Construction quality detection method and system based on image monitoring

Abstract

The invention discloses a construction quality detection method and system based on image monitoring, and relates to the technical field of image processing. The method comprises the following steps: constructing an intelligent sensing network; performing real-time three-dimensional digital twin modeling according to the data acquired by the intelligent sensing network; performing BIM model comparative analysis to generate a difference thermodynamic diagram; carrying out parallel processing on the multi-task deep learning network, and carrying out surface defect detection, structural deviation analysis and material performance evaluation; setting an abnormal response mechanism, grading the severity of the defects, and automatically generating a repair scheme knowledge base; and developing and mixing to realize a quality inspection interface, realizing visualization of defect positions, and automatically generating a detection report. According to the method, after the intelligent sensing network is constructed and the collected data is subjected to multi-source data space-time synchronization, three-dimensional digital twinning modeling is carried out, and intelligent quality is utilized to analyze construction abnormity and generate a quality detection report, so that the multi-scale defect detection precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a construction quality detection method and system based on image monitoring. Background Art

[0002] With the rapid development of intelligent construction technology, construction quality inspection methods based on image analysis have become an important research direction in the field of engineering monitoring. Existing technologies mainly use visual sensors to collect image data during the construction process and combine them with computer vision algorithms to identify surface defects. However, in complex engineering scenarios, traditional methods still face the following technical bottlenecks:

[0003] (1) Insufficient accuracy of multi-scale defect detection

[0004] Existing inspection systems often use single-scale convolutional networks (such as YOLO and Faster R-CNN), which struggle to simultaneously identify micron-scale cracks (<100μm) and macrostructural deformations (>10cm). This is particularly true under interference from lighting variations and surface contamination, as the lack of a cross-scale feature fusion mechanism results in insufficient sensitivity to hidden defects.

[0005] (2) Lack of prediction of dynamic damage evolution

[0006] Current methods mostly focus on static defect identification, fail to establish a correlation model between defect evolution and material performance degradation, and cannot reflect the damage propagation law under construction loads in real time.

[0007] (3) Low level of intelligent repair decision-making

[0008] Existing systems usually only provide defect location annotation and lack a hierarchical response mechanism.

[0009] (4) Weak multi-source data fusion capabilities

[0010] Mainstream detection systems mostly use single-modal visual data and fail to effectively integrate multi-dimensional information such as fiber optic sensing and acoustic emission.

[0011] (5) Imperfect real-time response mechanism

[0012] Existing technologies make it difficult to configure differentiated handling strategies for defects of different severity.

[0013] However, these improved solutions still have technical drawbacks such as high algorithm complexity, imperfect multi-physics field coupling modeling, and delayed knowledge base updates. Therefore, a construction quality inspection method and system based on image monitoring is urgently needed. Summary of the Invention

[0014] The purpose of the present invention is to provide a construction quality detection method and system based on image monitoring. By constructing an intelligent perception network and performing multi-source data spatiotemporal synchronization on the collected data, three-dimensional digital twin modeling is performed, and construction anomalies are analyzed using intelligent quality to generate quality inspection reports, thereby solving the problems of high algorithm complexity, imperfect multi-physical field coupling modeling, and delayed knowledge base updates in existing methods.

[0015] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0016] The present invention is a construction quality detection method based on image monitoring, comprising the following steps:

[0017] Step S1: construct an intelligent sensing network, dynamically deploy composite intelligent sensor nodes at the construction site, and autonomously locate the nodes;

[0018] Step S2: Perform real-time three-dimensional digital twin modeling based on the data collected by the intelligent perception network;

[0019] Step S3: Perform BIM model comparison analysis to generate a difference heat map;

[0020] Step S4: Multi-task deep learning network parallel processing to perform surface defect detection, structural deviation analysis and material performance evaluation;

[0021] Step S5: Set up an exception response mechanism, classify the severity of defects, and automatically generate a repair solution knowledge base;

[0022] Step S6: Develop a hybrid quality inspection interface to visualize defect locations and automatically generate inspection reports.

[0023] As a preferred technical solution, in step S1, the intelligent perception network construction process is as follows:

[0024] Step S11: Obtain the initial node distribution, environmental prior information and UAV cluster status;

[0025] Step S12: Using visual-inertial-lidar tightly coupled SLAM to construct a centimeter-level precision environment point cloud map and generate a semantic feature topology network;

[0026] Step S13: The UAV carries a UWB base station to form an aerial dynamic reference frame;

[0027] Step S14: performing spatiotemporal alignment of multi-source signal data;

[0028] Step S15: performing cross-module feature extraction to generate the optimal observation path for the UAV;

[0029] Step S16: Obtain the spatial three-dimensional coordinates of all nodes.

