Intelligent monitoring method and device for time sequence deformation of building components

Abstract

The application relates to the technical field of building informatization and computer vision, and discloses a kind of intelligent monitoring method and device for time sequence deformation of building component, comprising: obtaining multiple batches of three-dimensional point cloud data of building at different time stages;Coordinate remapping is carried out on each batch of point cloud data to establish a unified building coordinate system, so that the space of different batches of point cloud data is aligned;Extract the building wall point cloud in the point cloud data, and identify the point cloud area of the target component from the wall point cloud, and extract the geometric parameters of the target component;A fusion auto-encoder model containing point cloud and parameter feature encoder, fusion module and decoder is constructed and trained using normal state data;New scanning data is input into the model to calculate the reconstruction error, and according to the degree of deviation of the reconstruction error from the normal state reconstruction error distribution, it is judged whether the target component has deformed or not.The application realizes automatic alignment of multiple batches of data and component-level high-precision deformation detection.

Description

Technical Field

[0001] This invention relates to the field of building information technology and computer vision, and in particular to a method and device for intelligent monitoring of temporal deformation of building components. Background Technology

[0002] With the widespread adoption of Building Information Modeling (BIM) technology, the construction industry is gradually transforming towards digital operation and maintenance. However, the efficient and accurate identification of discrepancies between the actual conditions of the building site and existing digital models or historical records remains a persistent challenge during construction and long-term service. This is especially true for critical building envelope structures such as facades; failure to promptly detect and quantify even minor deformations of components can create safety hazards and increase maintenance costs.

[0003] Traditional building structure monitoring methods suffer from the following drawbacks: reliance on manual visual inspection or contact measurement leads to low efficiency, limited coverage, and difficulty in quantitatively describing minute deformations. When using 3D scanning technology for multi-batch data acquisition, point cloud data acquired at different times suffer from inherent coordinate system inconsistencies due to differences in equipment position and orientation, i.e., coordinate drift, resulting in a lack of reliable spatial benchmarks for change detection. Currently, there are point cloud processing methods for building facade components, such as coarse calibration based on principal component analysis, plane extraction based on random sampling consensus algorithms, and component segmentation schemes based on density clustering. However, these schemes still involve static geometric segmentation of single-scan data, only acquiring the spatial position and geometric contour of components, and lacking the ability to detect deformations across time periods. Furthermore, in the field of change detection, research has proposed a construction site change recognition framework based on deep learning semantic segmentation networks, but its technical approach focuses on scene-level semantic change discovery, failing to extend to the deformation-based output of specific building components, and the network architecture used is not optimized for component geometry and attribute reconstruction.

[0004] Existing detection methods are generally based on single-modal data representation. For example, relying solely on pure geometric deviation measurements of component point clouds is insufficiently sensitive to minute deformations in complex environments and is susceptible to false alarms due to scanning noise; while using statistical parameter comparisons alone ignores the completeness of spatial structural information. The limitations of this single-modal approach make it difficult to comprehensively reflect the true deformation state of building components. Summary of the Invention

[0005] Based on this, it is necessary to provide an intelligent monitoring method and device for the temporal deformation of building components to address the problems of inconsistent coordinate references in multiple batches of scanning data and the difficulty of accurately characterizing minute deformations of components using a single mode, so as to achieve automated, high-precision temporal deformation detection and quantitative analysis of building components.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for intelligent monitoring of temporal deformation of building components includes: Acquire multiple batches of 3D point cloud data of buildings at different time stages; Coordinate remapping is performed on the point cloud data of each batch to establish a unified building coordinate system and spatially align the point cloud data of different batches. Extract the point cloud of the building wall from the point cloud data, identify the point cloud region of the target component from the wall point cloud, and extract the geometric parameters of the target component. A fusion autoencoder model is constructed and trained using target component data under normal conditions. The fusion autoencoder model includes: a point cloud feature encoder for extracting geometric feature vectors from the point cloud region of the target component; a parametric feature encoder for extracting parametric feature vectors from the geometric parameters; a fusion module for mapping the geometric feature vectors and the parametric feature vectors to a latent space to obtain a joint feature representation; and a decoder for reconstructing the point cloud and geometric parameters based on the joint feature representation. The newly scanned target component data is input into the trained fusion autoencoder model to obtain the reconstruction result and calculate the reconstruction error. Based on the degree to which the reconstruction error deviates from the normal state reconstruction error distribution, it is determined whether the target component has undergone deformation.

