A feature enhancement neural network-based complex mechanical and electrical pipeline BIM collision detection method and system

Abstract

This invention relates to the field of Building Information Modeling (BIM), and discloses a method and system for collision detection of complex electromechanical pipelines based on a feature-enhanced neural network. The method includes: processing the electromechanical pipeline model file using a multi-source BIM data fusion and geometric-semantic joint feature extraction method to obtain initial feature representations of pipeline components; performing multi-layer convolution on the initial feature representations using a dynamic heterogeneous graph neural network method to obtain enhanced feature representations of the pipeline components; performing two-stream interaction on multiple pipeline component pairs in a candidate collision pair set based on the enhanced feature representations to obtain a collision pair set; and continuously optimizing the pipeline avoidance path based on the collision pair set to obtain a collision-free BIM model. This invention achieves high-precision identification and intelligent avoidance of multiple types of collisions in complex electromechanical scenarios.

Description

Technical Field

[0001] This invention relates to the field of building information technology, and in particular to a BIM collision detection method and system for complex electromechanical pipelines based on feature-enhanced neural networks. Background Technology

[0002] With the continuous expansion of building construction scale and the increasing complexity of electromechanical systems, Building Information Modeling (BIM) technology has been widely applied in the collaborative design and construction management of multiple professional pipelines, including HVAC, water supply and drainage, electrical systems, and fire protection. In the detailed design phase of complex electromechanical pipelines, various professional pipelines are densely intersecting within a limited space. Clash detection, as a core component to ensure construction quality and reduce rework costs, places extremely high demands on detection accuracy and processing efficiency.

[0003] Existing BIM collision detection methods for MEP (Mechanical, Electrical, and Plumbing) pipelines primarily rely on geometric bounding box (AABB) intersection tests, determining collision relationships by whether the bounding boxes of two components overlap. However, these methods only utilize the geometric dimensions of components for collision determination, completely ignoring semantic attributes such as pipeline system type, installation priority, and specification constraints. This leads to a large number of misclassified process connection nodes within the same MEP system as collision conflicts. Furthermore, they suffer from serious omissions in soft collision cases where the clearance is insufficient but the geometry does not intersect, failing to meet the accurate identification requirements of multiple collision types in complex MEP scenarios. Summary of the Invention

[0004] This invention provides a BIM collision detection method and system for complex electromechanical pipelines based on feature-enhanced neural networks, in order to overcome the shortcomings of existing technologies.

[0005] This invention provides a BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks, comprising: S1: The electromechanical pipeline model file is processed by multi-source BIM data fusion and geometric semantic joint feature extraction method to obtain the initial feature representation of pipeline components; S2: The initial feature representation is subjected to multi-layer convolution using a dynamic heterogeneous graph neural network method to obtain an enhanced feature representation of the pipeline component; S3: Based on the enhanced feature representation, perform two-flow interaction on multiple pipeline component pairs in the candidate collision pair set to obtain the collision pair set; S4: Continuously optimize the pipeline avoidance path based on the collision pair set to obtain a collision-free BIM model.

[0006] According to the present invention, a BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks is provided, wherein step S1 further includes: S11: Hierarchical voxel meshing is performed on the 3D space of the electromechanical pipeline model file using the octree adaptive voxelization method to obtain the standardized local point set within the non-empty voxels. S12: Extract the semantic attributes of the standardized local point set and the electromechanical pipeline model file through the geometric semantic joint embedding method to obtain the initial feature representation and spatial index structure of the pipeline components.

[0007] According to the present invention, a BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks is provided, wherein step S12 further includes: S121: Input the semantic attributes of the electromechanical pipeline model file into the learnable embedding layer to perform discrete semantic label mapping and obtain a continuous vector; S122: Local geometric features are extracted from the standardized local point set by using a multi-scale set abstraction layer based on PointNet++ to obtain local geometric features; S123: The continuous vector and the local geometric features are spliced ​​and fused in the channel dimension to obtain the initial feature representation and spatial index structure of the pipeline component.

[0008] According to the BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks provided by the present invention, step S122 further includes: S1221: Perform a sphere query on the standardized local point set based on multiple preset neighborhood radii to obtain a multi-scale local point cloud neighborhood set; S1222: For the multi-scale local point cloud neighborhood set, the geometric shape including pipeline endpoints, bend nodes and straight pipe sections is used as a priori spatial mask input to the PointNet++-based set abstraction layer to perform mask attention feature aggregation on the multi-scale local point cloud neighborhood set to obtain mask-aware multi-scale local features. S1223: A curvature normal vector joint feature module is constructed based on pipeline curvature change, cross-sectional profile and spatial orientation information. The curvature gradient is calculated and the normal is uniformly projected on the multi-scale local features to obtain local geometric features.

[0009] According to the present invention, a BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks is provided, wherein step S2 further includes: S21: Construct a dynamic heterogeneous scene graph based on the initial feature representation and the corresponding spatial location information; wherein, the dynamic heterogeneous scene graph includes pipeline component nodes, adjacent edges, and system topology edges; S22: Multi-layer message passing is performed on the dynamic heterogeneous scene graph through a multi-head attention feature-enhanced graph convolution algorithm to obtain multi-layer graph convolution output features; S23: The enhanced feature representation of pipeline components is obtained by weighted summation of the output features of multi-layer graph convolution through a multi-scale skip-connection fusion mechanism.

[0010] According to the present invention, a BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks is provided, wherein step S22 further includes: S221. Input the feature vectors of the neighboring edges and the system topology edges into independent linear transformation layers for dimension mapping to obtain edge features of a unified dimension. S222. Based on the edge features of a unified dimension, a multi-head attention mechanism is used to perform message passing calculation on the neighbor set of nodes in the dynamic heterogeneous scene graph under multiple heterogeneous edges, so as to obtain the convolution output features of each layer of graph.

[0011] According to the BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks provided by the present invention, step S3 further includes: S31: Based on the spatial index structure corresponding to the initial feature representation, perform axial alignment bounding box overlap test screening on spatially adjacent pipeline component pairs to obtain a candidate collision pair set; S32: Through dual-stream feature interaction, semantic symmetric interaction features and geometric precision features are calculated for pipeline components in the candidate collision pair set, and then concatenated with the normative constraint difference features and input into the classification head network to obtain the collision type probability distribution. S33: Classify and filter the candidate collision pair set according to the probability distribution of the collision type to obtain the collision pair set.

[0012] According to the BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks provided by the present invention, step S32 further includes: S321: Input the enhanced features of the two pipeline components in the candidate collision pair set into the semantic symmetric interaction flow, perform element-wise product and absolute difference calculation and multilayer perceptron compression processing to obtain semantic symmetric interaction features; S322: The original parameters of the two pipeline components are calculated by projection interval overlap and multilayer perceptron encoding through the differentiable separable axis theorem algorithm to obtain geometric precision features; S323: The semantic symmetric interaction features, the geometric calculation features, and the normative constraint difference features are concatenated and input into the classification head network to obtain the collision type probability distribution.