[0030] As a preferred technical solution, in step S12, the specific process of constructing a centimeter-level precision environment point cloud map and generating a semantic feature topology network is as follows:

[0031] Step S121: Establish a joint calibration model and implement lidar-camera-IMU extrinsic calibration through a checkerboard calibration plate; the joint calibration model is to establish a multi-sensor unified coordinate system through the checkerboard calibration plate and solve the optimal extrinsic matrix Minimize the spatial alignment error between the camera and LiDAR observation data. The specific formula is as follows:

[0032]

[0033] Where, is the rigid transformation matrix from LiDAR to camera, is the point cloud coordinate of the i-th frame laser radar scan, is the camera coordinate observation data corresponding to the i-th frame image, and Project is the operation of projecting the 3D point in the camera coordinate system to the LiDAR coordinate system;

[0034] Step S122: using a hardware-triggered synchronization mechanism to achieve multi-sensor time synchronization, with a time deviation of less than 1ms;

[0035] Step S123: Generate an environmental point cloud based on the environment scanned by the laser radar, including edge / plane feature extraction, motion distortion compensation, and point cloud standard error;

[0036] Step S124: Construct a joint optimization factor graph, and use global optimization to unify the pose estimation at different times and the multi-sensor observation data into a unified model, effectively suppressing the odometer drift problem and achieving centimeter-level positioning accuracy;

[0037] Step S125: Establish a construction element association model based on the graph neural network. The specific formula is as follows:

[0038] A ij =σ(W×[h i ||h j ]);

[0039] Where h i is the feature vector of node i, h j is the eigenvector of node j, A is the adjacency matrix, A ij is the association strength between node i and node j, σ is the activation function, and W is the learnable weight matrix;

[0040] Step S126: Use a multi-machine collaborative mapping algorithm to align multiple point cloud maps, use ORB-SLAM3 to extract ORB feature points, fuse semantic segmentation results (such as component edges, equipment contours) to construct an enhanced feature descriptor, extract edge / plane features based on the Loam-Livox algorithm, use curvature features to encode spatial geometric distribution, and fuse visual-laser features through a cross-modal attention mechanism to generate a hybrid descriptor with semantic labels.

[0041] As a preferred technical solution, in step S15, the specific steps of performing cross-module feature extraction to generate the optimal observation path for the UAV are as follows:

[0042] Step S151: SLAM generates a semantic point cloud map, divides the map environment into 10cm*10cm*10cm voxels, and labels the attributes of each voxel; the attributes of each voxel include: obstacle probability (based on historical observation data), signal propagation quality (calculated based on the building material attenuation model), and observation value score (GDOP impact factor calculated from the target node distribution);

[0043] Step S152: Generate an initial sparse path based on the current position of the UAV (obtained through VIO), with the end point of the path being the center point of the node area to be located;

[0044] Step S153: Dynamically adjust the sampling area weight according to the change of construction progress. The specific formula is as follows:

[0045]

[0046] Where, P sample is the sampling probability density, is in direct proportion, α and β are the geometric accuracy weight coefficient and coverage weight coefficient respectively, and GDOP is the geometric precision dilution;

[0047] Step S154: Evaluate the new node X from the perspectives of geometric accuracy cost, energy cost, and risk cost. new Conduct a multi-dimensional assessment, including:

[0048] The geometric accuracy cost calculation formula is: Where λ is the eigenvalue of the observation matrix;

[0049] The energy cost calculation formula is: C energy =k1×Δh+k2×||Δθ||; where Δh is the altitude change of the UAV and Δθ is the energy consumption of the UAV attitude adjustment;

[0050] The formula for calculating risk cost is: Where, γ i is the obstacle hazard level coefficient;

[0051] Step S155: Introduce the construction machinery motion prediction model to adaptively adjust the UAV's flight radius. The construction machinery motion prediction model establishes a time-series occupancy probability map for dynamic obstacles. When the risk probability of a detected path segment is greater than a threshold, a local replanning scheme is triggered. The formula for adaptively adjusting the flight radius is: Where r rewire is the re-movement radius after dynamic adjustment, r max is the maximum allowable motion radius, v current 、v max are the current flight speed and maximum flight speed of the UAV, r base is the basic movement radius;

[0052] Step S156: Perform curve smoothing on the discrete path points of the UAV flight; Input: The optimized discrete path points; Use 3rd order uniform B-spline for smoothing, add dynamic constraints (maximum curvature ≤ 0.15 rad / m), and ensure the continuity of the path derivative (CC continuity);

[0053] Step S157: Evaluate the discrete sampling points along the path; during the evaluation, the number of visible nodes must be greater than or equal to 3 to ensure the observability of the positioning, and the minimum GODP threshold must be less than 2.5.