[0007] In one embodiment, the coordinate remapping of each batch of point cloud data to establish a unified building coordinate system includes: Perform RANSAC plane fitting on the top floor plan of the building, extract the top floor plan, and calculate its normal vector. Using the normal vector as the new coordinate system The axial direction is determined by the right-hand rule. shaft and Establish a unified architectural coordinate system along the axial direction; according to Coordinate transformation was performed on each batch of point cloud data, where The original point cloud coordinates, The center point of the top floor plan, Let be a rotation matrix, such that axis and the normal vector Alignment.

[0008] In one embodiment, the step of extracting the building wall point cloud from the point cloud data and identifying the point cloud region of the target component from the wall point cloud includes: Multi-plane RANSAC segmentation is performed on the building point cloud after coordinate remapping to extract the planes of each wall surface. The wall point cloud is then separated into main plane parts and non-main plane parts by the plane normal vector direction. DBSCAN density clustering is performed on the non-primary planar portion to obtain candidate component regions; Project the 3D point cloud of the candidate component region onto the wall plane, calculate the 2D convex hull of the projected point cloud, find the largest inscribed rectangle within the convex hull, and extract the 3D dimensions of the axis-aligned bounding box of the candidate component region as the length. ,Width ,thick ; Threshold filters are set for various parameters to identify the point cloud region of the target component.

[0009] In one embodiment, the geometric parameters are seven-dimensional parameters, including: length ,Width ,thick , points ,density Tightness Area ratio ,in This represents the area of ​​the largest inscribed rectangle. Let be the area of ​​the two-dimensional convex hull.

[0010] In one embodiment, the point cloud feature encoder is a PointNet encoder with a three-layer shared MLP structure that outputs geometric feature vectors through global max pooling; the parametric feature encoder is a two-layer fully connected layer that outputs parametric feature vectors; the fusion module concatenates the geometric feature vectors and parametric feature vectors and maps them to the latent space through the MLP to obtain a joint feature representation; the decoder includes a point cloud decoding branch and a CSV decoding branch, which reconstruct the point cloud and geometric parameters from the joint feature representation, respectively.

[0011] In one embodiment, the fused autoencoder model is trained using target component data under normal conditions, employing a joint loss function: ,in For point cloud reconstruction loss, The mean square error of the geometric parameter reconstruction loss. This is the balance coefficient.

[0012] In one embodiment, determining whether the target component has deformed based on the degree to which the reconstruction error deviates from the normal reconstruction error distribution includes: Calculate the reconstruction error from the training data under normal conditions and calculate its mean. and standard deviation Set deformation threshold , This is the sensitivity coefficient; When the reconstruction error corresponds to the new scanned data When the time is right, it is determined that the component has deformed.

[0013] In one embodiment, when it is determined that the target component has deformed, the relative rate of change of each geometric parameter is calculated. , In the formula, The parameter values ​​corresponding to the new scan data. The parameter values ​​are those corresponding to the same component under normal historical conditions; the deformation level is output by combining the degree of deviation of reconstruction error and the relative change rate of each geometric parameter.