[0013] According to the BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks provided by the present invention, step S4 further includes: S41: Sort the collision pair set in descending order according to the comprehensive priority score to obtain an ordered processing queue; S42: The state space, action space and reward function of the pipeline avoidance path in the ordered processing queue are continuously optimized by the near-end policy optimization algorithm to obtain the corrected pipeline axis parameters and bend coordinates; S43: Write back the corrected pipeline axis parameters and elbow coordinates to the IFC model, and perform incremental local verification on the local sub-maps corresponding to the modified area to obtain a collision-free BIM model.

[0014] This invention provides a BIM clash detection system for complex electromechanical pipelines based on feature-enhanced neural networks, used to execute a BIM clash detection method for complex electromechanical pipelines based on feature-enhanced neural networks as described above, comprising: Extraction module: Used to process electromechanical pipeline model files through multi-source BIM data fusion and geometric semantic joint feature extraction method to obtain the initial feature representation of pipeline components; Convolution module: used to perform multi-layer convolution on the initial feature representation using a dynamic heterogeneous graph neural network method to obtain an enhanced feature representation of the pipeline components; Interaction module: used to perform two-stream interaction on multiple pipeline component pairs in the candidate collision pair set according to the enhanced feature representation, to obtain the collision pair set; Avoidance module: used to continuously optimize the pipeline avoidance path based on the collision pair set to obtain a collision-free BIM model.

[0015] This invention provides a BIM collision detection method and system for complex electromechanical pipelines based on feature-enhanced neural networks. By jointly encoding the geometric and semantic attributes of the electromechanical pipeline model file into an initial feature representation, the subsequent detection process simultaneously perceives the shape information and system identity information of the pipeline. This fundamentally avoids the problem of numerous misjudgments of process connection nodes within the same system as collisions due to neglecting semantic constraints, significantly reducing the false detection rate. Secondly, the introduction of heterogeneous scene graphs in this invention enables the classification and modeling of proximity relationships and system topology relationships. Combined with multi-head attention graph convolution and multi-scale hop connection fusion mechanisms, the enhanced feature representation of each pipeline component fully encodes the interaction information within the multi-hop spatial neighborhood, thereby enabling the model to detect collisions in the spatial neighborhood. The system exhibits enhanced global perception capabilities in areas with complex layouts and densely intersecting pipelines, effectively reducing missed detections due to insufficient local information. Furthermore, the introduction of the differentiable separating axis theorem in the dual-flow feature interaction module allows the penetration depth vector to participate in training as a continuous gradient signal, significantly improving the model's resolution in discriminating between hard and soft collision critical states. It can effectively identify soft collisions and gap conflicts where the net distance is insufficient but the geometry does not intersect. Finally, the path replanning framework based on near-end strategy optimization, combined with incremental local verification closed loop, ensures that the avoidance scheme does not introduce secondary collisions while meeting specification constraints, outputting a verified collision-free BIM model. Overall, this significantly shortens the manual review cycle for the detailed design of electromechanical pipelines. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A schematic diagram of the process for a BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks provided by the present invention; Figure 2 This invention provides a schematic diagram of a complex electromechanical pipeline BIM collision detection system based on a feature-enhanced neural network. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0019] To better understand this invention, the technical terms appearing in the embodiments of this invention will be explained below.

[0020] BIM (Building Information Modeling) is a digital building representation technology that uses a three-dimensional digital model as a carrier to integrate geometric information, material information, system ownership information, and engineering specification constraints of building components. Each component in a BIM model not only has a three-dimensional geometric shape, but also carries structured semantic attributes related to engineering design, construction, and operation and maintenance, supporting multi-disciplinary collaborative design and clash detection.

[0021] IFC (Industry Foundation Classes) is an open standard format for exchanging building information modeling (BIM) data, with the .ifc file extension. IFC files store the geometric representations (such as IfcPolyline and IfcShapeRepresentation), spatial relationships (such as IfcRelConnectsElements), and attribute sets of building and MEP (Mechanical, Electrical, and Plumbing) components in a hierarchical object structure, supporting model data interoperability between different BIM platforms such as Revit and MagiCAD. This invention uses IFC files as the unified input format for MEP pipeline model data.

[0022] An octree is a three-dimensional spatial recursive partitioning data structure. Starting from a root cube containing all target objects, an octree divides the current cube into eight equal sub-cubes along the X, Y, and Z axes. Each sub-cube containing objects continues to be recursively subdivided until a preset termination condition is met (e.g., the number of objects in a leaf node is below a threshold). Each leaf node of an octree corresponds to a voxel unit in three-dimensional space. Through hierarchical indexing, spatial proximity queries can be performed in logarithmic time complexity, avoiding the need to perform pairwise distance calculations for all objects.

[0023] PointNet++ is a hierarchical deep neural network for extracting local features from 3D point cloud data. The core operational unit of PointNet++ is the ensemble abstraction layer, which performs three operations sequentially: First, it selects several center points from the input point cloud through farthest point sampling, ensuring that the center points are distributed as evenly as possible across the point cloud space; then, it uses a ball query operation to collect all neighboring points within a fixed radius around each center point; finally, it applies a PointNet mini-network (composed of a multilayer perceptron and max pooling) to each set of neighboring points to extract local features, outputting a local feature vector represented by each center point. By stacking multiple ensemble abstraction layers, PointNet++ progressively expands the receptive field, gradually aggregating local detail features into global shape features.

[0024] Multilayer Perceptron (MLP): A feedforward neural network structure consisting of multiple fully connected layers stacked on top of each other. Each fully connected layer multiplies the input vector by a learnable weight matrix and adds a bias vector, then transforms it through a non-linear activation function (such as ReLU) before outputting it to the next layer. MLP maps the input feature vector to an output feature vector of a specified dimension through alternating layers of linear transformation and non-linear activation, and is widely used in feature compression, feature fusion, and classification output scenarios.

[0025] Graph Neural Networks (GNNs) are deep neural networks that take graph-structured data as input. A graph consists of a set of nodes and a set of edges. Each node carries a feature vector, and each edge may optionally carry an edge feature vector. GNNs iteratively update node features through a message-passing mechanism: in each iteration, each node collects the features of its neighbors and connected edges, aggregates them using an aggregation function, and then merges them with its own features to complete a feature update. After multiple iterations, the feature representation of each node encodes the structural information within its multi-hop neighborhood.

[0026] Heterogeneous graph: A graph structure containing multiple types of nodes or edges. Unlike homogeneous graphs, which contain only a single type of node and edge, heterogeneous graphs have edges of different types carrying different semantic relational information. During message passing, independent feature transformations must be performed on each type of edge to preserve the independent semantics of each type of relation. The dynamic heterogeneous scene graph constructed in this invention includes two types of heterogeneous edges: proximity edges and system topology edges, which respectively encode the spatial proximity relationships between pipeline components and the connection topology relationships within the electromechanical system.

