Automatic inspection path planning method based on BIM and unmanned aerial vehicle construction

Abstract

The invention belongs to the field of physical distance measurement, and particularly relates to an automatic inspection path planning method based on BIM and unmanned aerial vehicle construction. In order to solve the problems of large calculation amount, difficulty in obstacle avoidance and low coverage efficiency of routing inspection path planning in a complex and limited construction environment, the method comprises the following steps: firstly, analyzing building information model data and constructing a manifold space mapping relation; then key inspection targets and obstacles are identified to construct a topological mapping model, and a global topological mapping graph is generated; calculating a node priority by using a graph neural network, and generating a continuous flow pattern curve path meeting an optimal coverage requirement based on a functional extremum principle; and performing end-side real-time re-planning and obstacle avoidance in combination with environmental point cloud data in the flight process. According to the method, the non-Euclidean geometric measurement and topological mapping mechanism is introduced, so that the calculation complexity is reduced, and the inspection efficiency and the digital level of a construction site are improved while the flight smoothness and the operation safety of the unmanned aerial vehicle are ensured.

Description

Technical Field

[0001] This invention belongs to the field of physical distance measurement, specifically relating to an automatic inspection path planning method based on BIM and UAV construction. Background Technology

[0002] With the deep integration of Building Information Modeling (BIM) technology and drone automation technology, drone-based automated inspection of construction sites has become a key means to improve the efficiency of engineering quality monitoring and construction safety management. In the construction engineering field, integrating high-precision BIM data models provides fundamental support for digital supervision throughout the entire construction lifecycle, enabling managers to rehearse construction processes in a virtual environment. In large and complex construction projects, using drones equipped with various sensors for periodic inspections has significant practical value for real-time monitoring of project progress, identification of structural defects, and ensuring construction compliance.

[0003] Automated inspection path planning, a core component of intelligent UAV operations, aims to generate safe and comprehensive flight trajectories based on the geometric constraints provided by the BIM model. This technology requires the system to accurately analyze dynamic obstacle information at the construction site and, in conjunction with inspection task requirements, plan a flight scheme in real time that incorporates obstacle avoidance capabilities and path optimization. By comparing the construction blueprint with real-time collected sensor data, the path planning scheme directly determines the completeness of the inspection data and the UAV's survivability in complex and confined spaces.

[0004] Existing automated inspection solutions typically rely on Euclidean space coordinate navigation. However, in the narrow and complex internal spaces of super high-rise core structures, traditional GPS or visual obstacle avoidance algorithms are highly susceptible to signal multipath effects and electromagnetic interference, leading to positioning failures. Furthermore, existing 3D path planning methods often depend on complex geometric obstacle avoidance logic, making it difficult for UAVs to support high-intensity real-time replanning tasks. This results in excessive reliance on remote processing from ground stations, causing significant communication delays and command deviations. Conventional path generation often employs discrete straight-line splicing, which is ill-suited for complex construction topologies with manifold characteristics. It fails to achieve optimal, comprehensive coverage of the construction site in nonlinear space, resulting in low inspection efficiency and potential safety hazards. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic inspection path planning method based on BIM and UAV construction, which can solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an automatic inspection path planning method based on BIM and UAV construction, comprising the following specific steps: Step 1: Parse Building Information Model Data and Construct Manifold Spatial Mapping Relationship: Obtain the original building information model of the construction site, extract the geometric component attributes, spatial topological constraints and construction progress information, establish a three-dimensional spatial coordinate system, and transform the Euclidean spatial coordinates in the building information model into a manifold spatial representation with curvature characteristics through a non-Euclidean geometric transformation program. Step 2: Construct a topology mapping model based on graph neural networks: Identify key inspection targets and physical obstacles in the construction environment in the building information model, abstract the key inspection targets, construction scaffolding, multi-level intersections and building components as topology nodes, and establish connection edges between nodes based on the geodesic distance in the manifold space to generate a global topology mapping graph; Step 3: Generate the optimal inspection path based on manifold curves: Input the global topology map into the pre-trained graph neural network, use graph convolution operators and attention mechanisms to calculate the inspection priority and coverage weight of each node, and generate an inspection path that meets the optimal coverage requirements based on the principle of functional extrema in manifold space. The inspection path consists of a series of continuous manifold curves. Step 4: Perform real-time replanning and obstacle avoidance on the UAV side: During the flight of the UAV, the environmental point cloud data collected by the airborne sensors is acquired in real time. The environmental point cloud data is dynamically compared with the manifold space. If dynamic obstacles or signal interference areas are detected within a preset range, the manifold curve is replanned in real time with the support of computing power on the UAV side by updating the edge weights of the global topology mapping map. Step 5: Output inspection instructions and execute automated path guidance: Transform the planned manifold curve into control instructions for the UAV flight control system. By controlling the attitude and heading of the UAV, the UAV completes the inspection task along the manifold curve in a confined space, and the inspection data is associated and stored with the building information model.

[0007] Preferably, step 1, parsing the building information model data, includes: reading the building information model file through a standard data interface, extracting geometric entity information including beams, columns, slabs, walls, and core tubes; identifying the three-dimensional boundary coordinates of each type of geometric entity and normalizing them to a preset global coordinate system; extracting the construction site status at the current time node according to the construction schedule, removing data of components that have not yet been constructed, and retaining the structural information of completed or under construction as the geometric constraint basis for path planning.

[0008] Preferably, the non-Euclidean geometric transformation procedure includes: defining a local coordinate tangent space based on a three-dimensional coordinate system in Euclidean space; mapping discrete points in the three-dimensional space to the manifold surface using an exponential mapping operator; calculating the metric tensor of each sampling point on the manifold, wherein the metric tensor describes the length scaling ratio of the space in different directions; and using the metric tensor to characterize the narrow passages and complex corners of the construction site, so that the measurement of spatial distance conforms to the actual flight constraints of the UAV in complex and confined spaces.