[0014] The present invention also provides an intelligent monitoring device for time-series deformation of building components, comprising: The data acquisition module is used to acquire multiple batches of 3D point cloud data of buildings at different time stages; The coordinate remapping module is used to remap the coordinates of point cloud data from different batches to establish a unified building coordinate system and spatially align the point cloud data from different batches. The component recognition module is used to extract the point cloud of the building wall from the point cloud data, identify the point cloud region of the target component from the wall point cloud, and extract the geometric parameters of the target component. The fusion autoencoder model includes: a point cloud feature encoder for extracting geometric feature vectors from the point cloud region of the target component; a parametric feature encoder for extracting parametric feature vectors from the geometric parameters; a fusion module for mapping the geometric feature vectors and the parametric feature vectors to a latent space to obtain a joint feature representation; and a decoder for reconstructing the point cloud and geometric parameters based on the joint feature representation. The deformation detection module is used to input the newly scanned target component data into the trained fusion autoencoder model, calculate the reconstruction error, and determine whether the target component has deformed based on the degree to which the reconstruction error deviates from the normal state reconstruction error distribution.

[0015] In one embodiment, a visualization output module is also included for generating a three-dimensional visualization model and inspection report that distinguishes deformed components from normal components with different colors.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: The aforementioned intelligent monitoring method and device for time-series deformation of building components establishes a unified building coordinate system to remap the coordinates of point cloud data from different batches. This allows point cloud data acquired at different time stages to be directly compared under the same spatial reference, eliminating coordinate deviations introduced by differences in device position and orientation, and providing a reliable foundation for component-level time-series analysis. Based on this, the point cloud region of the target component is identified from the point cloud of the building wall, refining deformation detection from the overall building or wall level to individual components.

[0017] To address the shortcomings of existing methods that rely solely on single-modal data, resulting in insufficient sensitivity to minute deformations and susceptibility to noise interference, a fusion autoencoder model is constructed. This model extracts spatial geometric feature vectors from components using a point cloud feature encoder and geometric parameter feature vectors using a parametric feature encoder. A fusion module maps these two modal features to the same latent space to obtain a joint feature representation. This allows the model to simultaneously learn the normal state distribution of both the component's geometric shape and statistical properties. The decoder reconstructs the point cloud and geometric parameters based on this joint feature representation, ensuring that the reconstruction error comprehensively reflects the deviation of the component's spatial structure and statistical characteristics from its normal state. When newly scanned target component data is input into the trained fusion autoencoder model, if the calculated reconstruction error significantly deviates from the distribution range of normal state reconstruction errors, the component is determined to have deformed. Thus, based on unified coordinates and component recognition, and through multimodal feature fusion and reconstruction error-driven methods, highly sensitive automated detection and discrimination of minute deformations in building components is achieved. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the intelligent monitoring method for temporal deformation of building components according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the intelligent monitoring device for time-series deformation of building components according to an embodiment of the present invention; Figure 3 This is a diagram of the internal structure of the electronic device in this embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0022] The following is combined Figures 1-3 The present invention describes a method and apparatus for intelligent monitoring of temporal deformation of building components.

[0023] like Figure 1 As shown, in one embodiment, a method for intelligent monitoring of temporal deformation of building components combines geometric algorithms with deep learning models to achieve automated deformation detection of building components. The method includes the following steps: Step S110: Obtain multiple batches of 3D point cloud data of the building at different time stages.

[0024] Alternatively, the target building can be periodically photographed using drones or ground-based 3D laser scanners to obtain raw point cloud sets reflecting the building's state at different times. Due to differences in the attitude, position, and environmental factors of the acquisition equipment, point cloud data from different batches are usually in independent local coordinate systems. Direct comparison will produce systematic deviations, so subsequent spatial alignment processing is required.

[0025] Step S120: Perform coordinate remapping on each batch of point cloud data to establish a unified building coordinate system and spatially align the point cloud data from different batches.

[0026] The coordinate remapping process includes: performing RANSAC plane fitting on the building's top floor plan, extracting the top floor plan, and calculating its normal vector. Using the normal vector as the new coordinate system The axial direction is determined by the right-hand rule. shaft and Construct a unified architectural coordinate system along the axis; according to Coordinate transformation was performed on each batch of point cloud data, where The original point cloud coordinates, The center point of the top floor plan, Let be a rotation matrix, such that Axis and normal vector Alignment. The top floor plan exhibits the highest structural stability throughout the building's lifespan, making it the least affected by local renovations or decorative obstructions. The top floor plan normal vector fitted using the RANSAC (RANdom Sampling Consensus) algorithm accurately reflects the building's overall vertical orientation. The new coordinate system constructed using this normal vector has clear physical meaning and, compared to traditional PCA principal component analysis, effectively avoids axial offset caused by asymmetric point cloud distribution. Mapping all historical scan data to a unified building coordinate system according to the same transformation rules eliminates global coordinate drift caused by differences in acquisition conditions, providing a precise spatial reference for subsequent time-series deformation comparisons.