[0027] The Separating Axis Theorem (SAT) is a geometric algorithm for determining whether two convex polyhedra intersect. Its mathematical basis is: if there exists an axis such that the projection intervals of the two convex polyhedra on this axis do not overlap, then the two polyhedra do not intersect; if the projection intervals on all candidate separating axes overlap, then the two polyhedra intersect. The set of candidate separating axes includes the normal vectors of each face of the two polyhedra and the cross product vector of the direction vectors of each edge of the two polyhedra. Based on this, the differentiable separating axis theorem expresses the amount of overlap of the projection intervals as a continuously differentiable function, allowing the penetration depth value to be used as a gradient signal in the backpropagation training of a neural network.

[0028] Proximal Policy Optimization (PPO) is a reinforcement learning algorithm based on policy gradients. Reinforcement learning is based on the interaction between an agent and its environment. The agent selects an action based on its current state, and the environment returns a reward value and a new state. The agent updates the policy network parameters with the goal of maximizing the cumulative reward. PPO introduces a cutoff ratio mechanism during policy gradient updates, constraining the probability ratio of the new and old policies within a preset range. This prevents a sudden drop in policy performance due to excessively large single parameter updates, thus achieving a balance between exploring new actions and maintaining policy stability.

[0029] BCF (BIM Collaboration Format) is an open standard for exchanging clash detection problem reports, developed by the buildingSMART international organization. Files with the .bcf extension are used to store the spatial coordinates, component identifiers, problem description, processing status, and 3D viewport screenshots of clash problems in a structured XML format, supporting interoperability of clash problem information between different BIM platforms.

[0030] The embodiments of the present invention are described below with reference to the figures.

[0031] like Figure 1 As shown, this invention provides a BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks, comprising: S1: The electromechanical pipeline model file is processed by multi-source BIM data fusion and geometric semantic joint feature extraction method to obtain the initial feature representation of pipeline components.

[0032] In step S1, the present invention first exports IFC format MEP (Mechanical, Electrical, and Pipeline) model files from mainstream BIM platforms including Revit and MagiCAD. These model files cover component data from multiple professional systems, including HVAC ducts, water supply and drainage pipes, electrical cable trays, and fire protection pipelines. The files are then parsed. After parsing the IFC files, the geometric attributes (including pipe diameter, pipe length, cross-sectional shape, spatial coordinates, and direction vector) and semantic attributes (including system type, material, installation priority, and design specification constraints) of each pipeline component are extracted. These attributes are then uniformly encoded into structured feature vectors, which serve as the raw data input for subsequent voxelization and feature extraction.

[0033] Step S1 further includes: S11: The three-dimensional space of the electromechanical pipeline model file is divided into hierarchical voxel meshes using the octree adaptive voxelization method to obtain the standardized local point set within the non-empty voxels.

[0034] Furthermore, the octree adaptive voxelization method described in this invention aims to recursively divide the three-dimensional space into a spatial index structure of hierarchical cubic units. Specifically, this invention uses the three-dimensional scene space containing all pipeline components as the root node, and adaptively recursively subdivides it into eight child nodes according to the pipeline spatial distribution density, until the number of pipeline point clouds in the leaf node voxels meets a preset threshold. For each non-empty leaf node voxel, this invention subtracts the voxel center coordinates from the pipeline point cloud coordinates in the voxel and divides by the voxel side length to complete local coordinate normalization, obtaining a standardized local point set with coordinate values ​​normalized to the range of [-1, 1]. Finally, the above octree structure simultaneously records the spatial mapping relationship between each leaf node voxel and the pipeline component, forming a spatial index structure for quick querying during subsequent candidate collision pair screening.

[0035] S12: Extract the semantic attributes of the standardized local point set and the electromechanical pipeline model file through the geometric semantic joint embedding method to obtain the initial feature representation and spatial index structure of the pipeline components.

[0036] Step S12 further includes: S121: Input the semantic attributes of the electromechanical pipeline model file into the learnable embedding layer to perform discrete semantic label mapping and obtain a continuous vector.

[0037] In step S121, the present invention first inputs the semantic attributes (system type, material, installation priority, and specification constraint parameters) of each pipeline component parsed from the IFC file as discrete category labels into a learnable embedding layer. This learnable embedding layer maps discrete integer indices to a parameterized matrix of fixed-dimensional continuous real-valued vectors. Each discrete semantic category corresponds to a row vector in the matrix, which is adjusted during training as the gradient is updated. Subsequently, the present invention concatenates the embedding vectors corresponding to each semantic attribute, outputting a 128-dimensional continuous vector for each pipeline component, which encodes semantic information such as the component's system affiliation, material category, and installation priority.

[0038] S122: Local geometric features are extracted from the standardized local point set by using a multi-scale set abstraction layer based on PointNet++.

[0039] Furthermore, PointNet++ is a hierarchical local feature extraction network for 3D point cloud data, with its core operation being an ensemble abstraction layer. For this ensemble abstraction layer, it first selects several center points through farthest point sampling, then uses sphere queries to collect neighborhood point clouds within a fixed radius around each center point, and finally performs PointNet micro-network extraction on the neighborhood point clouds to extract local features. This invention improves upon this by simultaneously performing sphere queries with multiple preset neighborhood radii to extract multi-scale local geometric features of pipeline endpoints, bend nodes, and straight pipe sections from a standardized set of local points.

[0040] Step S122 further includes: S1221: Perform a sphere query on the standardized local point set based on multiple preset neighborhood radii to obtain a multi-scale local point cloud neighborhood set.

[0041] Specifically, this invention performs spherical query operations on standardized local point sets using three neighborhood radii of 0.05m, 0.15m, and 0.30m. This involves retrieving all points in the point cloud that fall within a spherical neighborhood, using a central point as the sphere's center and a preset radius as the sphere's radius. These three different radii correspond to three levels of neighborhood ranges: the local fine structure of the pipeline (endpoints and bends), the medium-scale cross-sectional profile, and the large-scale spatial orientation. This results in a multi-scale local point cloud neighborhood set containing point sets at these three scales.

[0042] S1222: For the multi-scale local point cloud neighborhood set, the geometric shapes including pipeline endpoints, bend nodes and straight pipe sections are used as prior spatial mask inputs to the PointNet++-based set abstraction layer to perform mask attention feature aggregation on the multi-scale local point cloud neighborhood set, thereby obtaining mask-aware multi-scale local features.

[0043] After obtaining the multi-scale local point cloud neighborhood set, in step S1222 of this invention, the geometric shape categories of pipeline endpoints, bend nodes, and straight pipe sections are labeled as prior spatial masks and input together with the multi-scale local point cloud neighborhood set into a PointNet++-based ensemble abstraction layer. The prior spatial mask, in binary matrix form, marks which geometric shape region each neighborhood point belongs to. Based on this, when performing feature aggregation on each neighborhood point, the ensemble abstraction layer assigns differentiated attention weights to points in different regions using the mask values, giving higher weight to points belonging to geometrically complex regions such as endpoints and bend nodes in feature aggregation, thereby outputting mask-aware multi-scale local features.

