Building collision detection method based on BIM cooperative multi-agent reinforcement learning

Abstract

The invention discloses a building collision detection method based on BIM cooperative multi-agent reinforcement learning, and belongs to the technical field of building engineering. The method comprises the following steps: firstly, acquiring geometric data and pipeline distribution information from a preset building information model, and determining model block granularity through analysis to obtain a sub-model set; then, a multi-agent system is adopted to distribute monitoring responsibilities, edge geometrical shape data and physical connection point locations are extracted, potential conflict areas with overlapped boundaries are judged, and a collision probability distribution diagram is generated; exchanging real-time data by setting an information synchronization period between intelligent agents to obtain local conflict distribution details, introducing real-time change data to train the intelligent agents by adopting a reinforcement learning algorithm, and updating a monitoring priority and a cooperative action sequence; and if the conflict points which are not solved still exist, local collision detection is carried out by adopting a distributed computing framework, and finally, efficient and accurate building information model collision detection is realized by adaptively adjusting sub-model division and task allocation rules.

Description

Technical Field

[0001] The present invention relates to the field of construction engineering technology, and in particular to a construction collision detection method based on BIM collaborative multi-agent reinforcement learning. Background Art

[0002] In large-scale construction projects, with the ever-increasing scale and complexity of projects, traditional clash detection methods are no longer able to meet the demands for efficiency and accuracy. Implementing dynamic clash detection for large-scale building information models (BIMs) in complex and ever-changing construction environments has become a key technical challenge that needs to be addressed. This challenge involves multiple aspects, including model segmentation, multi-agent collaboration, and real-time data processing. First, how to determine the appropriate model segmentation granularity based on pipeline density and regional complexity to balance computational efficiency and detection accuracy? Second, how to coordinate multiple agents in a distributed environment to efficiently exchange information and perform clash analysis between sub-models? Third, how to dynamically adjust detection strategies and priorities to ensure the timeliness and accuracy of detection results in the face of real-time changes during construction? Furthermore, when dealing with high-probability collision zones, how to precisely locate specific collision points and conduct in-depth analysis? Finally, how to integrate clash detection results with real-time construction data to achieve adaptive optimization of detection strategies? These sub-problems collectively constitute the technical challenges of dynamic clash detection for large-scale BIMs. The core issue is how to achieve efficient, accurate, and real-time clash detection in complex and ever-changing environments to support the smooth progress of large-scale construction projects. Summary of the Invention

[0003] The purpose of the present invention is to provide a construction collision detection method based on BIM collaborative multi-agent reinforcement learning. The method adopts a reinforcement learning algorithm to introduce real-time changing data to train the agents, update the monitoring priority and collaborative action sequence; if there are still unresolved conflict points, a distributed computing framework is used for local collision detection. Finally, by adaptively adjusting the sub-model division and task allocation rules, efficient and accurate building information model collision detection is achieved to solve the problems raised in the above background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a construction collision detection method based on BIM collaborative multi-agent reinforcement learning, comprising the following steps:

[0005] S1: Obtain geometric data and pipeline distribution information from the preset building information model, determine the granularity range of the initial model block by analyzing the pipeline density and regional complexity, and obtain the sub-model set after block division;

[0006] S2: For the sub-model set, a multi-agent system is used to assign monitoring responsibilities according to the complexity of the sub-models. Edge geometry data and physical connection points are extracted from each sub-model. Information is exchanged through a defined data exchange format to determine potential conflict areas with overlapping boundaries and obtain a preliminary collision probability distribution map.

[0007] S3: If there is a high-probability conflict area in the preliminary collision probability distribution map, obtain detailed geometric parameters and real-time construction change data from the high-probability conflict area, and determine the specific spatial location data of the collision point in the conflict area through boundary feature extraction and connection relationship analysis;

[0008] S4: Based on the spatial location data of the collision point, the specific sub-models involved are queried. By setting an information synchronization period between agents to exchange real-time data, the physical connection points between the associated sub-models of the collision point are determined to obtain the details of the local conflict distribution.

[0009] S5: Based on the details of local conflict distribution, a reinforcement learning algorithm is used to introduce real-time changing data to train the agent, update the monitoring priority weights and collaborative action sequences, and determine the adjusted agent behavior pattern;

[0010] S6: Extract edge geometry data from the sub-model set based on the adjusted agent behavior pattern. The agents exchange data according to the updated collaborative action sequence to determine the conflict area with overlapping boundaries and obtain the updated collision probability distribution map.