[0027] Step S130: Extract the point cloud of the building wall from the point cloud data, identify the point cloud region of the target component from the wall point cloud, and extract the geometric parameters of the target component.

[0028] To achieve refined analysis from the overall building down to individual components, this step further includes: performing multi-plane RANSAC segmentation on the coordinate-remapped building point cloud, extracting each wall plane, and separating the wall point cloud into primary and non-primary plane parts based on the plane normal vector direction; performing DBSCAN (Density-Based Spatial Clustering of Applications with Noise) density clustering on the non-primary plane parts to obtain candidate component regions; projecting the 3D point cloud of the candidate component region onto the corresponding wall plane, calculating the 2D convex hull of the projected point cloud, finding the largest inscribed rectangle within the convex hull, and extracting the 3D dimensions of the axis-aligned bounding box of the candidate component region as the length. ,Width ,thick Threshold filters are set for various parameters to identify the point cloud region of the target component.

[0029] Specifically, multi-planar RANSAC segmentation can adapt to complex scenarios where building facades have multiple facing walls. By iteratively extracting and removing identified planar point clouds, individual wall faces can be separated. The normal vector direction is used to distinguish between primary planes (the main body of the wall) and non-primary planes (windows, recesses, etc.), achieving coarse-grained semantic separation based on geometric priors. For non-primary planes, the DBSCAN density clustering algorithm is used, whose neighborhood radius eps can be adaptively adjusted according to the local point cloud density, and the minimum number of points minPts can be set to 100, thus effectively identifying discrete candidate component clusters. By projecting the 3D point cloud onto the 2D plane of the wall and calculating the convex hull and the maximum inscribed rectangle, the regular geometric boundaries of the components can be estimated relatively stably, and usable dimensional parameters can still be obtained even when the point cloud has local missing points or noise.

[0030] Based on this, the extracted geometric parameters are seven-dimensional parameters, including: length ,Width ,thick , points ,density Tightness Area ratio ,in This represents the area of ​​the largest inscribed rectangle. Let be the area of ​​the two-dimensional convex hull.

[0031] These seven parameters describe the target component from different dimensions. Length, width, and thickness describe the macroscopic scale of the component; the number of points reflects the scanning coverage; density and compactness reflect the integrity of the component's surface and the solid fill rate, with significant changes in density and compactness when the component is damaged or missing; and the area ratio measures the regularity of the component's shape, which can be used to distinguish windows from other irregular attachments. The geometric and statistical attribute parameters work together to provide complementary information for subsequent deep learning models, helping to overcome the limited information content of simple point cloud features when expressing minute deformations.

[0032] Step S140: Construct a fusion autoencoder model and train it using target component data under normal conditions.

[0033] The fusion autoencoder model includes: a point cloud feature encoder for extracting geometric feature vectors from the point cloud region of the target component; a parametric feature encoder for extracting parametric feature vectors from the geometric parameters; a fusion module for mapping the geometric feature vectors and parametric feature vectors to the latent space to obtain a joint feature representation; and a decoder for reconstructing the point cloud and geometric parameters based on the joint feature representation.

[0034] In its specific implementation, the point cloud feature encoder is a PointNet encoder with a three-layer shared MLP (Multi-Layer Perceptron) structure that outputs geometric feature vectors through global max pooling; the parametric feature encoder consists of two fully connected layers that output parametric feature vectors; the fusion module concatenates the geometric feature vectors and parametric feature vectors and maps them to the latent space through the MLP to obtain a joint feature representation; the decoder includes a point cloud decoding branch and a CSV decoding branch, which reconstruct the point cloud and geometric parameters from the joint feature representation, respectively.