[0009] Preferably, the curvature feature calculation in the manifold space includes: performing triangular meshing on the surface of the building information model, calculating the average curvature and Gaussian curvature of each vertex; identifying high curvature regions in the construction environment based on the curvature distribution, wherein the high curvature regions correspond to sharp corners or slender components at the construction site; and refining the mesh in the high curvature regions when generating the manifold representation to improve the geometric accuracy of path planning at complex structures.

[0010] Preferably, the identification of key inspection targets in step 2 includes: based on the attribute tags of the building information model, filtering out specific components that need to be inspected or monitored for quality, the specific components including embedded parts, rebar lap joints, concrete pouring surfaces and high-altitude scaffolding; assigning a unique topology identifier to each key inspection target, and setting its initial weight value in the topology mapping model according to the inspection frequency requirements.

[0011] Preferably, the generation of the global topology map includes: performing space-filling curve sampling within the manifold space to generate a candidate node set covering the entire construction area; applying a visibility algorithm to determine the visibility between nodes, and establishing a connection edge if there is no geometric occlusion between two nodes on the manifold surface; calculating the length of the connection edge, where the length is the shortest geodesic distance connecting two points on the manifold surface; and using the geodesic distance as the initial weight of the connection edge to construct a graph structure containing a set of nodes and a set of edges.

[0012] Preferably, the training process of the graph neural network in step 3 includes: constructing a deep network structure composed of multiple graph convolutional layers; defining an input feature vector, which includes the spatial location information of nodes, component type information, and associated inspection task level; in the hidden layer, performing node feature aggregation and updating, and extracting global topological features by learning the structural information of neighboring nodes; using a loss function to minimize the total path length and the area of ​​the uncovered region, and optimizing the network parameters through the backpropagation algorithm.

[0013] Preferably, the method of calculating the inspection priority using the attention mechanism includes: introducing a self-attention coefficient during the aggregation process of the graph neural network; the self-attention coefficient is calculated based on the similarity between the features of the current node and the features of the neighboring nodes; for nodes with high construction risk or delayed progress, their weight coefficient is increased through human intervention or automatic logic, so that the attention mechanism can guide the drone to pass through the node first during path planning, thereby achieving intensive inspection of key areas.

[0014] Preferably, the generation of inspection paths based on the functional extremum principle in the manifold space includes: defining an inspection energy functional, which is a weighted combination of a path length term, a coverage completeness term, and a smoothness term; solving for the minimum value of the inspection energy functional using the variational method to derive the Euler equation on the manifold; and solving the Euler equation numerically to obtain a continuous curve connecting each inspection node, wherein the continuous curve has second-order continuity on the manifold surface to ensure the dynamic smoothness of the UAV flight.

[0015] Preferably, the generated manifold curve has the following characteristics: the tangent direction of the curve at any point is consistent with the preset heading of the UAV; the radius of curvature of the curve is always greater than the preset threshold of the minimum turning radius of the UAV; when passing through narrow spaces such as inside the core tube, the curve automatically contracts towards the manifold centerline to increase the safe distance between the UAV and the wall.

[0016] Preferably, the environmental point cloud data processing in step 4, which is acquired by the airborne sensor, includes: acquiring real-time 3D point clouds using a lidar or depth camera mounted on the UAV; removing noise interference points from the point cloud using a statistical filtering algorithm; reducing the size of the point cloud using voxel mesh downsampling technology to meet the computing performance limitations of the UAV's edge processor; and quickly registering the reduced real-time point cloud with the building information model to identify dynamic objects not included in the model, such as mobile cranes, construction workers, or other temporary facilities.

[0017] Preferably, the real-time replanning manifold curve includes: when a dynamic obstacle intrudes into the predetermined safety radius of the current inspection path, immediately marking the area where the obstacle is located as an unreachable area in the global topology mapping; triggering a local path search procedure to find a sequence of alternative nodes in the manifold space to bypass the unreachable area; and using a manifold curve fitting algorithm to regenerate a local obstacle avoidance curve while keeping the connection direction of the start and end points unchanged, and seamlessly splicing it with the original inspection path.

[0018] Preferably, the computation process supported by the UAV edge computing power includes: quantizing and compressing the inference process of the graph neural network so that it runs in the UAV's onboard embedded computing unit; using a parallel computing architecture to accelerate the feature update process of the topology nodes; and narrowing the solution space range during replanning through a preset heuristic search strategy to ensure that the response time of the path update is less than a preset time threshold, thereby avoiding collisions during high-speed flight of the UAV.

[0019] Preferably, the conversion process of control commands in step 5 includes: sampling the planned manifold curve at equal intervals to obtain a series of waypoint coordinates and their corresponding desired velocity vectors and desired acceleration vectors; inputting the coordinates and vectors into the UAV's flight control operator, which calculates the motor speed command based on proportional-integral-derivative control logic or model predictive control logic; during the command output process, monitoring the UAV's battery level and communication signal strength in real time, and executing an autonomous return-to-home procedure if the battery level is lower than a predetermined percentage or the signal attenuation exceeds a predetermined range.

[0020] Preferably, the association storage of the inspection data with the building information model includes: recording the manifold spatial coordinates and camera pose information at the current moment while the UAV collects images, videos or point cloud data; using timestamp alignment technology to map the sensor data back to the corresponding components of the building information model; establishing a multi-dimensional database to store the inspection results, supporting the retrieval of corresponding historical inspection records by clicking on model components, and realizing the visual traceability of construction quality.