[0011] S7: If there are still unresolved conflict points in the updated collision probability distribution map, the spatial location data of the unresolved conflict points are obtained from the distribution map, the involved sub-models are determined through the associated sub-model query, and computing resources are allocated using the distributed computing framework to obtain the local collision detection results;

[0012] S8: For the local collision detection results, the intelligent agent exchanges data according to a set period to fuse the detection results and the collision probability distribution map, and determines whether the detection accuracy reaches the preset detection accuracy threshold to obtain the final local collision detection report;

[0013] S9: Based on the final local collision detection report, extract real-time change data from the construction environment data, and determine the block granularity range and intelligent agent monitoring responsibility allocation plan for the next detection by adaptively adjusting the sub-model division basis and task allocation rules.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] The present invention first obtains geometric data and pipeline distribution information from a preset building information model, determines the model block granularity through analysis, and obtains a sub-model set; then adopts a multi-agent system to allocate monitoring responsibilities, extracts edge geometry data and physical connection points, determines potential conflict areas with overlapping boundaries, and generates a collision probability distribution map; for high-probability conflict areas, obtains detailed parameters and real-time construction change data to determine the specific collision point location; by setting an information synchronization cycle between agents to exchange real-time data, the local conflict distribution details are obtained, and a reinforcement learning algorithm is used to introduce real-time change data to train the agents, and the monitoring priority and collaborative action sequence are updated; if there are still unresolved conflict points, a distributed computing framework is used for local collision detection, and finally, by adaptively adjusting the sub-model division and task allocation rules, efficient and accurate building information model collision detection is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] See also Figure 1 A construction collision detection method based on BIM collaborative multi-agent reinforcement learning in this embodiment includes the following steps:

[0019] S1. Obtain geometric data and pipeline distribution information from the preset building information model, determine the granularity range of the initial model block by analyzing the pipeline density and regional complexity, and obtain the sub-model set after block division.

[0020] The geometric data and pipeline distribution information are obtained through the preset building information model to obtain a complete data set for the initial model. The extracted geometric data and pipeline distribution information are used to analyze the pipeline density and obtain a density distribution map. Based on the density distribution map and regional complexity, the range of the block granularity is determined to obtain a set of granularity thresholds. The initial model is divided using the granularity threshold set to obtain a preliminary set of sub-models. For the preliminary set of sub-models, the degree of matching between pipeline density and regional complexity is analyzed to determine the boundary adjustment requirements of the sub-models and obtain an optimized set of sub-models. After obtaining the optimized set of sub-models, the consistency of the geometric data and pipeline distribution is verified through data extraction to obtain the final set of sub-models. The machine learning algorithm K-means is used to cluster the final set of sub-models to determine the spatial correlation between sub-models and obtain a classification set of block results.

[0021] Specifically, when extracting geometric data and pipeline distribution information from the preset building information model, the model file format (such as IFC or Revit) is first parsed to obtain the geometric properties (such as coordinates, dimensions, shapes) and pipeline information (such as diameter, material, and connection relationships) of components such as walls, floor slabs, and pipes.

[0022] For example, a building model contains 1,000 pipelines, each with a diameter ranging from 50 mm to 300 mm. Spatial indexing algorithms (such as R-trees) are used to organize these pipeline data into a spatial data structure for easy analysis. Next, a density clustering algorithm (such as DBSCAN) is used to analyze pipeline density, setting the neighborhood radius ε to 2 meters and the minimum number of samples (MinPts) to 5 to identify areas with dense pipelines.

[0023] For example, an area containing 10 pipelines within a 2-meter radius is marked as a high-density area. Furthermore, by calculating regional complexity indicators (such as the number of pipeline intersections and the diversity of component types), regional complexity is divided into three levels: low, medium, and high.

[0024] For example, an area containing 20 pipeline intersections and five component types is classified as a high-complexity area. Based on the pipeline density and regional complexity, an adaptive block algorithm is used to determine the granularity range of the initial model block.

[0025] For example, high-density and high-complexity areas use smaller block granularity (such as 5m x 5m), while low-density and low-complexity areas use larger block granularity (such as 20m x 20m). Ultimately, a spatial segmentation algorithm (such as a quadtree or octree) is used to divide the building model into several sub-model sets. Each sub-model contains geometric data and pipeline information within a specific range, providing a basis for subsequent refined analysis or optimization.

[0026] S2. For the sub-model set, a multi-agent system is used to allocate monitoring responsibilities according to the complexity of the sub-models. Edge geometry data and physical connection points are extracted from each sub-model. Information is exchanged through a defined data exchange format to determine potential conflict areas with overlapping boundaries and obtain a preliminary collision probability distribution map.

[0027] The multi-agent system partitions the sub-model set, and the monitoring responsibilities of each agent are determined based on the complexity allocation results, resulting in a responsibility allocation table. The edge geometry information of each sub-model is extracted from the responsibility allocation table, and a preset algorithm is used to calculate the shape data and physical connection points to obtain a connection point set. For this connection point set, the relationship between edge geometry and physical connections is processed using a data exchange format, and areas of boundary overlap are determined to obtain an overlapping area set. Based on the overlapping area set, the spatial distribution characteristics of the conflict areas are analyzed to determine the density of potential conflicts and obtain a conflict density map. If high-density areas exist in the conflict density map, the conflict areas are clustered using the K-means algorithm to obtain a clustered conflict classification set. Using the conflict classification set, the strategy for extracting points with overlapping boundaries is adjusted to determine the optimized physical connection points, resulting in an adjusted point distribution. The adjusted point distribution is used to verify the spatial consistency of the sub-model set, resulting in a final collision probability distribution map.