[0035] For example, the PointNet encoder can employ a three-layer shared MLP structure with dimensions ranging from 64 to 128 to 1024. Each layer is followed by BatchNorm and ReLU activation functions, ultimately obtaining a 1024-dimensional geometric feature vector through global max pooling. This structure exhibits permutation invariance and is suitable for processing unordered point clouds. The parametric feature encoder uses a two-layer fully connected network with dimensions ranging from 7 to 32 to 64, compressing the seven-dimensional geometric parameters into a 64-dimensional parametric feature vector. The fusion module concatenates the 1024-dimensional geometric features with the 64-dimensional parametric features to form a 1088-dimensional vector, which is then mapped to a 256-dimensional latent space through a 1088→512→256-dimensional MLP, forming a compact joint feature representation. The dual-branch decoder reconstructs the point cloud through a 256→512→1024→N×3-dimensional MLP and reconstructs the geometric parameters through a 256→128→7-dimensional MLP, respectively. Through the aforementioned fusion structure, the model simultaneously encodes spatial geometric information and statistical attribute information in the latent space. Compared with autoencoders that use only a single modality, it can more fully learn the feature distribution of normal components, which is beneficial to improving the ability to distinguish abnormal states.

[0036] During training, the fusion autoencoder model is trained using target component data under normal conditions, employing a joint loss function: , in For point cloud reconstruction loss, The mean square error of the geometric parameter reconstruction loss. This is the balance coefficient.

[0037] Specifically, Chamfer Distance measures the geometric distance between the reconstructed point cloud and the original point cloud, used to control the accuracy of shape reconstruction; MSE measures the numerical deviation between the reconstructed parameters and the original parameters, used to maintain the consistency of statistical properties. Balance coefficient. The weight ratio of the two losses is used to adjust the ratio, for example, it can be set to 0.5, to prevent one loss from dominating the gradient update direction. Through joint optimization, the model learns the intrinsic relationship between the geometric reconstruction of normal components and various geometric parameters during training. When a deformed component is input, both reconstruction errors will increase, which helps to improve detection sensitivity and reduce noise interference.

[0038] Step S150: Input the newly scanned target component data into the trained fusion autoencoder model to obtain the reconstruction result and calculate the reconstruction error. Based on the degree to which the reconstruction error deviates from the normal state reconstruction error distribution, determine whether the target component has deformed.

[0039] To avoid the subjectivity of manually setting fixed thresholds, the judgment process specifically includes: calculating the reconstruction error of the training data under normal conditions and calculating its mean. and standard deviation Set deformation threshold , This is the sensitivity coefficient; when the reconstruction error corresponds to the new scan data When the component is deformed, it is determined that it has deformed. This mechanism is based on statistical principles, assuming that the reconstruction error of a normal component follows an approximately normal distribution, and low-probability events exceeding this range are considered abnormal. The sensitivity coefficient can be adjusted according to actual engineering needs. For example, in scenarios with high safety requirements, a value of 2 can be used to improve the recall rate, while a value of 3 can be used in noisy environments to reduce the false alarm rate. This threshold setting method can be adjusted according to the noise levels of different buildings and different sensors, making it widely applicable.

[0040] When determining whether the target component has deformed, the relative rate of change of each geometric parameter is also calculated. , In the formula, Representing various geometric parameters, The parameter values ​​corresponding to the new scan data. The parameter values ​​correspond to the same component under normal historical conditions; the deformation level is output by combining the degree of reconstruction error deviation and the relative change rate of each geometric parameter. Specifically, in addition to qualitatively determining whether deformation has occurred, the severity of deformation also needs to be quantitatively assessed. By calculating the relative change rate of key dimensions such as length, width, and thickness, the physical deformation amplitude of the component can be intuitively reflected. This is combined with the multiple by which the reconstruction error exceeds the threshold (i.e., This allows for the establishment of a multi-dimensional deformation rating system, such as dividing it into four levels: normal, slight, moderate, and severe. The final output test results not only include deformation status indicators but also specific displacement information and percentage changes in attributes, providing maintenance personnel with interpretable and traceable decision-making support.