[0044] S1223: A curvature normal vector joint feature module is constructed based on pipeline curvature change, cross-sectional profile and spatial orientation information. The curvature gradient is calculated and the normal is uniformly projected on the multi-scale local features to obtain local geometric features.

[0045] After obtaining the mask-aware multi-scale local features, this invention further processes them through a curvature-normal-vector joint feature generation module. Specifically, this module first performs principal component analysis on the neighborhood point set of each center point in the multi-scale local features, estimating the normal vector using the eigenvector corresponding to the smallest eigenvalue. Then, it calculates the angular change in the normal vector between adjacent center points as the curvature gradient, and projects the normal vector of each center point onto the reference axis direction of its corresponding straight or curved pipe segment, calculating the projection consistency score as the normal consistency projection value. After calculation, the curvature gradient and normal consistency projection value are concatenated with the multi-scale local features, ultimately outputting local geometric features that encode pipeline curvature changes, cross-sectional profiles, and spatial orientation information.

[0046] S123: The continuous vector and the local geometric features are spliced ​​and fused in the channel dimension to obtain the initial feature representation and spatial index structure of the pipeline component.

[0047] In step S123, the present invention directly concatenates the 128-dimensional continuous vector output in step S121 with the local geometric features output in step S1223 along the channel dimension, that is, the two vectors are joined end to end into a longer vector, resulting in a 256-dimensional initial feature representation for each pipeline component. This representation carries both geometric shape description and system semantic information. Simultaneously, the octree spatial index structure constructed in step S11 is output together with the initial feature representation as a spatial query tool, for use during candidate collision pair enumeration in step S3.

[0048] Current collision detection methods typically separate geometric processing from semantic analysis. In the feature extraction stage, only the 3D bounding box or point cloud geometric features of the pipeline are extracted, with the attribute table only being consulted in the final judgment stage. This delayed processing approach, prioritizing geometry over semantics, has serious flaws. In complex electromechanical scenarios, there are numerous process connection nodes between straight pipes and elbows within the same system (such as fire protection pipe networks). Traditional pure geometric feature extraction methods cannot distinguish between real collisions across systems and process connections within the same system, misclassifying many normal connection nodes as collision conflicts. This results in an extremely high false detection rate and a significant increase in unnecessary manual review work.

[0049] Based on this, the present invention breaks with convention in step S1, pioneering the introduction of a geometric-semantic joint embedding mechanism at the front-end feature extraction stage. By mapping discrete semantic labels such as system type and priority into continuous vectors through a learnable embedding layer, and directly concatenating them with local geometric features in the channel dimension, the initial feature representation of pipeline components is transformed from a traditional pure geometric shape descriptor into a semantically perceptive descriptor carrying identity information. The inventive approach of pre-fusion of semantic information in this invention enables the subsequent neural network to perceive the system affiliation of the pipeline from the feature layer when performing receptive field aggregation and collision discrimination, fundamentally eliminating the root cause of misjudging process connections within the same system as collisions.

[0050] S2: The initial feature representation is subjected to multi-layer convolution using a dynamic heterogeneous graph neural network method to obtain an enhanced feature representation of the pipeline component.

[0051] Step S2 further includes: S21: Construct a dynamic heterogeneous scene graph based on the initial feature representation and the corresponding spatial location information; wherein, the dynamic heterogeneous scene graph includes pipeline component nodes, adjacent edges, and system topology edges.

[0052] In step S21, this invention treats each pipeline component as a node, using its 256-dimensional initial feature representation as the node's initial feature. Edge construction is divided into two categories: Neighboring edges are established by using an adaptive threshold determined by whether the shortest distance between the axes of two pipeline components is less than the sum of the outer diameters of the two pipelines and the installation spacing margin to establish an undirected connection. The relative distance, relative angle, estimated cross-sectional projection interference area, and normal penetration component are encoded as a 32-dimensional neighboring edge feature vector. System topology edges are established by using the IfcRelConnectsElements relationship in the IFC file to establish directed connections between adjacent components within the same electromechanical system. The installation priority difference, the cosine of the pipeline routing angle, and the specification constraint coefficient are encoded as system topology edge feature vectors. Finally, the two types of nodes and edges together constitute a dynamic heterogeneous scene graph.

[0053] S22: The dynamic heterogeneous scene graph is subjected to multi-layer message passing through a multi-head attention feature-enhanced graph convolution algorithm to obtain multi-layer graph convolution output features.

[0054] Step S22 further includes: S221. Input the feature vectors of the neighboring edges and the system topology edges into independent linear transformation layers for dimension mapping to obtain edge features of a unified dimension.

[0055] Since the neighboring edge feature vector is 32-dimensional and the system topology edge feature vector is 16-dimensional, the two types of edge features have inconsistent dimensions and cannot directly participate in the unified message passing calculation. Therefore, in step S221 of this invention, the two types of edge feature vectors are respectively input into their respective independent linear transformation layers, that is, each is multiplied by a learnable weight matrix, so that the two types of edge features are mapped to the same unified dimension, resulting in edge features of unified dimension, which then participate in subsequent message passing.

[0056] This invention argues that after constructing a scene graph encompassing spatial proximity and system topology, existing conventional graph neural networks typically concatenate or simply add all types of edge features, employing isomorphic graph convolution for unified message passing. However, this indiscriminate isomorphic processing suffers from severe information confusion. Proximity edge features primarily characterize physical spatial distances and geometric interference quantities (such as relative distances and penetration components), while system topology edge features primarily characterize engineering logic constraints (such as priority differences and canonical coefficients). Their numerical dimensions and physical meanings are drastically different. Direct mixing and aggregation can dilute spatial distance information with topological logic information, even generating mutually interfering gradient noise, preventing the model from accurately perceiving real collision threats in densely intersecting pipeline areas.

[0057] To address this deficiency, this invention innovatively designs a heterogeneous edge independent mapping and multi-head attention isolation mechanism in step S22. By inputting neighboring edges and system topology edges into independent linear transformation layers, not only is dimension alignment achieved, but also physical isolation and semantic decoupling of the two types of heterogeneous relationships are performed in the feature space. Subsequently, in the multi-head attention mechanism, attention weights are calculated separately for different types of edges and message passing is performed independently, ensuring that spatial proximity constraints and system topology constraints do not interfere with each other during graph convolution, thereby significantly improving the model's global perception accuracy and anti-interference capability in complex and dense pipeline environments.

[0058] S222. Based on the edge features of a unified dimension, a multi-head attention mechanism is used to perform message passing calculation on the neighbor set of nodes in the dynamic heterogeneous scene graph under multiple heterogeneous edges, so as to obtain the convolution output features of each layer of graph.

[0059] Furthermore, the multi-head attention mechanism aims to scale the dot product of the query vector and the key vector and then normalize it using softmax to obtain an attention calculation method that weights the features of each neighboring node. "Multi-head" refers to executing multiple independent attention calculations in parallel and then concatenating the results. In step S222, this invention calculates attention weights for two types of heterogeneous edges—nearby edges and system topology edges—on a dynamic heterogeneous scene graph. Weighted feature aggregation is performed on the neighbor sets of each node under both types of edges. The neighboring node features are concatenated with the corresponding edge features of the same dimension, multiplied by a learnable weight matrix, and then weighted and summed using 8-head attention weights. This sum is then concatenated with the current node's own features and transformed by a multilayer perceptron to obtain the graph convolution output features of the node in the current layer. Finally, this invention stacks four layers of the above graph convolutions, outputting the graph convolution output features of each pipeline component layer by layer, so that the receptive field of each node covers a spatial neighborhood of approximately 3 to 4 hops.