[0021] Preferably, the method further includes a step of evaluating the effectiveness of the inspection path: after the task is completed, the average deviation between the actual flight trajectory and the planned flow curve is calculated; the coverage rate of the inspection target is calculated, that is, the ratio of the number of observed nodes to the total number of nodes; the average computation time of path planning and replanning is recorded; based on the evaluation results, the weight parameters in the graph neural network are automatically adjusted to achieve self-evolution and optimization of subsequent inspection tasks.

[0022] Preferably, for enclosed spaces such as the core tube of ultra-high-rise buildings, the method adopts a positioning compensation logic based on the fusion of ultrasonic and lidar: an additional signal shielding field is defined in the manifold space, and when the UAV enters the area where the global positioning system fails, it automatically switches to a combined positioning mode based on visual inertial odometry and laser matching; by pre-setting high-precision feature nodes in the global topology map, the UAV is assisted in maintaining centimeter-level positioning accuracy in complex internal spaces.

[0023] Preferably, the dynamic update of the construction progress information includes: periodically comparing the real-time point cloud data obtained from the inspection with the standard geometric model in the building information model; using a semantic segmentation network to identify the types of components that have been constructed and their spatial locations; and when a deviation between the actual construction progress and the planned progress is detected, automatically triggering the update process of the path planning program to regenerate an inspection topology map adapted to the current site environment.

[0024] Preferably, the generation logic of the manifold curve in the intersection area is as follows: identify the intersection center of multiple construction channels and set it as the topological hub in the manifold space; when generating the curve, a smooth multi-branch connection strategy is adopted to ensure that the change in angular velocity of the UAV is less than the preset physical limit threshold when switching between different inspection branches, thereby improving flight stability.

[0025] Preferably, the method further includes establishing a cloud-based construction monitoring collaboration mechanism: when the drone is performing an inspection task, it transmits the compressed low-resolution video stream back to the ground station in real time via a wireless communication link; after receiving the video stream, the ground station projects it onto the 3D building information model interface for managers to perform real-time remote monitoring; if the managers discover an emergency safety hazard, they can issue a forced interruption command through the ground station to take over control of the drone or make it hover immediately.

[0026] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention breaks through the limitations of traditional Euclidean coordinate navigation by transforming building information models into manifold spatial representations. By introducing non-Euclidean geometric metrics, it can accurately describe the geometric constraints in complex and confined spaces such as the core tube of super high-rise buildings, solve the positioning failure problem caused by signal multipath effects and electromagnetic interference, and improve the survivability of UAVs in complex environments inside buildings.

[0027] 2. This invention simplifies the 3D geometric obstacle avoidance problem into graph theory computation in topological space using graph neural networks. Intelligent weight allocation for inspection targets is achieved through graph convolution operators and attention mechanisms, reducing the computational complexity of path planning. This topological mapping mechanism enables UAVs to complete dozens of levels of real-time obstacle avoidance replanning based on edge computing power, alleviating the system's dependence on high-latency communication with ground stations and ensuring the real-time performance and safety of inspection operations.

[0028] 3. The inspection path generated by this invention adopts a manifold curve based on manifold space optimization coverage. Compared with the traditional discrete straight line splicing path, the manifold curve has better dynamic smoothness and coverage efficiency. The path obtained by functional extremum principle can balance flight distance and energy consumption while meeting the requirement of all-round coverage of the construction site, thereby improving the inspection efficiency and operation continuity of UAVs at the construction site.

[0029] 4. This invention achieves deep integration of Building Information Modeling (BIM) data, UAV perception data, and inspection task logic. By dynamically updating the topology mapping model, it can adapt to the ever-changing environment of the construction site. The associated storage mechanism provides a precise digital foundation for construction quality traceability, promoting the transformation of building construction inspection from manual to automated, intelligent, and digital. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram of the core principle framework constructed based on building information model analysis, manifold space mapping and global topology mapping according to an embodiment of the present invention; Figure 3 A logical flowchart is provided for calculating inspection weights and generating the optimal manifold curve path using the attention mechanism based on a graph neural network according to an embodiment of the present invention. Figure 4 A flowchart illustrating real-time obstacle avoidance and environmental point cloud perception and manifold curve replanning performed on the UAV side according to an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the multi-level interaction relationship and data flow between the inspection command conversion and execution and the associated storage of inspection results with the building information model, according to an embodiment of the present invention. Detailed Implementation

[0031] Example 1: Please refer to Figures 1 to 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0032] When implementing an automated inspection path planning method based on Building Information Modeling (BIM) and UAV construction, this invention first performs step 1, which involves parsing the BIM data and constructing a manifold spatial mapping relationship. In this step, the system reads the BIM file through a standard data interface, such as an industrial basic class file interface. This parsing process involves extracting the attributes of all geometric entities in the model, including but not limited to beams, columns, floor slabs, walls, and core tubes in high-rise buildings. The system identifies the three-dimensional boundary coordinates of each type of geometric entity and extracts its vertex set and facet connection relationships in the local coordinate system. To achieve uniformity in subsequent path planning, the system normalizes all extracted coordinate data and transforms it into a preset global coordinate system. Through linear transformations of translation vectors and rotation matrices, it ensures that every geometric detail in the BIM model can be accurately located in the global space.

[0033] While acquiring geometric entities, this invention further integrates with the construction schedule to extract the current construction site status at any given time. The system identifies completed components, components under construction, and components yet to be constructed by comparing timestamp information in the construction management database. When constructing the geometric constraint basis for path planning, the system automatically removes data from components yet to be constructed, retaining only the structural information currently physically present on site. This dynamic filtering mechanism ensures the effectiveness of path planning and prevents drones from performing ineffective obstacle avoidance maneuvers in open areas.