[0054] As a preferred technical solution, in step S152, when generating the initial sparse path according to the current position of the UAV, the current root node is set to the current position of the UAV, and the target point is set to the center of mass of the node area; the parameters of the UAV are set as follows: search radius r = 0.5m, maximum number of iterations N = 2000, target bias probability p goal =0.1; dynamic obstacle avoidance is performed by performing temporal safety detection on the sampling points through random sampling. The mixed sampling strategy is: with probability p goal Directly sample the target point x goal , while other cases sample x uniformly in free space rand ; Find the nearest node in the existing tree nodes. The distance is measured using three-dimensional Euclidean distance, and a KD-Tree is used to establish a spatial index. Candidate nodes are generated according to the extension direction of the node and collision detection is performed. The collision detection performs voxel-by-voxel penetration detection on the path segment, and the dynamic obstacle occupancy probability threshold is less than 0.3. Search all reachable nodes and select the node with the smallest path cost as the optimal path. new Search all reachable nodes x with a radius of r as the center near , the optimal parent node selection formula is as follows: Among them, Cost(x) is the cumulative path cost from the root node to x; judge x near Check if the node in x newA better path is available.

[0055] As a preferred technical solution, in step S2, when modeling the three-dimensional digital twin, dual-channel position coding is used to construct a dual-branch network structure; the dual-channel position coding includes spatial coding and temporal coding;

[0056] The spatial encoding is γ(x,y,z)=[sin(2 0 πx),cos(2 0 πx),...,sin(2 L-1 πx)];

[0057] The time code is τ(t)=[sin(w0t),cos(w0t),...,sin(w n t)];

[0058] Where, L = 10, w n is a geometric frequency sequence;

[0059] The dual-branch network structure includes: a dynamic branch and a static branch; wherein, the dynamic branch is used to predict the deformation field of the object; the static branch is used to store the modeling background set and material properties, and realize the lighting consistency constraint between branches through differentiable rendering; for real-time data streams, a sliding time window mechanism is used to select key frames, and the window length adapts to the scene change rate. At the same time, a database of reflective properties of building materials is established, and material properties and surface textures are embedded in the produced three-dimensional digital twin model; and the mechanical motion trajectory is predicted in combination with the operating data of the construction machinery, and an LSTM motion prediction module is constructed to assist in the initialization of the deformation field.

[0060] As a preferred technical solution, in step S3, when performing BIM model comparative analysis to generate a difference heat map, a hierarchical registration strategy is used to extract difference features; the hierarchical registration strategy includes main structure registration, component node registration, and surface texture registration; the difference feature types include geometric deviation (size, shape, position), material deviation (color, texture, reflective characteristics), and process deviation (joint width and welding quality); the differences are quantitatively analyzed, and the degree of difference is mapped using the HSV color space; a multi-dimensional difference index is defined, and the specific formula is as follows:

[0061] D tatal =a×D genmetry +b×D material +c×D process ;

[0062] Where a, b, and c are weight coefficients that are dynamically adjusted according to the construction stage. When mapping the degree of difference in the HSV color space, hue indicates the type of deviation (red = geometric deviation, blue = material difference, etc.), saturation reflects the severity of the deviation, and lightness indicates the confidence level. A 4D difference cube is constructed to visualize the evolution of defects.

[0063] As a preferred technical solution, in step S4, when the multi-task deep learning network is processed in parallel, a multi-task quality analysis network is constructed, and shared features are extracted by ResNeXt-101, including surface defect analysis (segmentation network), structural deformation analysis (regression network) and material performance evaluation (graph neural network), and the material mechanics equation constraints are embedded in the loss function; the constraint equations include: Linear elastic stage: using the generalized Hooke's law σ ij =C ijkl ρ kl ; Damage stage: Introducing damage factor D: Plastic stage: Integrated J2 flow theory: Where C ijkl is the stiffness tensor, s ij is the adaptability tensor; the loss function is constructed as: L total =λ data L data +λ phys L phys +λ bc L bc Where, L data For experimental / sensor data matching, L phys is the residual term of the physical equation, L bc is the boundary condition constraint; data ,λ phys ,λ bc They are data item weights, physical item weights and boundary adjustment weights, and are adaptively adjusted according to the training stage; a strain and visual feature association model is established, regularization terms are introduced into the neural network, a strain-texture association knowledge graph is constructed, and a material defect pattern library is established; a dynamic fusion mechanism is adopted to design a gated neural network, the network nodes are strain sensors and visual feature areas, and the network edges are physical constraint relationships and spatial distances.