[0060] Furthermore, the expression for message passing in multi-head attention heterogeneous graph convolutions is: Since step S21 uses each pipeline component as a node in the dynamic heterogeneous scene graph during construction, and the initial features of the node represent the initial features of the pipeline component, there is a strict one-to-one mapping relationship between the two. Therefore, in the formulas of this invention, the subscripts... and This represents both the node number in the heterogeneous scene graph and the pipeline component number corresponding to that node; both representations point to the same entity object. When graph structure operations are involved (such as message passing and attention weight calculation), and It participates in graph convolution operations as a node; when geometric calculations are involved (such as axis coordinates, cross-sectional polygon vertices, and installation priority), and Provide raw parameters as a pipeline component.

[0061] in, This is the network layer index value; For nodes In the Layer feature representation; For nodes In the The set of neighbor nodes under the same edge class; The category index values ​​of the edges correspond to neighboring edges and system topology edges, respectively; For the first Layer The learnable weight matrix corresponding to each edge class; For nodes With nodes Between Feature vectors of edge classes; This represents a vector concatenation operation; For the first Layer Nodes under the same edge For neighboring nodes The attention weights are calculated using the following formula; MLP represents the multilayer perceptron nonlinear transformation.

[0062] The expression for calculating head attention weights is: in, For the first The query projection matrix of the layer; For the first The bond projection matrix of the layer; For each attention head's feature dimension, divide the total feature dimension by the number of attention heads, which is 8. Indicates a node All neighboring nodes The scores are normalized.

[0063] S23: The enhanced feature representation of pipeline components is obtained by weighted summation of the output features of multi-layer graph convolution through a multi-scale skip-connection fusion mechanism.

[0064] In step S23, the present invention inputs the features output by the four layers of graph convolution (i.e., the graph convolution output features from layers 1 to 4) into their respective independent linear projection layers. After mapping to the same dimension, the four layers of features are weighted and summed using a set of learnable hierarchical attention weights. These hierarchical attention weights are normalized using softmax to ensure that the sum of the weights is 1. Subsequently, the weighted sum is linearly transformed to output a 512-dimensional enhanced feature representation. This representation integrates shallow local geometric information and deep global topological information, serving as the input data for the dual-stream feature interaction module in step S3.

[0065] Furthermore, the expression for multi-scale skip-weighted fusion is: in, Pipeline components The 512-dimensional enhanced feature representation; For the first The independent linear projection matrix corresponding to each layer maps the output features of each layer to a unified dimension; For the first The attention weights of the learnable layers corresponding to each layer are summed to 1 after being normalized by softmax.

[0066] S3: Based on the enhanced feature representation, perform two-flow interaction on multiple pipeline component pairs in the candidate collision pair set to obtain the collision pair set.

[0067] Step S3 further includes: S31: Based on the spatial index structure corresponding to the initial feature representation, perform axial alignment bounding box overlap test screening on spatially adjacent pipeline component pairs to obtain a candidate collision pair set.

[0068] Furthermore, the Axially Aligned Bounding Box (AABB) is a geometric representation of a three-dimensional object that is approximately enclosed by a cuboid parallel to the coordinate axes. It is determined by six parameters, namely the minimum and maximum values ​​of the object in the X, Y, and Z axes.

[0069] In step S31, the present invention utilizes the octree spatial index structure from step S1 to enumerate component pairs for all pipeline components within each leaf node voxel and its 26 adjacent voxels. Then, for each component pair, the AABBs of each component are calculated, and it is determined whether the projection intervals of the two AABBs overlap simultaneously in the X, Y, and Z axes. If the projection intervals of all three axes overlap, the AABBs are considered to intersect. Component pairs whose AABBs do not intersect but whose shortest distance between component axes is less than an adaptive threshold are also retained. After the above filtering, component pairs that meet the conditions are retained, forming a candidate collision pair set. Compared to the quadratic computational cost of performing pairwise enumeration on all components, the octree index compresses the candidate pair size to a logarithmic linear level.

[0070] S32: Through dual-stream feature interaction, semantic symmetric interaction features and geometric precision features are calculated for pipeline components in the candidate collision pair set, and then concatenated with the normative constraint difference features and input into the classification head network to obtain the collision type probability distribution.

[0071] Step S32 further includes: S321: Input the enhanced features of the two pipeline components in the candidate collision pair set into the semantic symmetric interaction flow, perform element-wise product and absolute difference calculation and multilayer perceptron compression processing to obtain semantic symmetric interaction features.

[0072] In step S321, based on the semantic symmetric interactive flow method, this invention receives the 512-dimensional enhanced feature representations of each of the two pipeline components in the candidate collision pair, denoted as feature vector i and feature vector j, respectively. Subsequently, this invention performs element-wise multiplication on the two vectors to obtain a 512-dimensional element-wise product feature, which reflects the degree of collaboration between the two components in each feature dimension. Then, the absolute value of the element-wise difference between the two vectors is calculated to obtain a 512-dimensional absolute difference feature, which reflects the degree of difference between the two components in each feature dimension.

[0073] After obtaining the absolute difference feature, this invention concatenates the 512-dimensional element-product feature and the 512-dimensional absolute difference feature in the channel dimension into a 1024-dimensional vector, and inputs it into a three-layer multilayer perceptron. The output dimensions of the three layers are 1024-dimensional, 512-dimensional, and 256-dimensional, respectively. Each layer undergoes a nonlinear transformation through an activation function, and finally compresses and outputs a 256-dimensional semantic symmetric interaction feature. This feature encodes the compatibility of the two pipeline system types, the relative relationship of installation priorities, and the similarity of cross-sectional shapes.

[0074] S322: The original parameters of the two pipeline components are calculated by projection interval overlap and processed by multilayer perceptron encoding through the differentiable separable axis theorem algorithm to obtain geometric precision features.

[0075] Furthermore, the Separating Axis Theorem (SAT) employed in this invention aims to determine whether two convex polyhedra intersect. The principle is that if there exists an axis such that the projection intervals of the two objects on that axis do not overlap, then the two objects do not intersect; conversely, if the projection intervals on all candidate separating axes overlap, then the two objects intersect. In this case, the amount of overlap of the projection intervals on each axis is the penetration depth component. The Differentiable Separating Axis Theorem, based on this, calculates the projection interval overlap in a differentiable manner, allowing the penetration depth value to participate in backpropagation as a continuous gradient signal.