[0028] Specifically, when implementing a multi-agent system to allocate monitoring responsibilities, computing resources are first dynamically allocated based on the complexity of the submodels. For example, a weight allocation algorithm based on entropy weighting is used to assign a weight of 35 to submodel A, which contains 1,000 triangles, while a weight of 15 is assigned to submodel B, which contains only 200 triangles. When extracting edge geometry data using the OpenCASCADE geometry kernel, the 235 boundary edges extracted from submodel A are simplified using the Douglas-Peucker algorithm, retaining 187 key feature edges after setting a tolerance of 5 mm. Physical connection points are identified using a feature extraction method based on curvature analysis. The center coordinates of six bolt holes in the connection flange are detected with an accuracy of ±1 mm. Data exchange uses a custom JSON format, which contains geometric entity types, vertex coordinate sets (e.g., [13, 46, 79]), and a topological relationship matrix. During the boundary conflict detection phase, the AABB hierarchical bounding box algorithm is used for preliminary screening. Accurate GJK distance calculations are performed on overlapping areas (e.g., the X-axis overlap interval [104mm, 107mm]). Areas with a minimum distance less than the safety threshold of 2mm are marked as conflicting. Monte Carlo simulations are used to calculate collision probabilities. Over 2000 pose perturbation samplings, the frequency of collisions in a specific assembly area was determined to be 326, corresponding to a probability of 13%. The resulting distribution is presented as a heat map, with red highlighted areas (probability >15%) matching actual assembly difficulty areas with a 92% accuracy rate.

[0029] S3. If there is a high-probability conflict area in the preliminary collision probability distribution map, obtain detailed geometric parameters and real-time construction change data from the high-probability conflict area, and determine the specific spatial position data of the collision point in the conflict area through boundary feature extraction and connection relationship analysis.

[0030] Geometric parameters and real-time data are obtained from high-probability conflict zones. Boundary features are processed using a pre-set algorithm to determine the spatial distribution characteristics of the conflict zone, generating a boundary feature set. Connectivity data is extracted from the boundary feature set, and a clustering algorithm is used to analyze the density of the conflict zone, determine the distribution characteristics of potential collision points, and determine a candidate collision point set. For this candidate collision point set, the impact of construction changes on spatial location is verified using real-time data to obtain an adjusted collision point distribution. Based on this adjusted collision point distribution, the matching degree between boundary features and connectivity is analyzed to determine the dynamic trend of the conflict zone and obtain trend data. Based on this trend data, updated geometric parameters are obtained. Interpolation is used to ensure spatial continuity and determine the optimized collision point coordinates. Based on the optimized collision point coordinates, the boundary overlap of the conflict zone is analyzed using connectivity to determine the overlapping area distribution. Based on the overlapping area distribution, the consistency of the spatial location with the real-time data is determined to obtain the final collision point spatial location data.

[0031] Specifically, the DBSCAN algorithm, based on point cloud density clustering, was used to extract geometric parameters within high-probability conflict zones. A neighborhood radius of 3 mm and a minimum point count threshold of 8 were set to isolate feature-dense regions from the point cloud data within the conflict zone. For example, clustering of the point cloud of a flange edge resulted in 12 feature clusters. Real-time construction change data was collected at a 2Hz frequency using a laser scanner. Combined with the ICP point cloud registration algorithm, the newly acquired 3,000 point clouds were aligned with the design model, with a registration error within 5mm. During the boundary feature extraction phase, the 42 identified characteristic edges within the conflict zone were adaptively segmented using Gaussian curvature, with a curvature threshold set to 15mm⁻¹. This segmentation yielded eight high-curvature boundary segments for collision point location. Connectivity analysis was performed by constructing a topological graph model, with nodes representing the extracted bolt hole centers (e.g., coordinates [205, 78, 36]) and edges representing the connectivity matrix. The Tarjan strongly connected component algorithm was used to detect three groups of interfering components. The spatial location of the collision point uses precise calculations based on the SDF (Signed Distance Field) to determine the minimum distance within the voxel grid (resolution 1mm³) of the collision zone. When the SDF values ​​of both objects are simultaneously less than 0, the collision point is determined. In one calculation, four consecutive voxels were detected between the flange sealing surface and the pipe support in the Z-axis direction between 157mm and 159mm.

[0032] S4. Based on the spatial location data of the collision point, query the specific sub-models involved. By setting the information synchronization period between intelligent agents to exchange real-time data, determine the physical connection points between the associated sub-models of the collision point and obtain the details of the local conflict distribution.

[0033] The collision point data is used to query the sub-model set, obtain the associated spatial location information, and obtain a preliminary set of physical connection points. For this set of physical connection points, a synchronization cycle is used to configure data exchange between agents to obtain the point distribution after real-time data updates. Based on the real-time data, the matching relationship between the physical connection points and the sub-model set is analyzed to determine the preliminary scope of the local conflict. Based on the preliminary scope of the local conflict, the point analysis data is processed using the K-means clustering algorithm to obtain the distribution characteristics of the conflict-intensive area. Based on the distribution characteristics of the conflict-intensive area, the degree of fit between the spatial location and the real-time data is obtained to determine the adjusted local conflict boundary. Based on the adjusted local conflict boundary, the data consistency between agents is updated through exchange analysis to obtain the final distribution data. For the final distribution data, the continuity of the physical connection points is processed using interpolation to obtain the optimized local conflict distribution details.