[0064] As a preferred technical solution, in step S5, the severity of defects is graded as follows: Level I warning: automatically generate a monitoring report and mark the re-inspection area; Level II response: trigger drone fine scanning and digital twin deduction; Level III disposal: link construction machinery to perform emergency reinforcement and activate the plan.

[0065] The present invention is a construction quality inspection based on image monitoring, including an intelligent perception network, a three-dimensional digital twin modeling unit and an intelligent defect detection unit. The intelligent perception network includes heterogeneous sensors and a multi-source data spatiotemporal synchronization module; the heterogeneous sensors include infrared sensors, high-definition cameras and lidar sensors; the multi-source data spatiotemporal synchronization module is used to trigger the synchronization mechanism through hardware and perform spatiotemporal calibration based on graph optimization; the three-dimensional digital twin modeling unit includes a dynamic three-dimensional modeling module, a construction machinery motion trajectory compensation module and a BIM model comparison module; the dynamic three-dimensional modeling module is used to compare the data collected by the heterogeneous sensors with the BIM model Dynamic three-dimensional modeling is performed on the construction site; the construction machinery motion trajectory compensation module is used to dynamically position and calibrate the construction machinery; the BIM model comparison module is used to match the three-dimensional modeling scene using a real-time differential comparison algorithm; the intelligent defect detection unit includes a multi-scale attention defect detection network, an anomaly propagation anomaly model and a material performance degradation inference model; the multi-scale attention defect detection network is used to capture cross-scale defects in dynamic three-dimensional modeling; the anomaly propagation anomaly model is used to perform evolutionary path deduction and risk assessment on construction quality; the material performance degradation inference model is used to perform life assessment and maintenance decision-making on construction quality.

[0066] The present invention has the following beneficial effects:

[0067] (1) The present invention constructs an intelligent perception network and performs spatiotemporal synchronization of multi-source data on the collected data, performs three-dimensional digital twin modeling, uses intelligent quality analysis to analyze construction anomalies and generate quality inspection reports, thereby improving the accuracy of multi-scale defect detection.

[0068] (2) The present invention generates a three-dimensional digital twin model of the construction site in real time and compares it with the BIM model. It uses the HSV color space to map the degree of difference and constructs a 4D difference cube (3D space + time dimension). This achieves a technological leap from two-dimensional appearance analysis to four-dimensional space-time diagnosis in construction quality inspection, enables visual tracing of the defect evolution process, and improves the recognition rate of minor defects.

[0069] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0071] Figure 1This is a flow chart of a construction quality detection method based on image monitoring according to the present invention;

[0072] Figure 2 The figure is a structural diagram of a construction quality detection system based on image monitoring according to the present invention. DETAILED DESCRIPTION

[0073] 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 any creative efforts shall fall within the scope of protection of the present invention.

[0074] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0075] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figure 1-2 It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0076] Example 1

[0077] See also Figure 1 As shown, the present invention is a construction quality detection method based on image monitoring, comprising the following steps:

[0078] Step S1: construct an intelligent sensing network, dynamically deploy composite intelligent sensor nodes at the construction site, and autonomously locate the nodes;

[0079] Step S2: Perform real-time three-dimensional digital twin modeling based on the data collected by the intelligent perception network;

[0080] Step S3: Perform BIM model comparison analysis to generate a difference heat map;

[0081] Step S4: Multi-task deep learning network parallel processing to perform surface defect detection, structural deviation analysis and material performance evaluation;

[0082] Step S5: Set up an exception response mechanism, classify the severity of defects, and automatically generate a repair solution knowledge base;

[0083] Step S6: Develop a hybrid quality inspection interface to visualize defect locations and automatically generate inspection reports.