[0034] To address the complex geometric topology of the construction site, step 1 also includes the execution of a non-Euclidean geometric transformation procedure. Based on a three-dimensional coordinate system in Euclidean space, the system defines a local coordinate tangent space for each sampling point. Using an exponential mapping operator, discrete geometric points in the three-dimensional space are mapped to the manifold surface. To quantify the geometric properties of the manifold space, the system calculates the metric tensor for each sampling point on the manifold. This metric tensor characterizes the degree of physical constraints at the construction site by describing the length scaling ratio of space in different directions. For example, in narrow construction passages or complex corner areas, the component values ​​of the metric tensor will change accordingly, allowing the measurement of spatial distance under the manifold representation to accurately reflect the difficulty of UAV flight within confined spaces.

[0035] The calculation of curvature features in the manifold space is also a core component of step 1. The system triangulates the surface of the Building Information Model (BIM) into a mesh structure composed of numerous vertices and edges, discretizing the complex surface. The average curvature and Gaussian curvature of each mesh vertex are calculated, and the curvature distribution is used to identify high-curvature regions in the construction environment, such as sharp building corners or slender scaffolding members. When generating the manifold representation, the system automatically refines the mesh for these high-curvature regions. The refined mesh provides a higher density of geometric sampling points, ensuring that the inspection path maintains extremely high geometric accuracy in complex structures.

[0036] After constructing the manifold space, this invention proceeds to step 2, building a topology mapping model based on a graph neural network. The system identifies key inspection targets in the building information model. Based on the model's attribute labels, the system uses a semantic retrieval algorithm to filter out specific components such as embedded parts locations, rebar lap joints, concrete pouring surfaces, and high-altitude scaffolding. A unique topology identifier is assigned to each key inspection target, consisting of a component type code, floor code, and sequence number. According to the inspection frequency required by construction specifications, the system sets an initial weight value for each target, which determines the importance of the node in the global topology graph.

[0037] The system generates a global topology map. Within the manifold space, the system performs space-filling curve sampling, such as Hilbert curve sampling, to generate a candidate node set covering the entire construction area. To ensure node validity, a visibility algorithm is applied to determine the visibility between nodes. The system emits virtual rays from one candidate node to another. If the ray is not obstructed by any building geometry on the manifold surface, visibility between the two points is determined, and a connecting edge is established in the graph. The length of the connecting edge is no longer based on the traditional Euclidean straight-line distance, but rather on the shortest geodesic distance connecting the two points on the manifold surface. The geodesic distance is calculated by solving the path integral on the manifold, and it serves as the initial weight of the connecting edge, reflecting the actual cost of the UAV moving in the curved space. Thus, the system constructs a global topology map containing both the node set and the edge set, transforming the complex 3D obstacle avoidance problem into a pathfinding problem on the topology map.

[0038] Next, the invention executes step 3 to generate an optimized inspection path based on the manifold curve. The system inputs the global topology map into a pre-trained graph neural network. This graph neural network contains multiple graph convolutional layers, each layer performing node feature aggregation and updating. The input feature vector includes the three-dimensional spatial location components of the nodes, component type encoding, and the associated inspection task urgency level. In the hidden layers, nodes extract global topology features by learning the structural information of their neighboring nodes. The invention utilizes an attention mechanism to calculate inspection priority, introducing a self-attention coefficient during the graph convolution aggregation process. This coefficient is calculated based on the cosine similarity between the current node's features and the features of its neighboring nodes. For key nodes with high construction risks or delayed progress, the system increases their weight coefficients through manually set bias terms, enabling the attention mechanism to prioritize guiding the drone through these areas during path planning.

[0039] Path generation is based on the functional extremum principle within the manifold space. The system defines an inspection energy functional, which is a weighted combination of path length, coverage integrity, and dynamic smoothness terms. The path length term represents the cost of the total inspection journey, the coverage integrity term represents the proportion of key inspection targets detected by the sensor's field of view, and the smoothness term constrains the rate of change of the path's curvature. The system uses a variational method to solve for the minimum value of this energy functional, deriving the Euler equation on the manifold. By numerically iterating through this equation, a continuous manifold curve connecting each inspection node is obtained. The generated manifold curve has second-order continuity, ensuring smooth acceleration changes during UAV flight. Characteristically, the tangent direction of the curve at any point is consistent with the UAV's preset heading, and the radius of curvature of the curve is always greater than the preset threshold of the UAV's physical minimum turning radius. When the path passes through narrow spaces such as inside the core tube, the manifold curve automatically contracts towards the centerline of the manifold, increasing the safe physical distance between the UAV and the side walls.

[0040] During the drone's mission, this invention executes step 4, namely, real-time replanning and obstacle avoidance at the edge. The lidar and depth camera onboard the drone acquire 3D point cloud data of the construction environment in real time. To meet the real-time requirements of the edge computing unit, the system first uses a statistical filtering algorithm to remove outlier noise points from the point cloud, based on whether the average distance between a sampled point and its neighboring points exceeds a preset standard deviation multiple. Subsequently, voxel mesh downsampling technology is applied to discretize the point cloud into fixed-size voxel cubes, significantly reducing the amount of data to be processed.

[0041] The processed real-time point cloud is rapidly registered with the original building information model. The system identifies dynamic objects not included in the model, such as moving crane booms, construction site personnel, or temporarily piled building materials. When these dynamic obstacles are detected encroaching on the predetermined safety radius of the current inspection path, the system immediately marks the area occupied by the obstacle as an unreachable region in the global topology map. At this point, a local path search procedure is triggered to find a new node sequence that bypasses the unreachable region within the manifold space. Using a manifold curve fitting algorithm, while keeping the connection direction between the start and end points of the current path unchanged, a new local obstacle avoidance curve is regenerated and seamlessly stitched with the original inspection path. This calculation process is accelerated by the UAV's onboard parallel computing architecture. By quantizing and compressing the graph neural network inference process, its response time in the embedded computing unit is less than a preset time threshold, ensuring that the UAV can safely avoid obstacles during high-speed inspections.