[0034] Specifically, in the processing of the spatial position data of the collision point, the collision point is first quickly located through a spatial indexing algorithm (such as an R-tree or a quadtree). Assuming that the coordinates of the collision point are (x=15, y=3, z=7), the system will query the sub-model where the point is located.

[0035] For example, when using an R-tree index, the system will traverse the tree structure to find the minimum bounding rectangle (MBR) containing the point, thereby determining that the specific sub-model involved is "sub-model A". Next, the system sets the information synchronization cycle between agents to 100 milliseconds, and exchanges data in real time through message queues (such as Kafka). During the synchronization process, the agents will send their respective sub-model status information, such as the current displacement and rotation angle of "sub-model A". The system uses collision detection algorithms (such as the GJK algorithm) to determine the physical connection points between the associated sub-models of the collision point. Assuming that "sub-model A" collides with "sub-model B" at point (x=15, y=3, z=7), the system will calculate the normal vectors and contact forces of the two sub-models at this point, and generate local conflict distribution details.

[0036] For example, the system will output that the contact force between "sub-model A" and "sub-model B" at the collision point is 50N and the normal vector is (7, 3, 6), thereby providing data support for subsequent physical simulation or optimization.

[0037] S5. Based on the details of local conflict distribution, a reinforcement learning algorithm is used to introduce real-time change data to train the intelligent agent, update the monitoring priority weights and collaborative action sequences, and determine the adjusted intelligent agent behavior pattern.

[0038] The real-time change data acquisition module acquires dynamic input, allowing the trained agent to update its behavior pattern. Based on the behavioral characteristics output by the trained agent, the monitoring priority weights are adjusted to determine the collaborative action sequence. Interpolation is used to maintain continuity between actions within the spatial distribution of the generated collaborative action sequence, resulting in a smoothed behavioral pattern. If the monitoring priority weight exceeds the preset threshold, the data introduction module re-collects real-time change data and updates the trained agent. Based on the updated trained agent results, an adjusted action sequence is obtained to determine the changing trend of the behavioral pattern. Through trend analysis, a reinforcement learning algorithm is used to optimize weight updates and determine the final collaborative action sequence. For the final collaborative action sequence, detailed distribution of local conflicts is obtained to determine the adjusted agent behavior pattern.

[0039] Specifically, the Deep Q-Network (DQN) algorithm was used to train the agent using real-time changing data in the local conflict distribution. First, sensors collected real-time data from the conflict area, including temperature, humidity, and occupancy density. This data was normalized and then fed into the DQN network. The network structure consisted of three fully connected layers, with 128, 256, and 128 neurons, respectively. Reinforced Luminance (ReLU) was used as the activation function, and the output layer used a Softmax function to generate action probability distributions. During training, a learning rate of 0.01 and a discount factor of 9 were set. An experience replay mechanism was used to store 10,000 historical data points, with 32 randomly sampled for update during each training session. The agent selected an action based on its current state, such as "prioritize monitoring area A" or "collaboratively adjust resource allocation in area B." The effectiveness of the action was evaluated using a reward function designed as a weighted sum of the degree of conflict reduction, with weights of 6 and 4, respectively. After 1,000 iterations of training, the agent was able to dynamically adjust monitoring priority weights based on real-time data, for example, raising the priority of Area A from 7 to 9, while also generating new collaborative action sequences, such as "increasing the patrol frequency of Area B to 3 times per hour." Ultimately, the agent's behavioral patterns were optimized, enabling it to effectively handle local conflicts in complex environments and improving overall monitoring efficiency.

[0040] S6. According to the adjusted agent behavior pattern, edge geometry data is extracted from the sub-model set. The agents exchange data according to the updated collaborative action sequence to determine the conflict area with overlapping boundaries and obtain the updated collision probability distribution map.

[0041] Using the adjusted agent behavior pattern, edge geometry data is extracted from the sub-model set. The data exchange module is then used to process the collaborative action sequence, yielding a preliminary determination of boundary overlap. Based on this preliminary determination of boundary overlap, the spatial distribution characteristics of the conflict area are determined, and an initial distribution map of collision probability is determined. If the spatial distribution characteristics of the conflict area exceed the preset threshold, the agent re-executes data exchange, updates the action sequence, and obtains adjusted boundary overlap data. Interpolation is used to maintain the continuity of the conflict area based on the adjusted boundary overlap data, yielding a smoothed probability distribution map. Using the updated probability distribution map, the changing trend of the action sequence is determined, and a reinforcement learning algorithm is used to optimize the execution sequence of collaborative actions. Based on the optimized execution sequence of collaborative actions, the final distribution details of the conflict area are obtained, and an adjusted collision probability distribution map is determined. Updated characteristics of the agent behavior are extracted from the adjusted collision probability distribution map, yielding an optimized behavior pattern.