[0084] In step S12, the specific process of constructing a centimeter-level precision environment point cloud map and generating a semantic feature topology network is as follows:

[0085] Step S121: Establish a joint calibration model and implement lidar-camera-IMU extrinsic calibration through a checkerboard calibration plate; the joint calibration model is to establish a multi-sensor unified coordinate system through the checkerboard calibration plate and solve the optimal extrinsic matrix Minimize the spatial alignment error between the camera and LiDAR observation data. The specific formula is as follows:

[0086]

[0087] Where, is the rigid transformation matrix from LiDAR to camera, is the point cloud coordinate of the i-th frame laser radar scan, is the camera coordinate observation data corresponding to the i-th frame image, and Project is the operation of projecting the 3D point in the camera coordinate system to the LiDAR coordinate system;

[0088] Use a checkerboard calibration plate with a surface area covering the LiDAR vertical field of view (e.g., 1m × 1m) to ensure that both the LiDAR point cloud and the camera image are fully captured. Perform multi-angle data acquisition by changing the calibration plate's pose (at least 10 different poses) within the common field of view of the LiDAR and camera, and synchronously acquire data including the LiDAR point cloud (including the calibration plate's point cloud), camera images (including the calibration plate's corner points), and IMU inertial measurement data (for time synchronization compensation).

[0089] Preprocess the collected camera data and LiDAR data. The specific preprocessing steps include:

[0090] Call OpenCV's image processing function on the camera data to perform fisheye distortion correction and extract sub-pixel corner coordinates;

[0091] The preprocessing of LiDAR data includes: (1) point cloud filtering: using statistical outlier removal; (2) calibration plate segmentation: extracting calibration plate point cloud based on plane fitting; (3) corner point calculation: using calibration plate set features to solve the coordinates of point cloud corner points;

[0092] When performing LiDAR-Camera-IMU external parameter calibration, the specific process is as follows:

[0093] Step 1: Use direct linear transformation to solve the initial external parameters:

[0094]

[0095] Where, P camis the coordinate of the corner point of the calibration plate in the camera coordinate system, P lidar is the coordinate corresponding to the LiDAR coordinate system;

[0096] Step 2: Construct the reprojection error function:

[0097]

[0098] Where π(·) is the camera projection model, is the pixel coordinate of the image corner point;

[0099] Step 3: Introduce IMU pre-integration constraints and expand the objective function:

[0100]

[0101] Where, is the camera-IMU rotation external parameter, α is the weight coefficient;

[0102] Step S122: using a hardware-triggered synchronization mechanism to achieve multi-sensor time synchronization, with a time deviation of less than 1ms;

[0103] Step S123: Generate an environmental point cloud based on the environment scanned by the laser radar, including edge / plane feature extraction, motion distortion compensation, and point cloud standard error; wherein,

[0104] When extracting edge / plane features, invalid points are filtered based on distance constraints, low-confidence points are eliminated, and then edge feature extraction and plane feature extraction are performed;

[0105] When compensating for motion distortion, we first establish a time synchronization model between the IMU and LiDAR, split the single-frame point cloud into several subframes, calculate the pose of each subframe through IMU pre-integration, and apply pose transformation to each point to complete motion distortion compensation.

[0106] When the point cloud has standard error, feature matching is performed to build a distance residual model: Plane fitting residual calculation formula: Q is a plane reference point, which makes the registration error of the point cloud less than 2 cm;

[0107] Step S124: Construct a joint optimization factor graph. Through global optimization, pose estimates at different times are unified with multi-sensor observation data to model effectively, effectively suppressing odometry drift and achieving centimeter-level positioning accuracy. Establish a unified mathematical framework to integrate the following observation constraints into factor nodes: visual reprojection error, IMU pre-integration constraint, LiDAR point cloud matching residual, and closed-loop detection constraint. This achieves tightly coupled optimization of heterogeneous sensors. Through an incremental optimization mechanism, the factor graph structure is updated in real time to support dynamic obstacle removal and online map correction.

[0108] Step S125: Establish a construction element association model based on the graph neural network. The specific formula is as follows:

[0109] A ij =σ(W×[h i ||h j ]);

[0110] Where h i is the feature vector of node i, h j is the eigenvector of node j, A is the adjacency matrix, A ij is the association strength between node i and node j, σ is the activation function, W is the learnable weight matrix, and || is the splicing operation; by establishing a construction element association model to generate a dynamic topological relationship, h i and h j The features of are spliced ​​to form a joint representation, and the weight matrix W is used to project the spliced ​​features to extract high-order interaction information; σ normalizes the output to quantify the correlation strength between nodes;

[0111] Step S126: A multi-machine collaborative mapping algorithm is used to align multiple point cloud maps, ORB-SLAM3 is used to extract ORB feature points, and semantic segmentation results (such as component edges and equipment contours) are fused to construct an enhanced feature descriptor. Edge / plane features are extracted based on the Loam-Livox algorithm, and spatial geometric distribution is encoded using curvature features. Vision-laser features are fused through a cross-modal attention mechanism to generate a hybrid descriptor with semantic labels. Three groups of matching points are iteratively selected to calculate the SE(3) transformation, and the hypothetical model with an inlier rate of >70% is retained. The improved ICP algorithm is used to achieve sub-pixel registration, and the consistency of component topological relationships is verified through GNN to eliminate conflicting areas.