[0041] This embodiment covers processing steps such as coordinate benchmark unification, component segmentation, multimodal feature fusion, and deformation transformation. The coordinate system established using the top-level plane normal vector solves the problem of coordinate inconsistencies between scan data from different periods; the joint description of seven-dimensional geometric parameters and point cloud features compensates for the limited information of a single modality; and the threshold setting method based on statistical distribution enables the detection process to adapt to different environmental conditions. The entire process requires no manual annotation and is suitable for long-term health monitoring of large-scale building complexes.

[0042] The intelligent monitoring device for time-series deformation of building components provided by the present invention is described below. The intelligent monitoring device for time-series deformation of building components described below can be referred to in correspondence with the intelligent monitoring method for time-series deformation of building components described above.

[0043] like Figure 2 As shown, in one embodiment, a building component temporal deformation intelligent monitoring device completes the processing from data acquisition to deformation early warning through the cooperation of various functional modules. The device includes a data acquisition module 210, a coordinate remapping module 220, a component identification module 230, a fusion autoencoder model 240, a deformation detection module 250, and a visualization output module 260.

[0044] The data acquisition module 210 is used to acquire multiple batches of 3D point cloud data of a building at different time stages. Specifically, this module can be a physical interface connecting to a 3D laser scanner or UAV storage medium, or a file parsing unit that reads local or cloud databases. This module can automatically identify and load point cloud file sequences with timestamp information, organizing discrete raw scan data into an ordered time-series dataset, providing standardized data input for subsequent comparative analysis. In actual deployment, this module can also perform integrity verification and format conversion on the point cloud data, ensuring that the data entering subsequent processing conforms to preset specifications.

[0045] The coordinate remapping module 220 is used to remap the coordinates of point cloud data from different batches to establish a unified building coordinate system, thus aligning the spatial data of different batches. This module executes the coordinate transformation logic described in the aforementioned embodiment. Specifically, it integrates a RANSAC plane fitting algorithm unit and a matrix operation unit, which can automatically calculate the normal vector of the building's top-floor plane from the original point cloud. and center point And construct a rotation matrix accordingly. By performing operations on each batch of point clouds The transformation module eliminates global coordinate drift caused by differences in the pose of the acquisition device, and outputs point cloud data that is precisely aligned under a unified physical reference, thereby enabling spatial consistency in temporal deformation detection.

[0046] The component identification module 230 is used to extract the point cloud of building walls from the point cloud data, identify the point cloud region of the target component from the wall point cloud, and extract the geometric parameters of the target component. This module receives the aligned point cloud after coordinate remapping and is used to execute the segmentation and parameter extraction process described in the aforementioned embodiment. Specifically, this module includes a multi-plane segmentation sub-unit, a density clustering sub-unit, and a geometric calculation sub-unit. The multi-plane segmentation sub-unit separates independent wall surfaces through iterative RANSAC; the density clustering sub-unit uses the DBSCAN algorithm to locate candidate component clusters from non-primary planes; and the geometric calculation sub-unit outputs a parameter containing the length of the wall surface through two-dimensional projection, convex hull calculation, and maximum inscribed rectangle search. ,Width ,thick The system extracts seven geometric parameters, including point count, density, compactness, and area ratio. By employing the above parameter extraction method, the unstructured point cloud is transformed into a structured parameter vector, making it easier for deep learning models to use directly.