[0042] Finally, the invention executes step 5, outputting inspection commands and performing automated path guidance. The system performs equidistant sampling on the planned manifold curve, extracting a series of high-frequency waypoint coordinates. Each waypoint contains not only three-dimensional position information but also the desired velocity vector and desired acceleration vector. This data is input into the UAV's flight control operator, which, based on model predictive control logic, calculates the rotational speed commands for each motor in real time. During command output, the system closely monitors the UAV's operating status, including battery voltage percentage, propeller speed fluctuations, and communication signal strength. If the battery level is detected to be below a safe threshold or the wireless signal attenuation exceeds a predetermined range, the system will immediately interrupt the current task and execute an autonomous return-to-home procedure.

[0043] During inspections, the system associates and stores the collected high-definition images, infrared videos, or laser point cloud data with the Building Information Model (BIM). Utilizing high-precision timestamp alignment technology, the system binds each frame of sensor data with its current manifold spatial coordinates and camera pose information, mapping the data back to the corresponding components in the BIM through a spatial projection algorithm. The resulting multidimensional database allows users to directly retrieve historical inspection records by clicking on virtual components in the model, enabling full-process visual traceability of construction quality.

[0044] After the mission is completed, this invention also includes an effectiveness evaluation of the inspection path. The system calculates the average spatial deviation between the actual flight trajectory and the planned manifold curve, and counts the proportion of observed nodes to the total number of nodes. Based on these evaluation results, the system automatically adjusts the weight parameters in the graph neural network using a backpropagation algorithm. This self-evolutionary mechanism enables the system to continuously optimize subsequent inspection path planning strategies according to the characteristics of different construction sites, thereby improving operational efficiency.

[0045] For areas where GPS signals are lost, such as the core tube of ultra-high-rise buildings, this invention employs a positioning compensation logic based on the fusion of ultrasonic and lidar. The system defines a signal masking field within the manifold space. When a UAV enters this field, it automatically switches to a combined positioning mode based on visual inertial odometry and laser matching. By using pre-defined high-precision feature nodes in the global topology map as spatial anchor points, the system can maintain centimeter-level positioning accuracy within enclosed spaces, ensuring that inspection tasks are not affected by the lack of external positioning sources.

[0046] For complex intersection areas, the manifold curve generation logic involves identifying the intersection center of multiple construction channels and setting it as the topological hub in the manifold space. A smooth, multi-branch connection strategy is employed when generating the curve. By calculating the change in angular velocity when entering and leaving the intersection, it is ensured that this change is always less than a preset physical limit threshold for the UAV, thus avoiding the risk of the UAV rolling over during sharp turns and improving flight stability.

[0047] This invention also establishes a cloud-based construction monitoring and collaboration mechanism. The drone transmits compressed, low-resolution video streams in real-time to the ground monitoring station via a wireless communication link. Upon receiving the video stream, the ground station projects it onto a 3D building information model interface using virtual reality or augmented reality technology. Management personnel can conduct remote inspections based on the real-time transmitted images. If an emergency safety hazard is detected, management personnel can issue a forced shutdown command through the ground station, directly taking control of the drone or putting it into emergency hover mode, thereby improving the level of safety management during on-site construction.

[0048] Example 2 demonstrates the applicability of this invention in ultra-large span spaces during the construction inspection of a large steel structure stadium. In this example, the geometric entities extracted in step 1 focus on complex steel structure nodes and space frame members. Due to the extremely high precision requirements of steel structure construction, the system increases the sampling frequency of the metric tensor to twice that of conventional buildings when constructing the manifold space to capture minute deflection deformation trends of the members.

[0049] Following the logic of step 2, the system abstracts each bolt connection node as a topological node. In large-scale grid structures, the number of nodes can reach tens of thousands. To improve computational efficiency, the system adopts a hierarchical graph neural network architecture, first performing topological mapping of the various zones of the stadium at a macroscopic level, and then refining the specific nodes of each zone at a microscopic level. When calculating the connection edges between nodes, the system specifically introduces the electromagnetic interference field distribution as a weighting correction factor. Because the steel structure has a strong reflection and shielding effect on electromagnetic waves, the geodesic distance will be weighted in conjunction with the signal field strength distribution to guide the drone to avoid signal blind spots or areas with severe multipath effects.

[0050] In step 3, when generating the inspection path, considering the tall and open interior of the stadium, the weight of the smoothness term in the inspection energy functional is appropriately increased. This results in a manifold curve with longer straight lines and gentler turns, which helps the UAV maintain a constant inspection speed and improves the clarity of image acquisition. The path generated by the system, while satisfying full coverage, can reduce the UAV's battery consumption through variational solution.

[0051] In the real-time obstacle avoidance process of step 4, for mobile hoisting platforms frequently used in steel structure construction, the system uses a semantic segmentation network to identify the hoisting platform category in the point cloud data in real time. Since the motion path of the hoisting platform has a certain predictability, when replanning the manifold curve, the system not only considers the current position of the obstacle, but also combines its motion vector to generate a dynamic no-fly zone that changes over time in the manifold space, achieving early avoidance and avoiding the dynamic impact caused by emergency obstacle avoidance.

[0052] In the execution phase of step 5, the association and storage of inspection data with the building information model (BIM) focuses more on recording stress and deformation data. The system performs high-precision registration between the laser scanning data collected by the UAV and the design model, calculating the deviation vector between the measured members and the theoretical positions. These deviation data are annotated in real time on the digital twin of the BIM, providing accurate data support for the unloading scheme of the steel structure.