[0112] In step S152, when generating the initial sparse path based on the current position of the drone, the current root node is set to the current position of the drone, and the target point is set to the center of mass of the node area; the parameters of the drone are set as follows: search radius r = 0.5m, maximum number of iterations N = 2000, target bias probability p goal =0.1; dynamic obstacle avoidance is performed by performing temporal safety detection on the sampling points through random sampling. The mixed sampling strategy is: with probability p goal Directly sample the target point x goal , while other cases sample x uniformly in free space rand ; Find the nearest node in the existing tree nodes. The distance is measured using three-dimensional Euclidean distance, and a KD-Tree is used to establish a spatial index. Candidate nodes are generated according to the extension direction of the node and collision detection is performed. The collision detection performs voxel-by-voxel penetration detection on the path segment, and the dynamic obstacle occupancy probability threshold is less than 0.3. Search all reachable nodes and select the node with the smallest path cost as the optimal path.new Search all reachable nodes x with a radius of r as the center near , the optimal parent node selection formula is as follows: Among them, Cost(x) is the cumulative path cost from the root node to x; judge x near Check if the node in x new A better path is available.

[0113] In step S2, when modeling the three-dimensional digital twin, dual-channel position coding is used to construct a dual-branch network structure; the dual-channel position coding includes spatial coding and temporal coding;

[0114] The spatial encoding is γ(x,y,z)=[sin(2 0 πx),cos(2 0 πx),...,sin(2 L-1 πx)];

[0115] The time code is τ(t) = [sin(w0t), cos(w0t), ..., sin(w n t)];

[0116] Where, L = 10, w n is a geometric frequency sequence;

[0117] The dual-branch network structure includes: dynamic branch and static branch; among them, the dynamic branch is used to predict the deformation field of the object; the static branch is used to store the modeling background set and material properties, and realize the lighting consistency constraint between branches through differentiable rendering; for real-time data streams, a sliding time window mechanism is used to select key frames, and the window length adapts to the scene change rate. At the same time, a database of building material reflection characteristics is established, and the material properties and surface textures of the produced three-dimensional digital twin model are embedded; and the operation data of construction machinery are combined to predict the mechanical motion trajectory, and an LSTM motion prediction module is constructed to assist in the initialization of the deformation field.

[0118] In step S3, when performing BIM model comparison analysis to generate a difference heat map, a hierarchical registration strategy is used to extract difference features. The hierarchical registration strategy includes main structure registration, component node registration, and surface texture registration. The difference feature types include geometric deviation (size, shape, position), material deviation (color, texture, reflective characteristics), and process deviation (joint width and welding quality). The differences are quantitatively analyzed and the degree of difference is mapped using the HSV color space. A multi-dimensional difference index is defined, and the specific formula is as follows:

[0119] D tatal =a×D genmetry +b×D material +c×D process ;

[0120] Where a, b, and c are weight coefficients that are dynamically adjusted according to the construction stage. When mapping the degree of difference in the HSV color space, hue indicates the type of deviation (red = geometric deviation, blue = material difference, etc.), saturation reflects the severity of the deviation, and lightness indicates the confidence level. A 4D difference cube is constructed to visualize the evolution of defects.

[0121] In step S4, when the multi-task deep learning network is processed in parallel, a multi-task quality analysis network is constructed, and shared features are extracted by ResNeXt-101, including surface defect analysis (segmentation network), structural deformation analysis (regression network) and material performance evaluation (graph neural network), and the material mechanics equation constraints are embedded in the loss function; the constraint equations include: Linear elastic stage: using the generalized Hooke's law σ ij =C ijkl ρ kl ; Damage stage: Introducing damage factor D: Plastic stage: Integrated J2 flow theory: Where C ijkl is the stiffness tensor, s ij is the adaptability tensor; the loss function is constructed as: L total =λ data L data +λ phys L phys +λ bc L bc Where, L data For experimental / sensor data matching, L phys is the residual term of the physical equation, L bc is the boundary condition constraint; data ,λ phys ,λ bc They are data item weights, physical item weights and boundary adjustment weights, and are adaptively adjusted according to the training stage; a strain and visual feature association model is established, regularization terms are introduced into the neural network, a strain-texture association knowledge graph is constructed, and a material defect pattern library is established; a dynamic fusion mechanism is adopted to design a gated neural network, the network nodes are strain sensors and visual feature areas, and the network edges are physical constraint relationships and spatial distances.