[0047] The fusion autoencoder model 240 includes a point cloud feature encoder, a parametric feature encoder, a fusion module, and a decoder. This model is loaded into the device's memory as a pre-trained model. The point cloud feature encoder extracts geometric feature vectors from the point cloud region of the target component; the parametric feature encoder extracts parametric feature vectors from the geometric parameters; the fusion module maps the geometric feature vectors and parametric feature vectors to the latent space to obtain a joint feature representation; and the decoder reconstructs the point cloud and geometric parameters based on the joint feature representation. Specifically, upon receiving the point cloud fragment and seven-dimensional parameters output by the component recognition module, the model performs parallel dual-path encoding and fuses spatial topological information and statistical attribute information into a unified latent variable representation in the latent space. Subsequently, the decoder reconstructs the ideal point cloud and parameters under normal conditions based on this latent variable. Through this fusion structure, the device can not only detect macroscopic dimensional changes but also identify local geometric distortions, which is beneficial for improving the ability to recognize complex deformation patterns.

[0048] The deformation detection module 250 is used to input newly scanned target component data into the trained fusion autoencoder model, obtain reconstruction results, calculate reconstruction error, and determine whether the target component has deformed based on the degree to which the reconstruction error deviates from the normal reconstruction error distribution. This module is used to execute the anomaly detection and quantization logic of the aforementioned embodiment. Specifically, it internally maintains the mean error value obtained based on historical normal data statistics. and standard deviation And based on the sensitivity coefficient Dynamically generate deformation threshold When the reconstruction error is calculated in real time When the threshold is exceeded, the module triggers a deformation alarm and further calculates the relative rate of change of each geometric parameter (e.g., ...). The device comprehensively assesses the deformation level. This threshold discrimination method based on statistical distribution allows the device to adjust its detection sensitivity according to the environmental noise level without human intervention, thus balancing the control of false alarm and false alarm rates.

[0049] The visualization output module 260 is used to generate 3D visualization models and inspection reports that distinguish deformed components from normal components using different colors or representational methods. Figure 2 The 3D visualization model uses a color-coding system to visually identify component states: the left frame represents deformed components, and the grayscale entities on the right represent normal components. As the final point of contact for human-computer interaction, this module transforms detected values ​​into intuitive visual information. Specifically, the generated 3D visualization model uses green to visually indicate component states: for example, green is used to render deformed components, and blue to render components in a normal state, allowing maintenance personnel to easily locate risk points. Simultaneously, the module automatically generates structured detection reports (such as the window_changes.csv file), covering indicators such as component number, deformation level, displacement, and percentage changes in various attributes. The combined output of the visualization model and the detection reports makes the detection results easier to understand and archive.

[0050] Through the coordination and data flow of the aforementioned modules, the device provided in this embodiment realizes the technical logic of the intelligent monitoring method for temporal deformation of building components and implements it as a deployable product. The functional boundaries between each module are clear, facilitating independent upgrades and maintenance. It can perform automated processing such as coordinate alignment, feature extraction, fusion reconstruction, and visualization output, and can be used for long-term health monitoring of building structures.

[0051] In a specific embodiment, the time-series deformation monitoring of window components on a building facade is used as a specific application scenario to further illustrate the intelligent monitoring method for time-series deformation of building components described in the aforementioned embodiment. It should be understood that the specific parameter settings and object selections in this embodiment are merely illustrative examples, intended to verify the feasibility of the technical solution, and do not constitute a limitation on the scope of protection of this invention.

[0052] In this application scenario, step S110 is first executed to acquire multiple batches of 3D point cloud data of the building at different time stages. Specifically, a drone equipped with a 3D laser scanner is used to periodically scan the exterior of the target building, with the point cloud sampling volume for a single scan set at 5,000,000 points. This sampling density, while ensuring data coverage integrity, can control the computational load of subsequent processing, providing data conditions for subsequent detection of subtle geometric changes in window components.

[0053] Then, step S120 is executed to remap the coordinates of each batch of point cloud data to establish a unified building coordinate system, thus aligning the point cloud data from different batches in space. In this scenario, the RANSAC algorithm is used to fit the top floor plan of the building and calculate the normal vector. and center point And construct the rotation matrix Using the transformation formula The process unifies 5 million point cloud data collected at different times under the same physical benchmark. This process eliminates coordinate drift caused by differences in UAV flight attitude, ensures spatial consistency for subsequent time-series comparisons, and provides a unified geometric benchmark for deformation detection at different times.