[0053] Furthermore, this embodiment also involves a multi-UAV collaborative inspection mode. The system executes a task decomposition algorithm on the global topology map, dividing the massive inspection task into multiple non-overlapping sub-topology regions. Utilizing a distributed attention mechanism, multiple UAVs independently plan manifold curves within their respective manifold subspaces and share dynamic obstacle positions through a cloud-based collaborative mechanism. This parallel operation mode shortens the inspection cycle in the construction of large venues and ensures real-time monitoring of construction progress.

[0054] For construction processes with extremely long cycles, the dynamic progress update function in step 1 played a crucial role. The system uses point cloud data acquired weekly through drone scans of the entire site and employs a deep learning model to automatically identify the proportion of installed grid structures. When the system detects that the actual installation progress is ahead of or behind schedule, it automatically triggers a topology map reconstruction process, reallocating the priority of inspection nodes to ensure that inspection resources are always concentrated in the most active construction areas.

[0055] In terms of intersection logic processing, for the complex junctions of catwalks and maintenance platforms in large venues, the system achieves smooth transitions across multiple levels of paths through central mapping of manifold space. By calculating the changes in normal vectors within three-dimensional space, the system guides the drone to switch altitudes more smoothly in the vertical direction, avoiding instantaneous altitude drops caused by fluctuations in barometer data.

[0056] Example 3: In the specific implementation of underground utility tunnel or tunnel construction and inspection, this invention demonstrates its technical advantages under extremely confined spaces. In step 1, due to the extremely high similarity of the geometric features inside the utility tunnel, the construction of the manifold space adopts a mapping algorithm based on dual constraints of texture features and geometric features. The system extracts pre-embedded channels, fire pipe supports, and lighting fixtures on the utility tunnel wall as auxiliary positioning points, and transforms these physical entities into key anchor points in the manifold space.

[0057] When constructing the topology map in step 2, considering the dim lighting and lack of GPS signal in the underground environment, the system forcibly adds visual feature descriptor vectors to the topology nodes. In addition to geodesic distance, the weight calculation for connecting edges also incorporates the visual positioning robustness evaluation score. If a path lacks sufficient feature points under the manifold representation, potentially causing visual positioning failure, the system automatically increases the weight of that path, forcing the path planning program to find alternative routes with richer features.

[0058] In the manifold curve generated in step 3, the system utilizes the safety margin constraint from the functional extremum principle to address potential slender obstacles such as cables and ventilation ducts within the utility tunnel. This results in a tube-within-a-tube shape, where the path is strictly confined to the safe central region of the utility tunnel's cross-section. The Euler equation solution obtained through numerical iteration ensures that the displacement deviation caused by centrifugal force when the UAV passes through curves remains within the safe envelope.

[0059] During the replanning process in step 4, the system uses an airborne depth camera to conduct real-time detection of temporary construction vehicles commonly found within the utility tunnel. Due to the extremely confined space, the replanning procedure prioritizes searching for obstacle avoidance paths in the vertical direction. If vertical space is insufficient, the system automatically searches for the nearest obstacle avoidance hole or widened turning area in the global topology map and generates a hovering waiting command at that location until the obstacle disappears or is removed from the inspection path.

[0060] Step 5, the command conversion process, incorporates an airflow disturbance compensation algorithm. Due to the piston effect or ground effect that occurs during flight within narrow tunnels, the flight control operator senses abnormal fluctuations in motor output power in real time. The system quickly counteracts the influence of airflow on the UAV's attitude by increasing the gain of the derivative term in the proportional-integral-derivative control.

[0061] The inspection results are stored in the underground scene, focusing on structural safety hazards such as water leakage and cracks. The system uses a semantic segmentation network to analyze the acquired images in real time, identify the geometric outlines of suspected cracks, and accurately project them onto the inner wall of the building information model. Through the time series analysis function of the multidimensional database, managers can observe the widening of cracks at different construction stages, realizing digital early warning for underground projects.

[0062] In this embodiment, the positioning compensation logic based on the fusion of ultrasonic and lidar becomes the core. The UAV uses an ultrasonic array to perceive the precise distance to the surrounding walls in real time within the utility tunnel. A multi-sensor fusion filter (such as an extended Kalman filter) is used to deeply couple the ultrasonic ranging data, visual inertial odometry, and lidar matching results. This tightly coupled positioning method can provide high-precision pose updates at a frequency of no less than 100 Hz even in completely obscured environments, ensuring the UAV's fully autonomous inspection capability within complex utility tunnels.

[0063] Furthermore, the system also features a communication relay logic based on manifold topology. During long-distance utility tunnel inspections, when the signal between the main inspection drone and the exit base station weakens to a critical value, the system assigns a second drone as a communication relay point. The second drone automatically calculates the optimal relay node coordinates based on the manifold position of the first drone and moves along the manifold curve to that position, establishing a stable multi-hop wireless link and ensuring real-time transmission of inspection data.