[0122] In step S5, the severity of defects is graded as follows: Level I warning: automatically generate a monitoring report and mark the re-inspection area; Level II response: trigger drone fine scanning and digital twin simulation; Level III disposal: link construction machinery to perform emergency reinforcement and activate the plan.

[0123] Example 2

[0124] See Figure 2As shown, the present invention is a construction quality detection system based on image monitoring, which can be used to execute the method content of Example 1 of the present invention, including: an intelligent perception network, a three-dimensional digital twin modeling unit and an intelligent defect detection unit. The intelligent perception network includes heterogeneous sensors and a multi-source data spatiotemporal synchronization module; the heterogeneous sensors include infrared sensors, high-definition cameras and lidar sensors; the multi-source data spatiotemporal synchronization module is used to trigger the synchronization mechanism through hardware, and perform spatiotemporal calibration based on graph optimization; the three-dimensional digital twin modeling unit includes a dynamic three-dimensional modeling module, a construction machinery motion trajectory compensation module and a BIM model comparison module; the dynamic three-dimensional modeling .... It is used to dynamically model the construction site based on the data collected by heterogeneous sensors; the construction machinery motion trajectory compensation module is used to dynamically position and calibrate the construction machinery; the BIM model comparison module is used to match the 3D modeling scene using a real-time differential comparison algorithm; the intelligent defect detection unit includes a multi-scale attention defect detection network, an anomaly propagation anomaly model and a material performance degradation inference model; the multi-scale attention defect detection network is used to capture cross-scale defects in dynamic 3D modeling; the anomaly propagation anomaly model is used to deduce the evolution path and conduct risk assessment of construction quality; the material performance degradation inference model is used to conduct life assessment and maintenance decision-making on construction quality.

[0125] It is worth noting that in the above system embodiment, the various units included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of the functional units are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention.

[0126] In addition, those skilled in the art will appreciate that all or part of the steps in the above-mentioned embodiments can be accomplished by instructing related hardware through a program, and the corresponding program can be stored in a computer-readable storage medium.

[0127] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A construction quality detection method based on image monitoring, characterized in that: The steps include: Step S1: construct an intelligent sensing network, dynamically deploy composite intelligent sensor nodes at the construction site, and autonomously locate the nodes; Step S2: Perform real-time three-dimensional digital twin modeling based on the data collected by the intelligent perception network; Step S3: Perform BIM model comparison analysis to generate a difference heat map; Step S4: Multi-task deep learning network parallel processing to perform surface defect detection, structural deviation analysis and material performance evaluation; Step S5: Set up an exception response mechanism, classify the severity of defects, and automatically generate a repair solution knowledge base; Step S6: Develop a hybrid quality inspection interface to visualize defect locations and automatically generate inspection reports.

2. A construction quality detection method based on image monitoring according to claim 1, characterized in that: In step S1, the intelligent perception network construction process is as follows: Step S11: Obtain the initial node distribution, environmental prior information and UAV cluster status; Step S12: Using visual-inertial-lidar tightly coupled SLAM to construct a centimeter-level precision environment point cloud map and generate a semantic feature topology network; Step S13: The UAV carries a UWB base station to form an aerial dynamic reference frame; Step S14: performing spatiotemporal alignment of multi-source signal data; Step S15: performing cross-module feature extraction to generate the optimal observation path for the UAV; Step S16: Obtain the spatial three-dimensional coordinates of all nodes.

3. A construction quality detection method based on image monitoring according to claim 2, characterized in that: In step S12, the specific process of constructing a centimeter-level precision environment point cloud map and generating a semantic feature topology network is as follows: Step S121: Establish a joint calibration model and implement lidar-camera-IMU extrinsic parameter calibration using a checkerboard calibration plate; Step S122: using a hardware-triggered synchronization mechanism to achieve multi-sensor time synchronization; Step S123: generating an environment point cloud based on the environment scanned by the laser radar; Step S124: constructing a joint optimization factor graph; Step S125: Establishing a construction element association model based on the graph neural network; Step S126: using a multi-machine collaborative mapping algorithm to align multiple point cloud maps.