[0076] Specifically, in step S322, this invention takes the axis equations, cross-sectional polygon vertex sets, and spatial transformation matrix of the two pipeline components in the candidate collision pair as input, and enumerates all candidate separation axes. Each candidate separation axis includes the normal vector of each face of the two components and the cross product vector of the direction vectors of each side of the two components. For each candidate separation axis, this invention projects the cross-sectional polygon vertex sets of the two pipeline components onto that axis, takes the maximum and minimum values ​​of their respective projections to determine the projection interval, and calculates the overlap between the two projection intervals. If the projection intervals do not overlap, the overlap is zero. Subsequently, the overlap of all candidate separation axes is combined to form a penetration depth vector containing six components: the penetration depth along the three axes and the cosine of the included angles along the three axes. Finally, this penetration depth vector is input into an independent multilayer perceptron for encoding, outputting a 64-dimensional geometric precision feature.

[0077] Furthermore, the expression for the overlap of the projection intervals on a single candidate separation axis is: in, For the two pipeline components in the first Candidate separation shafts The amount of overlap of the projection intervals on the surface can be negative or zero; the larger the absolute value, the deeper the penetration. , Pipeline components With components The coordinate vector of the midpoint of the axis; For the first The unit direction vector of each candidate separation axis, the set of candidate separation axes includes the normal vector of each face of the two components and the cross product vector of the direction vectors of each side of the two components; , Components With components exist The projection radius on the axis, i.e., the maximum absolute value of the projection of the vertices of the respective cross-sectional polygons; when When it means there is no penetration in the direction of that axis, when Its absolute value is the penetration depth component along that axis. This is combined with the results of all candidate separation axes. The values ​​are combined with the penetration depth along the three axes and the cosine of the angles between the three axes to form a 6-dimensional penetration depth vector. .

[0078] In refined collision type classification, most existing deep learning-based collision detection methods employ end-to-end pure data-driven networks (e.g., directly inputting point cloud pairs into a 3D convolutional network to output probabilities). This pure black-box model suffers from a difficult-to-overcome technical bottleneck: it can effectively identify hard collisions (intersecting objects), but for the critical state of soft collisions (non-intersecting objects but with insufficient net distance), due to the lack of explicit geometric physical quantity constraints, the model can only rely on the distribution probability of the training data to guess, resulting in extremely low resolution in discriminating the critical state of net distance and often missing soft collisions.

[0079] To overcome the ambiguity of purely data-driven models, this invention creatively introduces the differentiable separating axis theorem in step S322. The traditional separating axis theorem (SAT) is a discrete algorithm used in computational geometry to determine the intersection of convex bodies, outputting only black-and-white Boolean values, which cannot participate in the backpropagation of neural networks. This invention transforms the traditional Boolean intersection determination into the calculation of a continuous penetration depth vector by rewriting the calculation logic of the projection interval overlap with a continuously differentiable function. This invention injects precise prior knowledge of geometric physics (separating axis projection) as an explicit constraint into the neural network, allowing the penetration depth vector to directly participate in the backpropagation of the loss function as a continuous gradient signal. Based on this, the model no longer performs probabilistic fitting of fuzzy boundaries during training, but is instead guided by precise geometric gradients, greatly improving the network's perception of the critical states of hard and soft collisions, and achieving accurate quantitative identification of gap conflicts such as insufficient net distance.

[0080] S323: The semantic symmetric interaction features, the geometric calculation features, and the normative constraint difference features are concatenated and input into the classification head network to obtain the collision type probability distribution.

[0081] In step S323, the present invention sequentially concatenates 256-dimensional semantic symmetric interaction features, 64-dimensional geometric calculation features, and 16-dimensional specification constraint difference features along the channel dimension to obtain a 336-dimensional joint feature vector. The specification constraint difference features are calculated from the component pairs specification parameters extracted in step S1, including 16 scalar values: minimum installation spacing specification value, maintenance space requirement, and structural reserved opening size constraint. Subsequently, the 336-dimensional joint feature vector is input into a three-layer classification head network. The three layers output dimensions are 256-dimensional, 128-dimensional, and C-dimensional, respectively, with C=4 corresponding to the four categories of hard collision, soft collision, gap conflict, and no collision. Finally, after Softmax normalization, the probability distribution of each component pair belonging to the four collision types is output.

[0082] For collision types, hard collision refers to the geometric intersection of the solid components and the penetration depth is greater than zero; soft collision refers to the clear distance between components being greater than zero but not meeting the minimum spacing required by the installation specifications; gap conflict refers to insufficient clear distance between pipelines and reserved openings in the structure; and no collision refers to the component spacing meeting all specification requirements.

[0083] S33: Classify and filter the candidate collision pair set according to the probability distribution of the collision type to obtain the collision pair set.

[0084] In step S33, the present invention takes the category corresponding to the maximum probability distribution of the collision type for each component pair in the candidate collision pair set as the collision type label, and uses the maximum probability value as the confidence score, while retaining the penetration depth vector output in step S322. Subsequently, component pairs with collision type labels of hard collision, soft collision, or gap conflict are selected to form a collision pair set, which is then input into step S4 for path optimization processing.

[0085] During the training phase of the classification head network, considering the extremely sparse nature of collision samples in electromechanical engineering (hard collisions typically account for less than 2%) and the significant differences in the severity of different collision types, this invention designs a semantically aware hierarchical focus loss function to optimize the classification head network parameters. The standard cross-entropy loss function assigns equal weights to all samples. When trained on real-world engineering datasets with very few collision samples and a vast majority of non-collision samples, the classification head network tends to predict all component pairs as non-collision to minimize the overall loss, resulting in extremely low recall rates for hard and soft collisions. The focus loss function introduces a modulation factor based on the standard cross-entropy loss function. This allows the classification head network to automatically reduce the loss weight of easily classified samples that can be correctly classified, and concentrate gradient updates on collision samples that are difficult to classify.

[0086] This invention further introduces a class weight coefficient based on the focus loss. With semantic priority penalty coefficient : Based on the reciprocal of the frequency of each collision type in the training set, rare categories such as hard collisions are given higher weight in the loss calculation. The normalized installation priority value of pipeline components and the L2 norm of the penetration depth vector are incorporated into the loss calculation. This significantly amplifies the loss term corresponding to hard collisions with large penetration depths in high-priority pipelines (such as fire mains), forcing the classification head network to prioritize the detection accuracy of high-safety-level collisions during parameter updates. These three factors together constitute the semantic-aware hierarchical focus loss function shown in the formula below. The classification head network uses the gradient signal of this loss function to perform backpropagation and parameter updates until the collision detection accuracy on the validation set converges.

[0087] Furthermore, the expression for the semantically aware hierarchical focus loss function is: in, This represents the total number of candidate collision pairs in the set. This is an index for collision types, corresponding to hard collision, soft collision, gap collision, and no collision in sequence. For component pairs In category The actual labels on the screen are encoded using one-hot encoding. For classifying head network components Predicted as category The probability output; These are the category weight coefficients; The category-adaptive focusing index is set to 1.5 for hard collisions, 2.0 for soft collisions, 2.5 for gap conflicts, and 0.5 for no collisions. This is the semantic priority penalty coefficient.