[0064] In summary, this invention transforms a Building Information Model (BIM) into a manifold space with non-Euclidean geometry and utilizes graph neural networks for topological mapping and path generation, enabling intelligent, high-precision, and high-efficiency inspection of UAVs in various complex construction environments. This method not only theoretically overcomes the limitations of traditional Euclidean spatial coordinate navigation but also provides a complete technical foundation for automated supervision of modern construction projects through real-time edge-side replanning, multi-source sensor fusion positioning, and deep integration with the BIM model.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An automatic inspection path planning method based on Building Information Modeling (BIM) and UAV construction, characterized in that, Includes the following steps: Analyze Building Information Model (BIM) data and construct manifold spatial mapping relationships: Obtain the original BIM model of the construction site, extract the geometric component attributes, spatial topological constraints and construction progress information, establish a three-dimensional spatial coordinate system, and transform the Euclidean spatial coordinates in the BIM model into a manifold spatial representation with curvature characteristics through a non-Euclidean geometric transformation program. Construct a topology mapping model based on graph neural networks: identify key inspection targets in the building information model and physical obstacles in the construction environment, abstract the key inspection targets, construction scaffolding, multi-level intersections and building components as topology nodes, and establish connection edges between nodes based on geodesic distances in the manifold space to generate a global topology mapping graph; Generate an optimal inspection path based on manifold curves: Input the global topology map into a pre-trained graph neural network, use graph convolution operators and attention mechanisms to calculate the inspection priority and coverage weight of each node, and generate an inspection path that meets the optimal coverage requirements based on the principle of functional extrema in manifold space. The inspection path consists of a series of continuous manifold curves. Perform real-time replanning and obstacle avoidance on the UAV side: During the flight of the UAV, environmental point cloud data collected by airborne sensors is acquired in real time. The environmental point cloud data is dynamically compared with the manifold space. If dynamic obstacles or signal interference areas are detected within a preset range, the edge weights of the global topology mapping map are updated, and the manifold curve is replanned in real time with the support of computing power on the UAV side. Output inspection commands and execute automated path guidance: The planned manifold curve is converted into control commands for the UAV flight control system. By controlling the attitude and heading of the UAV, the UAV completes the inspection task along the manifold curve in a confined space, and the inspection data is associated and stored with the building information model.

2. The automatic inspection path planning method based on Building Information Modeling and UAV construction as described in claim 1, characterized in that, The process of parsing building information model data and constructing manifold spatial mapping relationships includes: reading building information models in industrial basic class file format through a standard data interface, and extracting geometric entity information including beams, columns, floor slabs, walls and core tubes; Identify the three-dimensional boundary coordinates of each type of geometric entity, and extract the vertex set and facet connection relationship of the geometric entity in the local coordinate system; By using linear transformations of translation vectors and rotation matrices, the identified 3D boundary coordinates are normalized to a preset global coordinate system; Based on the construction schedule, compare the timestamp information in the construction management database, extract the construction site status at the current time node, and identify completed components, components under construction, and components that have not yet been constructed. Automatically remove component data that has not yet been constructed, and retain the structural information of completed and under construction as the basis for geometric constraints of path planning; The non-Euclidean geometric transformation procedure includes: defining a local coordinate tangent space for each sampling point based on the three-dimensional coordinate system of Euclidean space; The exponential mapping operator maps discrete geometric points in three-dimensional space to the surface of a manifold. The metric tensor of each sampling point on the manifold is calculated. The metric tensor characterizes the degree of physical constraints at the construction site by describing the length scaling ratio of space in different directions, so that the spatial distance metric truly reflects the flight constraints of the UAV in the confined space under the manifold representation.

3. The automatic inspection path planning method based on Building Information Modeling and UAV construction as described in claim 2, characterized in that, The calculation of curvature features in the manifold space includes: performing triangulation processing on the surface of the building information model, discretizing the three-dimensional surface into a grid structure composed of vertices and edges; Calculate the average curvature and Gaussian curvature of each grid vertex; High curvature areas in the construction environment are identified by utilizing curvature distribution; these high curvature areas correspond to sharp corners or slender components at the construction site. When generating manifold representations, mesh refinement is performed on the high curvature regions to provide high-density geometric sampling points by increasing the mesh density in the high curvature regions, thereby improving the geometric accuracy of path planning at complex structures. When calculating the metric tensor, the component values ​​of the metric tensor are adjusted according to the local geometric features of the high curvature region, so that the generated manifold space can characterize the nonlinear constraints of narrow passages and complex turns at the construction site on the flight attitude of the UAV.

4. The automatic inspection path planning method based on Building Information Modeling and UAV construction as described in claim 3, characterized in that, The construction of the topology mapping model based on graph neural network includes: based on the attribute tags of the building information model, using a semantic retrieval algorithm to filter out components that need to be inspected for quality or monitored for safety, the components including embedded parts, steel bar lap joints, concrete pouring surfaces and high-altitude scaffolding. Each key inspection target is assigned a unique topology identifier, which consists of a component type code, a floor code, and a serial number. Based on the inspection frequency required by the construction specifications, initial weight values ​​are set for each key inspection target in the topology mapping model. The generation of the global topology map includes: performing space-filling curve sampling within the manifold space to generate a candidate node set covering the entire construction area; The visibility algorithm is used to determine the visibility between nodes. A virtual ray is emitted from one candidate node to another. If the ray is not obscured by any building component geometry on the manifold surface, it is determined that there is visibility between the two points, and a connecting edge is established in the graph. Calculate the length of the connecting edge, which is the shortest geodesic distance between two points on the surface of the manifold, and the geodesic distance is obtained by solving the path integral on the manifold. Using the geodesic distance as the initial weight of the connecting edges, a graph structure containing a set of nodes and a set of edges is constructed.

5. The automatic inspection path planning method based on Building Information Modeling and UAV construction as described in claim 4, characterized in that, In the process of generating the optimal inspection path based on the manifold curve, the training process of the graph neural network includes: constructing a deep network structure composed of multiple graph convolutional layers, and performing aggregation and updating of node features in each layer; Define an input feature vector, which includes the three-dimensional spatial location components of the node, the component type code, and the associated inspection task urgency level; In the hidden layer, nodes extract global topological features by learning the structural information of their neighboring nodes; The loss function is used to minimize the total path length and the area of ​​the uncovered region, and the network parameters are optimized through the backpropagation algorithm. Calculating inspection priority using attention mechanisms includes introducing self-attention coefficients during the aggregation process of graph convolutions; The self-attention coefficient is calculated based on the cosine similarity between the features of the current node and the features of the neighboring nodes; For nodes with high construction risks or delayed progress, their weight coefficients are increased by manually setting bias terms, so that the attention mechanism prioritizes guiding the drone through the nodes during path planning.