4. The construction quality detection method based on image monitoring according to claim 2 is characterized in that: In step S15, the specific steps of performing cross-module feature extraction to generate the optimal observation path for the UAV are as follows: Step S151: SLAM generates a semantic point cloud map, divides the map environment into voxels, and labels the attributes of each voxel; Step S152: generating an initial sparse path according to the current position of the UAV; Step S153: Dynamically adjust the sampling area weight according to changes in construction progress; Step S154: Perform multi-dimensional evaluation on the new node from three aspects: geometric accuracy cost, energy cost, and risk cost; Step S155: introducing a construction machinery motion prediction model to adaptively adjust the UAV flight radius; Step S156: performing curve smoothing processing on the discrete path points of the UAV flight; Step S157: Evaluate the discrete sampling points along the path.

5. The construction quality detection method based on image monitoring according to claim 4 is characterized in that: In step S152, when generating the initial sparse path based on the current position of the UAV, the current root node is set to the current position of the UAV, and the target point is set to the center of mass of the node area; dynamic obstacle avoidance is performed by performing temporal safety detection on the sampling points through random sampling; Find the nearest node in the existing tree nodes and establish a spatial index; generate candidate nodes based on the extension direction of the node and perform collision detection; search all reachable nodes and select the node with the smallest path cost as the optimal path.

6. The construction quality detection method based on image monitoring according to claim 1 is characterized in that: In step S2, when modeling the three-dimensional digital twin, a dual-channel position encoding is used to construct a dual-branch network structure; The dual-branch network structure includes: a dynamic branch and a static branch; the dynamic branch is used to predict the deformation field of the object; the static branch is used to store the modeling background set and material properties; a sliding time window mechanism is used to select key frames for real-time data streams, and a database of building material reflection characteristics is established at the same time, and material properties and surface textures are embedded in the produced three-dimensional digital twin model; and the mechanical motion trajectory is predicted in combination with the operating data of the construction machinery.

7. The construction quality detection method based on image monitoring according to claim 1 is characterized in that: In step S3, when performing BIM model comparative analysis to generate a difference heat map, a hierarchical registration strategy is used to extract difference features, where the difference feature types include geometric deviation, material deviation, and process deviation. The differences are quantitatively analyzed, and the degree of difference is mapped using the HSV color space. A 4D difference cube is constructed to achieve visual tracing of the defect evolution process.

8. The construction quality detection method based on image monitoring according to claim 1 is characterized in that: In step S4, when the multi-task deep learning network is processed in parallel, a multi-task quality analysis network is constructed, including surface defect analysis, structural deformation analysis and material performance evaluation, and the material mechanics equation constraints are embedded in the loss function to establish a strain and visual feature correlation model.

9. The construction quality detection method based on image monitoring according to claim 1 is characterized in that: In step S5, the severity of the defects is graded as follows: Level I warning: automatically generate a monitoring report and mark the re-inspection area; Level II response: trigger drone fine scanning and digital twin deduction; Level III disposal: link construction machinery to perform emergency reinforcement and activate the plan.

10. A construction quality inspection system based on image monitoring, comprising an intelligent perception network, a three-dimensional digital twin modeling unit and an intelligent defect detection unit, characterized in that: The intelligent perception network includes heterogeneous sensors and multi-source data spatiotemporal synchronization modules; the heterogeneous sensors include infrared sensors, high-definition cameras and lidar sensors; the multi-source data spatiotemporal synchronization module is used to trigger the synchronization mechanism through hardware, and perform spatiotemporal calibration based on graph optimization; the three-dimensional digital twin modeling unit includes a dynamic three-dimensional modeling module, a construction machinery motion trajectory compensation module and a BIM model comparison module; the dynamic three-dimensional modeling module is used to perform dynamic three-dimensional modeling of the construction site based on data collected by heterogeneous sensors; the construction machinery motion trajectory compensation module is used to perform dynamic positioning and calibration of construction machinery; the BIM model comparison module is used to match the three-dimensional modeling scene using a real-time differential comparison algorithm; the intelligent defect detection unit includes a multi-scale attention defect detection network, an anomaly propagation anomaly model and a material performance degradation inference model; the multi-scale attention defect detection network is used to capture cross-scale defects in dynamic three-dimensional modeling; the anomaly propagation anomaly model is used to perform evolutionary path deduction and risk assessment of construction quality; the material performance degradation inference model is used to perform life assessment and maintenance decision-making on construction quality.

Images