[0088] The expression for the category weight coefficient is as follows: in, For categories in the training set The total number of samples; The total number of collision types is set to 4; The loss weight is calculated based on the reciprocal of the frequency of each category in the training set, giving higher loss weights to rare categories such as hard collisions.

[0089] The expression for the semantic priority penalty coefficient is as follows: in, This is an indicator function that takes a value of 1 only when the collision type is hard or soft, and a value of 0 for other types. , The components extracted in step S1 are respectively With components The installation priority normalization value is close to 1 for high-priority systems such as fire pipelines and main air ducts, and close to 0 for low-priority systems such as low-voltage cable trays. Let L2 norm be the penetration depth vector; The adaptive collision detection threshold; This is a priority penalty intensity hyperparameter, with a default value of 1.5.

[0090] S4: Continuously optimize the pipeline avoidance path based on the collision pair set to obtain a collision-free BIM model.

[0091] Step S4 further includes: S41: Sort the collision pair set in descending order according to the comprehensive priority score to obtain an ordered processing queue.

[0092] Furthermore, in step S41, the present invention first calculates a comprehensive priority score for each component pair in the collision pair set. This score consists of a weighted sum of four factors: the L2 norm of the penetration depth vector (reflecting the severity of the collision), the normalized value of the higher installation priority among the two components, the detection confidence score output in step S33, and the degree centrality of the component pair in the heterogeneous scene graph constructed in step S21. The degree centrality refers to the proportion of edges directly connected to a node in the graph structure to the maximum possible number of connections in the entire graph. Nodes with higher degree centrality have a greater cascading influence on surrounding pipelines. Subsequently, the four scores are multiplied by configurable weight coefficients and summed to obtain the comprehensive priority score for each component pair. Then, all collision component pairs are arranged in descending order of this score to form an ordered processing queue.

[0093] S42: The state space, action space and reward function of the pipeline avoidance path in the ordered processing queue are continuously optimized by the near-end policy optimization algorithm to obtain the corrected pipeline axis parameters and bend coordinates.

[0094] Furthermore, the aforementioned proximal policy optimization is a reinforcement learning algorithm based on policy gradient. The core mechanism is to limit the change range between the old and new policies at each policy update, and to constrain the policy update step size by means of a truncation ratio, thereby achieving a balance between exploration and stability.

[0095] In step S42, this invention uses the heterogeneous graph encoding of the current collision scenario as a basis, and adds the spatial coordinates of the pipeline to be adjusted, the remaining number of collisions, and the consumed adjustment margin to jointly form a state vector input strategy network. The strategy network outputs the probability distribution of three types of actions: axis translation action, which selects discrete displacements in the three-axis direction with a step size of 0.05m; elbow insertion action, which inserts a standard elbow at a specified position and updates the downstream pipeline direction vector; and pipe diameter downgrading action, which replaces large-diameter pipelines with a multi-path small-diameter parallel arrangement scheme under the premise of satisfying flow constraints.

[0096] After each action is performed, the present invention calculates a reward value, which consists of four components: the reduction in the L2 norm of the penetration depth vector output in step S322 as a collision elimination reward; the increase in the total pipeline length as a negative penalty; whether the net distance between the adjusted components meets the minimum installation spacing specification as a spacing satisfaction reward; and whether the new path collides with structural components as a structural collision penalty. During execution, the policy network continuously updates its parameters based on the accumulated reward signal, continuously optimizing the path replanning process. The pre-trained policy network converges on average after no more than 50 iterations in the new scenario, and finally outputs the corrected pipeline axis parameters and elbow coordinates that meet the specification constraints.

[0097] Furthermore, the path replanning reward function The expression is: in, For collision elimination rewards, The amount of reduction in the L2 norm of the penetration depth vector before and after the action is performed; the greater the reduction, the higher the reward. Penalty for pipeline length increments, The increment of the total pipeline length after the action is performed relative to the original length; the larger the increment, the greater the penalty. To meet the spacing requirements and receive rewards, This refers to the actual net distance between the two components after the action is performed. The value is 1 if the minimum installation spacing required by the specification is met, otherwise it is 0. For structural collision penalties, This represents the newly planned path and structural components; a value of 1 is used when a collision occurs, otherwise a value of 0 is used. , , , These are the weighting coefficients for each reward item.

[0098] S43: Write back the corrected pipeline axis parameters and elbow coordinates to the IFC model, and perform incremental local verification on the local sub-maps corresponding to the modified area to obtain a collision-free BIM model.

[0099] Furthermore, this invention writes back the corrected pipeline axis parameters output in step S42 to the IfcPolyline field of the corresponding component in the IFC model, and writes back the elbow coordinates to the IfcShapeRepresentation field, completing the geometric update of the IFC model. Subsequently, this invention locates the leaf node and its neighboring leaf nodes corresponding to the modified area in the octree structure of step S1, and only re-executes the feature extraction and collision detection process of steps S1 to S3 on the pipeline components in the above local area, extracting the local subgraph of the modified area and recalculating the enhanced feature representation and collision detection results of the affected nodes. If the local verification result shows that there are still residual collision component pairs in the modified area, the component pair is added back to the ordered processing queue, triggering a new round of optimization in step S42; if the local verification result shows that all component pairs in the modified area pass the collision detection, the corresponding component pair is removed from the collision pair set and marked as resolved. After all collision component pairs have been processed, the final output is a collision-free BIM model confirmed by incremental verification and a complete collision detection and optimization processing report.

[0100] The collision-free BIM model generated by this invention can be used to guide the prefabrication and on-site installation of electromechanical pipeline components without rework, and serves as a cross-platform, non-destructive, interoperable digital twin base for subsequent engineering quantity calculation and operation and maintenance management. At the same time, the complete collision detection and optimization report generated by this invention includes a three-level hierarchical aggregation result based on collision type, professional system, and floor zoning, as well as a standard BCF format file. The BCF format file is automatically filled with structured fields such as component identifier, collision spatial coordinates, penetration depth value, specification item number, and suggested processing priority, and is overlaid with a 3D viewport screenshot that uses color mapping to distinguish the collision level and a 3D arrow indicating the penetration direction.

[0101] This invention provides a BIM collision detection system for complex electromechanical pipelines based on feature-enhanced neural networks, comprising: Extraction module 100: Used to process electromechanical pipeline model files through multi-source BIM data fusion and geometric semantic joint feature extraction method to obtain the initial feature representation of pipeline components; Convolution module 200: used to perform multi-layer convolution on the initial feature representation through a dynamic heterogeneous graph neural network method to obtain an enhanced feature representation of the pipeline component; Interaction module 300: used to perform two-stream interaction on multiple pipeline component pairs in the candidate collision pair set according to the enhanced feature representation, to obtain the collision pair set; Avoidance module 400: used to continuously optimize the pipeline avoidance path based on the collision pair set to obtain a collision-free BIM model.