6. The automatic inspection path planning method based on Building Information Modeling and UAV construction as described in claim 5, characterized in that, The method for generating inspection paths based on the principle of functional extrema in manifold space includes: defining an inspection energy functional, which is a weighted combination of a path length term, a coverage completeness term, and a dynamic smoothness term; The path length term represents the cost of the total inspection journey, the coverage completeness term represents the proportion of key inspection targets detected by the sensor's field of view, and the smoothness term constrains the rate of change of the path's curvature. The minimum value of the inspection energy functional is solved by using the variational method, and the Euler equation on the manifold is derived. The Euler equation is solved by numerical iteration to obtain a continuous manifold curve connecting each inspection node. The continuous manifold curve has second-order continuity on the manifold surface. The flow curve has the following dynamic characteristics: the tangent direction of the curve at any point is consistent with the preset heading of the UAV; The radius of curvature of the curve is always greater than the minimum turning radius physical threshold of the drone. As the path passes through the narrow space inside the core tube, the manifold curve automatically contracts towards the manifold centerline to increase the safe physical distance between the drone and the wall.

7. The automatic inspection path planning method based on Building Information Modeling and UAV construction as described in claim 6, characterized in that, The execution of real-time replanning and obstacle avoidance on the UAV side includes: acquiring real-time 3D point cloud data through the lidar and depth camera carried by the UAV; Statistical filtering algorithms are used to remove outlier noise points in the point cloud, and the criterion is whether the average distance between the sampling point and its neighboring points exceeds a preset standard deviation multiple. By applying voxel mesh downsampling technology, the point cloud is discretized into voxel cubes of fixed size, thus reducing the size of the processed point cloud data. The simplified real-time point cloud is quickly registered with the building information model to identify dynamic objects not included in the model, including mobile cranes, construction workers, and temporarily stacked materials. When the dynamic object is detected to have intruded into the predetermined safety radius of the current inspection path, the area occupied by the dynamic object is immediately marked as an unreachable area in the global topology mapping map; Trigger a local path search procedure to find alternative node sequences within the manifold space that bypass unreachable regions; Using a manifold curve fitting algorithm, a local obstacle avoidance curve is regenerated while keeping the connection direction between the start and end points unchanged, and then seamlessly spliced ​​with the original inspection path.

8. The automatic inspection path planning method based on Building Information Modeling and UAV construction according to claim 7, characterized in that, The computation process supported by the UAV edge computing power includes: quantizing and compressing the inference process of the graph neural network so that it runs in the UAV's onboard embedded computing unit; Accelerate the feature update process of topology nodes using a parallel computing architecture; The solution space during replanning is narrowed by a preset heuristic search strategy to ensure that the response time of path update is less than a preset time threshold. For the signal failure area of ​​the global positioning system in the core tube of the super high-rise building, a positioning compensation logic based on the fusion of ultrasonic and lidar is adopted: an additional signal shielding field is defined in the manifold space, and when the UAV enters the signal failure area, it automatically switches to a combined positioning mode based on visual inertial odometry and laser matching. By pre-setting high-precision feature nodes as spatial anchors in the global topology mapping graph, centimeter-level positioning accuracy is maintained in a closed space. In the construction and inspection of underground utility tunnels, the pre-embedded channels, fire pipe supports and lighting fixtures on the tunnel walls are extracted as auxiliary positioning points and transformed into key anchor points in the manifold space. Visual feature descriptor vectors are added to the topological nodes, and the weight calculation of the connecting edges combines the geodesic distance with the visual positioning robustness evaluation score.

9. The automatic inspection path planning method based on Building Information Modeling and UAV construction according to claim 8, characterized in that, In the process of outputting inspection commands and executing automated path guidance, the conversion process of the control commands includes: sampling the planned flow pattern curve at equal intervals to obtain a series of high-frequency waypoint coordinates and their corresponding desired velocity vectors and desired acceleration vectors; The coordinates and vectors are input into the flight control operator of the UAV, and the flight control operator calculates the motor speed command based on model predictive control logic; During the command output process, the drone's operating status is monitored in real time, including battery voltage percentage, propeller speed fluctuations, and communication signal strength. If the battery level is below the safety threshold or the signal attenuation exceeds the predetermined range, the autonomous return procedure will be executed. During long-distance inspections, when the signal between the main inspection drone and the exit base station weakens to a critical value, a second drone is assigned as a communication relay point. The second drone calculates the relay node coordinates based on the manifold position of the main inspection drone and moves to the relay position along the manifold curve to establish a multi-hop wireless link.

10. The automatic inspection path planning method based on Building Information Modeling and UAV construction according to claim 9, characterized in that, The associated storage of the inspection data and the building information model includes: recording the manifold spatial coordinates and camera pose information at the current moment while the UAV is collecting images, videos or point cloud data; Using timestamp alignment technology, sensor data is mapped back to the corresponding components in the building information model; Establish a multidimensional database to store inspection results, and support the retrieval of corresponding historical inspection records by clicking on model components; The method also includes a performance evaluation and dynamic update step: calculating the average spatial deviation between the actual flight trajectory and the planned flow pattern curve; The coverage rate of the statistical inspection target is the ratio of the number of observed nodes to the total number of nodes. Based on the evaluation results, the weight parameters in the graph neural network are automatically adjusted using the backpropagation algorithm. Regularly compare the real-time point cloud data obtained from inspections with the standard geometric model in the building information model, and use a semantic segmentation network to identify the types of components that have been constructed and their spatial locations. When a deviation between the actual construction progress and the planned progress is detected, the path planning program is automatically triggered to update the process and regenerate an inspection topology map that is adapted to the current site environment. In the intersection area, the intersection center of multiple construction channels is identified and set as the topological hub in the manifold space. A smooth multi-branch connection strategy is adopted to ensure that the change in angular velocity of the UAV when switching between different inspection branches is less than the preset physical limit threshold.

Images