[0042] Specifically, after adjusting the agent's behavior, an improved α-shape algorithm (with parameter α = 5) is first used to extract edge geometry features based on the 3D point cloud data in the sub-model set (e.g., 10,000 sampling points with an accuracy of ±1 mm). This algorithm then generates a polygonal mesh containing curvature information. This updated sequence of collaborative actions is then executed across a distributed agent network (consisting of 50 computing nodes). Using a timestamp-based incremental data exchange protocol (with a synchronization period of 200 milliseconds), each agent transmits the boundary coordinates of its jurisdiction (e.g., the boundary point set {(x1, y1, z1)…(x100, y100, z100)} for region A). When it is detected that the boundary polygons transmitted by two intelligent agents have overlapping areas (for example, the boundaries of area A and area B overlap by 15%), the conflict detection module is triggered, and a Monte Carlo simulation (10,000 iterations) is used to calculate the collision probability, taking into account the geometric error (standard deviation σ = 0.5 mm) and motion trajectory uncertainty (Brownian motion parameter μ = 2). Finally, an updated collision probability heat map is generated (resolution 512×512 pixels, probability value range 0-0), in which high-conflict areas (probability > 8) are marked in red, and the probability distribution boundary is smoothed using a spatial interpolation algorithm (radial basis function kernel radius r = 5 mm).

[0043] S7. If there are still unresolved conflict points in the updated collision probability distribution map, the spatial position data of the unresolved conflict points are obtained from the distribution map, the involved sub-models are determined through the associated sub-model query, and the computing resources are allocated using the distributed computing framework to obtain the local collision detection results.

[0044] If there are unresolved conflict points in the collision probability distribution map, data extraction is used to obtain the spatial location data of the conflict points and determine the involved sub-models. Relevant geometric features are extracted from the sub-models through association queries, and computing resources are allocated using a distributed framework to obtain preliminary local collision detection results. Based on the local collision detection results, the boundary data of the conflict points in the distribution map is obtained. Interpolation is used to ensure the continuity of the boundary data to obtain a smoothed collision point distribution. Based on the smoothed collision point distribution, the overlapping areas between the sub-models are determined. A Monte Carlo algorithm is used to simulate the changing trend of collision probability and determine the adjusted distribution characteristics. If the number of conflict points in the adjusted distribution characteristics still exceeds the preset threshold, computing resources are reallocated through the distributed framework to obtain updated local collision data. Based on the updated local collision data, the correlation strength between the sub-models is determined, and the data exchange order is adjusted using a reinforcement learning algorithm to obtain an optimized collision point distribution. Based on the optimized collision point distribution, the final collision probability distribution map is determined, completing the collision point processing process.

[0045] Specifically, if unresolved conflict points are detected in the updated collision probability distribution map, the coordinates of these conflict points are first obtained through a spatial location data extraction algorithm. For example, in a three-dimensional coordinate system, the coordinates of conflict point A are (x=125, y=83, z=42). Next, the associated sub-model query algorithm is used to determine the involved sub-models based on the spatial location data of the conflict points.

[0046] For example, using spatial indexing technology, we can find that conflict point A is located at the intersection of sub-models M1 and M2. Then, using a distributed computing framework such as Apache Spark, we can distribute the computing tasks to multiple computing nodes, with each node responsible for processing collision detection in a local area.

[0047] For example, the collision detection tasks for sub-models M1 and M2 are assigned to nodes 1 and 2, respectively. Node 1 uses a collision detection algorithm based on AABB bounding boxes, while node 2 uses a collision detection algorithm based on the GJK algorithm. Through parallel computing, nodes 1 and 2 each obtain local collision detection results. For example, node 1 detects a collision with sub-model M1 at collision point A, and node 2 detects a collision with sub-model M2 at collision point A. Finally, the local collision detection results are summarized to generate a global collision detection report, providing data support for subsequent conflict resolution.

[0048] S8. For the local collision detection results, the intelligent agent exchanges data according to the set period to fuse the detection results and the collision probability distribution map, judges whether the detection accuracy reaches the preset detection accuracy threshold, and obtains the final local collision detection report.

[0049] The intelligent agent obtains the detection results of local collisions and periodically exchanges data to obtain a preliminary distribution of fused detections. Based on the preliminary distribution of fused detections, it determines whether the detection accuracy has reached the preset threshold and obtains accuracy assessment data. If the accuracy assessment data is lower than the preset threshold, the intelligent agent triggers a data update to obtain new local collision data. For the new local collision data, the distributed framework is used to fuse the detection results with the probability distribution to obtain adjusted distribution characteristics. Based on the adjusted distribution characteristics, the deviation between the detection accuracy and the preset threshold is determined to obtain deviation analysis results. Based on the deviation analysis results, a reinforcement learning algorithm is used to adjust the order of data exchange to obtain an optimized probability distribution. Based on the optimized probability distribution, the final state of detection accuracy is determined, completing the local collision processing process.

