The safety distance threshold for electrical conduits and water supply / drainage conduits is 0.3m, and the safety distance threshold for conduits of the same type is 0.1m. The inspection revealed 28 pairs of conduits with spatial distances less than the safety distance threshold. Among these, 18 pairs were conflicts between electrical conduits and water supply / drainage conduits, and 10 pairs were conflicts between conduits of the same type. Adjustments were made based on pipeline priority: fire protection conduits had the highest priority, followed by water supply / drainage conduits, then electrical conduits, and ventilation conduits had the lowest priority. For lower-priority conduit paths, the starting or ending points were shifted vertically by the safety distance threshold, and the improved shortest path pipeline optimization algorithm was re-executed. The adjusted integrated pipeline layout scheme satisfied all safety distance constraints.

[0148] The technical team updated the Building Information Model (BIM) data based on the integrated pipeline layout results, writing the optimized pipeline path coordinates into the geometric and spatial location attributes of the corresponding components. The updated BIM data triggered a re-execution of the process, regenerating the 3D scene rendering and construction drawings. The generated construction drawings underwent automated verification, checking for unclosed outlines, complete dimensions, and compliance with drafting rules. The automated verification identified five non-conformities: two unclosed outlines, one missing dimension, and two incorrect layer settings. After correcting the BIM data and drafting parameters based on the verification report, the optimization process was re-executed, ultimately generating 28 construction drawings that met the specifications, reducing the overall drawing turnaround time to 12 working days.

[0149] It should be noted that the technological advancements of this invention compared to traditional methods are mainly reflected in three aspects. First, the spatial octree index structure recursively divides the three-dimensional space and records the component identifiers contained in each node, enabling frustum clipping to quickly locate visible components without having to determine the visibility of each component individually. From an algorithm complexity perspective, this reduces the time complexity of component visibility determination from linear to logarithmic, significantly improving the rendering performance of large-scale building information models. Second, the component feature classification model intelligently allocates processor resources based on the geometric complexity and attribute data volume of the components. This allows geometrically dense components to fully utilize the parallel computing capabilities of the graphics processor, while attribute query components are processed serially on the central processing unit, avoiding unnecessary data transmission overhead. This heterogeneous computing architecture achieves a match between tasks and processor characteristics from the perspective of computing resource utilization, solving the problem of uneven distribution of computing resources in traditional single-processor solutions. Finally, the improved shortest path pipeline optimization algorithm optimizes the pipeline path planning process from the perspective of path search strategy by introducing heuristic distance estimation and dynamic weight adjustment mechanism. Heuristic distance estimation can guide the algorithm to explore path nodes closer to the target first, thereby reducing the search space. The dynamic weight adjustment mechanism adjusts the edge connection weights according to the path history information to avoid congestion caused by multiple pipeline paths being concentrated in a few paths. The two-level game model realizes the synergy between path length optimization and cross-conflict optimization from the perspective of multi-objective optimization. These improvements enhance the rationality and constructability of pipeline integrated layout from the algorithm principle level.

[0150] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing the efficiency of design and construction drawing output based on BIM, characterized in that, This process includes: collecting building information model (BIM) data and establishing a spatial octree index structure; obtaining viewpoint positions and view frustum parameters and determining component visibility; performing distance-level processing on retained components; classifying components using a component feature classification model and assigning them to a graphics processor (GPU) or central processing unit (CPU); instantiating and rendering components assigned to the GPU; establishing a component semantic recognition network to identify component categories in the 3D model and output semantic category labels and expression priority weights; querying the drafting rule base based on semantic category labels to obtain layer configuration parameters and line style parameters; using a depth buffer sorting algorithm to handle component occlusion relationships and determine the drawing order; performing orthogonal projection transformation on the 3D model according to the projection direction and sequentially drawing the 2D outlines of the components to complete the construction drawing generation; establishing a pipeline path node network and using an improved shortest path pipeline optimization algorithm for integrated pipeline layout; updating BIM data based on the integrated pipeline layout results and triggering re-execution until the construction drawings meet the specifications.

2. The method according to claim 1, characterized in that, Building Information Modeling (BIM) data includes the geometric attributes and spatial location attributes of components. Geometric attributes include the vertex coordinates, triangular facets, and normal vectors of components. Spatial location attributes include the center point coordinates and the bounding box boundary coordinates of components.

3. The method according to claim 2, characterized in that, The spatial octree index structure recursively divides the cubic space into eight sub-cubes. Each octree node stores a list of component identifiers that fall within its spatial range, and the number of components contained in a leaf node does not exceed a component count threshold.

4. The method according to claim 3, characterized in that, The parameters of the view frustum include viewpoint coordinates, line-of-sight vector, view angle, near clipping plane distance, and far clipping plane distance. The method for determining the intersection relationship between the view frustum and the octree node is to extract the coordinates of the eight vertices of the octree node's bounding box and project the coordinates of the eight vertices onto the six clipping planes of the view frustum.

5. The method according to claim 4, characterized in that, The calculation method for the multi-level detail level coefficient is to calculate the Euclidean distance from the center point of the component to the viewpoint, divide the Euclidean distance by the distance standard value to obtain the normalized distance ratio, and determine the multi-level detail level coefficient based on the interval in which the normalized distance ratio is located.

6. The method according to claim 5, characterized in that, The component feature classification model uses a classification decision function for judgment. When the product of the number of triangles divided by the standard value of the number of triangles and the number of vertices divided by the standard value of the number of vertices is greater than the number of attribute fields divided by the standard value of the number of attribute fields, the processor allocation identifier is output as the graphics processor identifier; otherwise, the processor allocation identifier is output as the central processing unit identifier.

7. The method according to claim 6, characterized in that, The parallel rendering process of the graphics processor adopts a thread block partitioning strategy, which calculates the required number of thread blocks based on the total number of geometry-dense components allocated to the graphics processor and the number of threads in a single thread block.

8. The method according to claim 7, characterized in that, The shared geometry data for instantiated rendering processing includes an array of vertex positions, an array of normal vectors, and an array of texture coordinates. The transformation matrix is ​​a composite result of translation, rotation, and scaling matrices.

9. The method according to claim 8, characterized in that, The structure of the component semantic recognition network is as follows: the input layer receives the point cloud data of the component, and the local geometric features are extracted through three convolutional layers. The number of convolutional kernels in each convolutional layer is 64, 128 and 256 respectively, and the kernel size is 3 by 3. After the convolutional layers, a max pooling layer is connected to reduce the feature dimension, and the pooling window size is 2 by 2.

10. The method according to claim 9, characterized in that, The component semantic recognition network adopts the principle of the egg-dropping problem to optimize the network depth search strategy. The number of network layers is regarded as the height of a building, and the network performance threshold is regarded as the critical point at which the egg breaks. The optimal layer exploration sequence is determined by dynamic programming to find the shallowest network structure that meets the network performance threshold with the fewest training times.

Images