[0102] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0103] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0104] This invention represents a paradigm shift from passive detection to proactive closed-loop optimization, completely overcoming the technical bottlenecks in in-depth design. By deeply integrating reinforcement learning-based intelligent optimization with incremental local verification, this invention breaks through the efficiency bottleneck of traditional manual adjustments to pipeline paths. It eliminates real conflicts in milliseconds while strictly controlling the risk of secondary collisions caused by path modifications. The fully automated closed-loop mechanism of this invention not only reduces the trial-and-error cost of massive pipeline layout to an extremely low level but also endows the BIM model with adaptive correction capabilities, providing a highly reliable intelligent underlying engine for achieving zero-collision, seamless delivery of large and complex building projects.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A feature-enhanced neural network-based complex electromechanical pipeline BIM collision detection method, characterized in that, include: S1: The electromechanical pipeline model file is processed by multi-source BIM data fusion and geometric semantic joint feature extraction method to obtain the initial feature representation of pipeline components; S2: The initial feature representation is subjected to multi-layer convolution using a dynamic heterogeneous graph neural network method to obtain an enhanced feature representation of the pipeline component; S3: Based on the enhanced feature representation, perform two-flow interaction on multiple pipeline component pairs in the candidate collision pair set to obtain the collision pair set; S4: Continuously optimize the pipeline avoidance path based on the collision pair set to obtain a collision-free BIM model.

2. The complex mechanical and electrical pipeline BIM collision detection method based on feature enhancement neural network according to claim 1, characterized in that, Step S1 further includes: S11: Hierarchical voxel meshing is performed on the 3D space of the electromechanical pipeline model file using the octree adaptive voxelization method to obtain the standardized local point set within the non-empty voxels. S12: Extract the semantic attributes of the standardized local point set and the electromechanical pipeline model file through the geometric semantic joint embedding method to obtain the initial feature representation and spatial index structure of the pipeline components.

3. The complex mechanical and electrical pipeline BIM collision detection method based on feature enhancement neural network according to claim 2, characterized in that, Step S12 further includes: S121: Input the semantic attributes of the electromechanical pipeline model file into the learnable embedding layer to perform discrete semantic label mapping and obtain a continuous vector; S122: Local geometric features are extracted from the standardized local point set by using a multi-scale set abstraction layer based on PointNet++ to obtain local geometric features; S123: The continuous vector and the local geometric features are spliced ​​and fused in the channel dimension to obtain the initial feature representation and spatial index structure of the pipeline component.

4. The complex mechanical and electrical pipeline BIM collision detection method based on feature enhancement neural network according to claim 3, characterized in that, Step S122 further includes: S1221: Perform a sphere query on the standardized local point set based on multiple preset neighborhood radii to obtain a multi-scale local point cloud neighborhood set; S1222: For the multi-scale local point cloud neighborhood set, the geometric shape including pipeline endpoints, bend nodes and straight pipe sections is used as a priori spatial mask input to the PointNet++-based set abstraction layer to perform mask attention feature aggregation on the multi-scale local point cloud neighborhood set to obtain mask-aware multi-scale local features. S1223: A curvature normal vector joint feature module is constructed based on pipeline curvature change, cross-sectional profile and spatial orientation information. The curvature gradient is calculated and the normal is uniformly projected on the multi-scale local features to obtain local geometric features.

5. The complex mechanical and electrical pipeline BIM collision detection method based on feature enhancement neural network according to claim 1, characterized in that, Step S2 further includes: S21: Construct a dynamic heterogeneous scene graph based on the initial feature representation and the corresponding spatial location information; wherein, the dynamic heterogeneous scene graph includes pipeline component nodes, adjacent edges, and system topology edges; S22: Multi-layer message passing is performed on the dynamic heterogeneous scene graph through a multi-head attention feature-enhanced graph convolution algorithm to obtain multi-layer graph convolution output features; S23: The enhanced feature representation of pipeline components is obtained by weighted summation of the output features of multi-layer graph convolution through a multi-scale skip-connection fusion mechanism.

6. The BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks according to claim 5, characterized in that, Step S22 further includes: S221. Input the feature vectors of the neighboring edges and the system topology edges into independent linear transformation layers for dimension mapping to obtain edge features of a unified dimension. S222. Based on the edge features of a unified dimension, a multi-head attention mechanism is used to perform message passing calculation on the neighbor set of nodes in the dynamic heterogeneous scene graph under multiple heterogeneous edges, and obtain the convolution output features of each layer of the graph.

7. The BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks according to claim 1, characterized in that, Step S3 further includes: S31: Based on the spatial index structure corresponding to the initial feature representation, perform axial alignment bounding box overlap test screening on spatially adjacent pipeline component pairs to obtain a candidate collision pair set; S32: Through dual-stream feature interaction, semantic symmetric interaction features and geometric precision features are calculated for pipeline components in the candidate collision pair set, and then concatenated with the normative constraint difference features and input into the classification head network to obtain the collision type probability distribution. S33: Classify and filter the candidate collision pair set according to the probability distribution of the collision type to obtain the collision pair set.

8. The BIM collision detection method for complex electromechanical pipelines based on feature-enhanced neural networks according to claim 7, characterized in that, Step S32 further includes: S321: Input the enhanced features of the two pipeline components in the candidate collision pair set into the semantic symmetric interaction flow, perform element-wise product and absolute difference calculation and multilayer perceptron compression processing to obtain semantic symmetric interaction features; S322: The original parameters of the two pipeline components are calculated by projection interval overlap and processed by multilayer perceptron encoding through the differentiable separable axis theorem algorithm to obtain geometric precision features; S323: The semantic symmetric interaction features, the geometric calculation features, and the normative constraint difference features are concatenated and input into the classification head network to obtain the collision type probability distribution.

9. The BIM clash detection method for complex electromechanical pipelines based on feature-enhanced neural networks according to claim 1, characterized in that, Step S4 further includes: S41: Sort the collision pair set in descending order according to the comprehensive priority score to obtain an ordered processing queue; S42: The state space, action space and reward function of the pipeline avoidance path in the ordered processing queue are continuously optimized by the near-end policy optimization algorithm to obtain the corrected pipeline axis parameters and bend coordinates; S43: Write back the corrected pipeline axis parameters and elbow coordinates to the IFC model, and perform incremental local verification on the local sub-maps corresponding to the modified area to obtain a collision-free BIM model.

10. A BIM clash detection system for complex electromechanical pipelines based on feature-enhanced neural networks, used to execute the BIM clash detection method for complex electromechanical pipelines based on feature-enhanced neural networks as described in any one of claims 1-9, characterized in that, include: Extraction module: Used to process electromechanical pipeline model files through multi-source BIM data fusion and geometric semantic joint feature extraction method to obtain the initial feature representation of pipeline components; Convolution module: used to perform multi-layer convolution on the initial feature representation using a dynamic heterogeneous graph neural network method to obtain an enhanced feature representation of the pipeline components; Interaction module: used to perform two-stream interaction on multiple pipeline component pairs in the candidate collision pair set according to the enhanced feature representation, to obtain the collision pair set; Avoidance module: used to continuously optimize the pipeline avoidance path based on the collision pair set to obtain a collision-free BIM model.

Images