[0050] Specifically, after the local collision detection results are generated, the agent adopts a time window-based data synchronization mechanism to exchange detection data with a period of 200 milliseconds.

[0051] For example, node 1 detects a collision probability of 92 for submodel N3 at coordinates (x=215, y=147, z=68), while node 2 detects a collision probability of 88 for submodel N4 at the same coordinates. Using a Kalman filter-based data fusion algorithm, the detection results from the two nodes are weighted and fused, with a weight coefficient of 6 for node 1 and a weight coefficient of 4 for node 2. The resulting fused collision probability is 904. This value is compared with the preset detection accuracy threshold of 9 at the corresponding location in the collision probability distribution map. Since 904 is greater than 9, the collision detection result at that location is considered valid. For areas where the detection accuracy falls below the threshold, such as the fused collision probability of 87 at coordinates (x=198, y=165, z=72), which is lower than the threshold of 9, an iterative optimization algorithm based on gradient descent is triggered to adjust the collision detection parameters of submodels N5 and N6. After three iterations, the collision probability is improved to 91, meeting the accuracy requirement. Finally, all local detection results are summarized through a fast indexing mechanism based on a hash table to form a structured detection report containing fields such as coordinates, collision probability, and confidence.

[0052] S9. Based on the final local collision detection report, extract real-time change data from the construction environment data, and determine the block granularity range and intelligent agent monitoring responsibility allocation plan for the next detection by adaptively adjusting the sub-model division basis and task allocation rules.

[0053] By extracting real-time change data from the construction environment, an adaptive adjustment method is used to obtain a preliminary basis for sub-model division. Based on the preliminary basis for sub-model division, the adjustment content of the task allocation rules is obtained, and the distribution plan of monitoring responsibilities is determined. Based on the distribution plan of monitoring responsibilities, a preliminary range of block granularity is obtained through the allocation of intelligent agent responsibilities. From the preliminary range of block granularity, the random forest algorithm is used to determine the boundary characteristics of the detection range and obtain the adjusted detection range. If the adjusted detection range exceeds the preset threshold, the sub-model division basis is updated through adaptive adjustment to obtain a new block granularity range. Based on the new block granularity range, the optimized allocation plan of intelligent agent responsibilities is determined to obtain the responsibility distribution for the next inspection. The real-time change data of the construction environment is obtained through the responsibility distribution, and the matching status of the detection range and the allocation plan is determined.

[0054] Specifically, based on the confidence distribution in local collision detection reports, areas with a mean speed of 2 meters per second and a standard deviation of 35 degrees for dynamic obstacles were extracted from the construction environment point cloud data. A DBSCAN-based spatial clustering algorithm was used, with a neighborhood radius of ε = 8 meters and a minimum sample number minPts = 5, to divide the construction area into 12 dynamically changing subspaces. For subspaces S7 and S9, which exhibited densities exceeding 150 points per cubic meter, a weighted allocation model based on information entropy was used to prioritize the sub-models. Area S7, with an entropy value of 7 due to the intersection of three types of mobile device trajectories, was divided into five fine grids with a side length of 5 meters. The static material accumulation area S9, with an entropy value of 9, was merged into two coarse-grained grids with a side length of 2 meters. Agent task allocation uses a modified Hungarian algorithm. Taking into account each agent's historical detection accuracy (for example, Agent 3 achieved an 89% accuracy rate for high-speed moving target detection) and current load (CPU utilization below 65%), Agent 3 is assigned three highly dynamic grids in region S7. Region S9 is assigned to Agent 5, which is in standby mode. The gridding results are updated to the global spatial database via an R-tree index, with a synchronization period of 150 milliseconds to accommodate dynamic environmental changes.

[0055] The above is a schematic description of the present invention and its embodiments, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs a structure and embodiment similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A construction collision detection method based on BIM collaborative multi-agent reinforcement learning, characterized by: The following steps are involved: S1: Obtain geometric data and pipeline distribution information from the preset building information model, determine the granularity range of the initial model block by analyzing the pipeline density and regional complexity, and obtain the sub-model set after block division; S2: For the sub-model set, a multi-agent system is used to assign monitoring responsibilities according to the complexity of the sub-models. Edge geometry data and physical connection points are extracted from each sub-model. Information is exchanged through a defined data exchange format to determine potential conflict areas with overlapping boundaries and obtain a preliminary collision probability distribution map. S3: If there is a high-probability conflict area in the preliminary collision probability distribution map, obtain detailed geometric parameters and real-time construction change data from the high-probability conflict area, and determine the specific spatial location data of the collision point in the conflict area through boundary feature extraction and connection relationship analysis; S4: Based on the spatial location data of the collision point, the specific sub-models involved are queried. By setting an information synchronization period between agents to exchange real-time data, the physical connection points between the associated sub-models of the collision point are determined to obtain the details of the local conflict distribution. S5: Based on the details of local conflict distribution, a reinforcement learning algorithm is used to introduce real-time changing data to train the agent, update the monitoring priority weights and collaborative action sequences, and determine the adjusted agent behavior pattern; S6: Extract edge geometry data from the sub-model set based on the adjusted agent behavior pattern. The agents exchange data according to the updated collaborative action sequence to determine the conflict area with overlapping boundaries and obtain the updated collision probability distribution map. S7: If there are still unresolved conflict points in the updated collision probability distribution map, the spatial location data of the unresolved conflict points are obtained from the distribution map, the involved sub-models are determined through the associated sub-model query, and computing resources are allocated using the distributed computing framework to obtain the local collision detection results; S8: For the local collision detection results, the intelligent agent exchanges data according to a set period to fuse the detection results and the collision probability distribution map, and determines whether the detection accuracy reaches the preset detection accuracy threshold to obtain the final local collision detection report; S9: Based on the final local collision detection report, extract real-time change data from the construction environment data, and determine the block granularity range and intelligent agent monitoring responsibility allocation plan for the next detection by adaptively adjusting the sub-model division basis and task allocation rules.

2. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1 is characterized by: The specific steps of S1 are: obtaining geometric data and pipeline distribution information through a preset building information model to obtain a complete data set of the initial model; using the extracted geometric data and pipeline distribution information, analyzing pipeline density to obtain a density distribution map; determining the range of block granularity based on the density distribution map and regional complexity, and obtaining a granularity threshold set; The initial model is divided using a set of granularity thresholds to obtain a preliminary set of sub-models. For this preliminary set of sub-models, the matching degree between pipeline density and regional complexity is analyzed to determine the need for sub-model boundary adjustment, thereby obtaining an optimized set of sub-models. After obtaining the optimized set of sub-models, the consistency of geometric data and pipeline distribution is verified through data extraction to obtain the final set of sub-models. The machine learning algorithm K-means is used to cluster the final sub-model set, determine the spatial correlation between the sub-models, and obtain a classification set of block results.

3. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1 is characterized by: The specific steps of S2 are: dividing the sub-model set through a multi-agent system, determining the monitoring responsibility scope of each agent according to the complexity distribution result, and obtaining a responsibility allocation table; extracting the edge geometry information of each sub-model from the responsibility allocation table, and using a preset algorithm to calculate the shape data and physical connection points to obtain a connection point set; for the connection point set, processing the relationship between the edge geometry and the physical connection through the data exchange format, judging the area of ​​overlapping boundaries, and obtaining an overlapping area set; based on the overlapping area set, analyzing the spatial distribution characteristics of the conflict area, determining the density of potential conflicts, and obtaining a conflict density map; if there are high-density areas in the conflict density map, clustering the conflict areas through the K-means algorithm to obtain a clustered conflict classification set; using the conflict classification set, adjusting the boundary overlapping point extraction strategy, determining the optimized physical connection points, and obtaining an adjusted point distribution; verifying the spatial consistency of the sub-model set through the adjusted point distribution, and obtaining a final collision probability distribution map.

4. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1 is characterized by: The specific steps of S3 are as follows: obtaining geometric parameters and real-time data from the high-probability conflict area, processing boundary features through a preset algorithm, determining the spatial distribution characteristics of the conflict area, and obtaining a boundary feature set; extracting connection relationship data through the boundary feature set, analyzing the density of the conflict area using a clustering algorithm, judging the distribution characteristics of potential collision points, and determining a candidate collision point set; for the candidate collision point set, verifying the impact of construction changes on the spatial position through real-time data, and obtaining an adjusted collision point distribution; based on the adjusted collision point distribution, analyzing the matching degree between boundary features and connection relationships, judging the dynamic change trend of the conflict area, and obtaining change trend data; obtaining updated values ​​of geometric parameters through change trend data, processing the continuity of spatial position using interpolation method, and determining the optimized collision point coordinates; for the optimized collision point coordinates, analyzing the boundary overlap of the conflict area through connection relationships, and obtaining the overlapping area distribution; based on the overlapping area distribution, judging the degree of fit between the spatial position and the real-time data, and obtaining the final collision point spatial position data.

5. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1 is characterized by: The specific steps of S4 are: querying the sub-model set through the collision point data, obtaining the associated spatial position information, and obtaining a preliminary physical connection point set; for the physical connection point set, using a synchronous cycle to configure data exchange between intelligent agents, and obtaining the point distribution after the real-time data is updated; based on the real-time data, analyzing the matching relationship between the physical connection points and the sub-model set, and determining the preliminary scope of the local conflict; based on the preliminary scope of the local conflict, using the K-means clustering algorithm to process the point analysis data, and obtain the distribution characteristics of the conflict-intensive area; based on the distribution characteristics of the conflict-intensive area, obtaining the degree of fit between the spatial position and the real-time data, and determining the adjusted local conflict boundary; based on the adjusted local conflict boundary, updating the data consistency between the intelligent agents through exchange analysis, and obtaining the final distribution data; for the final distribution data, using the interpolation method to process the continuity of the physical connection points, and obtaining the optimized local conflict distribution details.

6. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1 is characterized by: The specific steps of S5 are as follows: obtaining dynamic input through the real-time change data acquisition module to train the intelligent agent to update the behavior pattern; adjusting the monitoring priority weight based on the behavioral characteristics output by the trained intelligent agent to determine the collaborative action sequence; using the interpolation method to process the continuity between actions based on the spatial distribution generated by the collaborative action sequence to obtain a smooth behavior pattern; if the monitoring priority weight exceeds the preset threshold, re-collecting the real-time change data through the data introduction module to update the trained intelligent agent; Based on the updated training results of the intelligent agent, the adjusted action sequence is obtained to determine the changing trend of the behavior pattern; through trend analysis, the reinforcement learning algorithm is used to optimize the weight update and determine the final collaborative action sequence; for the final collaborative action sequence, the distribution details of local conflicts are obtained to obtain the adjusted intelligent agent behavior pattern.

7. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1 is characterized by: The specific steps of S6 are: extracting edge geometry data from the sub-model set through the adjusted intelligent agent behavior pattern, using the data exchange module to process the collaborative action sequence, and obtaining a preliminary judgment result of boundary overlap; based on the preliminary judgment result of boundary overlap, obtaining the spatial distribution characteristics of the conflict area, and determining the initial distribution map of the collision probability; if the spatial distribution characteristics of the conflict area exceed the preset threshold, re-executing the data exchange through the intelligent agent, updating the action sequence, and obtaining adjusted boundary overlap data; for the adjusted boundary overlap data, using the interpolation method to process the continuity of the conflict area, and obtaining a smooth probability distribution map; judging the change trend of the action sequence through the updated probability distribution map, and optimizing the execution order of the collaborative action using the reinforcement learning algorithm; obtaining the final distribution details of the conflict area based on the optimized execution order of the collaborative action, and determining the adjusted collision probability distribution map; The updated features of the agent's behavior are extracted from the adjusted collision probability distribution graph to obtain the optimized behavior pattern.

8. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1 is characterized by: The specific steps of S7 are as follows: if there are unresolved conflict points in the collision probability distribution map, spatial location data of the conflict points are obtained through data extraction to determine the sub-models involved; relevant geometric features are extracted from the sub-models through association query, computing resources are allocated using a distributed framework, and preliminary local collision detection results are obtained; based on the local collision detection results, boundary data of the conflict points in the distribution map are obtained, and the continuity of the boundary data is processed by interpolation to obtain a smooth distribution of conflict points; based on the smooth distribution of conflict points, the overlapping areas between the sub-models are determined, and the Monte Carlo algorithm is used to simulate the changing trend of the collision probability to determine the adjusted distribution characteristics; If the number of conflict points in the adjusted distribution characteristics still exceeds the preset threshold, the computing resources are reallocated through the distributed framework to obtain updated local collision data. Based on the updated local collision data, the correlation strength between the sub-models is obtained, and the data exchange order is adjusted through the reinforcement learning algorithm to obtain an optimized conflict point distribution. Based on the optimized conflict point distribution, the final collision probability distribution map is determined to complete the conflict point processing process.

9. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1, characterized in that: The specific steps of S8 are: obtaining the detection results of local collisions through the intelligent agent, and exchanging data according to a periodic setting to obtain the preliminary distribution of fused detection; judging whether the detection accuracy reaches a preset threshold based on the preliminary distribution of fused detection, and obtaining accuracy evaluation data; if the accuracy evaluation data is lower than the preset threshold, triggering data update through the intelligent agent to obtain new data of local collisions; for the new data of local collisions, using a distributed framework to fuse the detection results and the probability distribution to obtain adjusted distribution characteristics; judging the deviation of the detection accuracy from the preset threshold based on the adjusted distribution characteristics, and obtaining the deviation analysis results; based on the deviation analysis results, using the reinforcement learning algorithm to adjust the order of data exchange to obtain an optimized probability distribution; judging the final state of the detection accuracy based on the optimized probability distribution, and completing the processing flow of local collisions.

10. The construction collision detection method based on BIM collaborative multi-agent reinforcement learning according to claim 1, characterized in that: The specific steps of S9 are: extracting real-time change data from the construction environment, and using an adaptive adjustment method to obtain a preliminary basis for sub-model division; based on the preliminary basis for sub-model division, obtaining the rule adjustment content of task allocation, and determining the distribution scheme of monitoring responsibilities; for the distribution scheme of monitoring responsibilities, obtaining a preliminary range of block granularity through the distribution of intelligent agent responsibilities; from the preliminary range of block granularity, using a random forest algorithm to determine the boundary characteristics of the detection range, and obtaining an adjusted detection range; if the adjusted detection range exceeds the preset threshold, updating the sub-model division basis through adaptive adjustment to obtain a new block granularity range; based on the new block granularity range, determining the optimal distribution scheme of intelligent agent responsibilities, and obtaining the responsibility distribution for the next detection; obtaining real-time change data of the construction environment through the responsibility distribution, and judging the matching status of the detection range and the distribution scheme.

Images