[0054] Next, step S130 is executed to extract the building wall point cloud from the point cloud data, identify the point cloud region of the target component from the wall point cloud, and extract the geometric parameters of the target component. For facade window identification, when performing DBSCAN density clustering on non-primary planar parts, the neighborhood radius eps adopts an adaptive strategy based on local point cloud density to address the density unevenness caused by changes in UAV scanning distance; the minimum number of points minPts is set to 100, which can effectively filter out wall noise and interference from small attachments, and locate candidate window regions. Based on this, the extracted geometric parameters are seven-dimensional parameters, including length... ,Width ,thick , points ,density Tightness Area ratio This set of parameters includes not only the macroscopic size information of the window, but also quantifies the integrity and regularity of the components through density, compactness, and area ratio. It provides multi-dimensional statistical information for subsequent deep learning models, which helps to make up for the limited information content of pure point cloud features when expressing small deformations.

[0055] In the model building and training phase, step S140 is executed to construct the fused autoencoder model and train it using window data under normal conditions. In this application scenario, the PointNet encoder outputs a 1024-dimensional geometric feature vector, the parametric feature encoder outputs a 64-dimensional parametric feature vector, and the fusion module maps them to a 256-dimensional latent space. The joint loss function used during training is... The balance coefficient is set to 0.5. This value is determined based on the dimensional differences between the geometric features and statistical properties of window components. This ensures that the model focuses on a balanced approach to point cloud shape reconstruction and parameter value reconstruction during training, preventing the loss of one aspect from dominating gradient updates and leading to insufficient learning of other modal features. Through this joint training, the model can learn the inherent correlation between the spatial geometry and statistical properties of a normal window, which is beneficial for improving the sensitivity of deformation detection.

[0056] In the deformation detection and quantization stage, step S150 is executed, where newly scanned window data is input into the trained fusion autoencoder model. The model determines whether deformation has occurred based on the degree of reconstruction error deviation and calculates the relative rate of change. Specifically, a deformation threshold is set based on the statistical distribution of reconstruction errors from normal window samples. The sensitivity coefficient In this scenario, the parameter is adjustable and takes a value between 2 and 3. When the system exhibits a higher recall rate for minute deformations, it is suitable for safety-sensitive monitoring; when At this time, the system has stronger noise resistance and is suitable for scenarios with significant environmental interference. For windows determined to be deformed, the relative change rate of each geometric parameter is further calculated, such as... The system then comprehensively reconstructs the deformation level by calculating the error deviation factor. Finally, the system automatically generates a detection report named window_changes.csv, which includes the component number, deformation level, displacement, and percentage change of each attribute. In the 3D visualization model, deformed windows are marked in green and normal windows are marked in blue.

[0057] Through the implementation of the above specific application scenarios, the combination of geometric algorithms and deep learning has been validated. The sampling of 5 million point clouds and adaptive DBSCAN parameters ensures a relatively stable component extraction process; the fusion of seven-dimensional geometric parameters and point cloud features improves the model's ability to identify minute deformations; and the threshold setting method based on statistical distribution allows for adjustment of detection sensitivity according to actual needs. The entire process, from data acquisition to report output, is automated, requiring no manual annotation or intervention. It solves problems such as incomparability of multiple batches of data, difficulty in detecting minute deformations, and low efficiency of manual inspections in traditional methods, making it suitable for long-term health monitoring of building facade components.

[0058] Figure 3 This example illustrates a schematic diagram of the physical structure of an electronic device, which can be a smart terminal. Its internal structure diagram can be as follows: Figure 3 As shown, the electronic device includes a processor, memory, and network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the intelligent monitoring method for temporal deformation of building components according to any of the above embodiments.

[0059] Those skilled in the art will understand that Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0060] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0061] By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for intelligent monitoring of temporal deformation of building components, characterized in that, include: Acquire multiple batches of 3D point cloud data of buildings at different time stages; Coordinate remapping is performed on the point cloud data from each batch to establish a unified building coordinate system, thereby aligning the point cloud data from different batches in space. This includes: performing RANSAC plane fitting on the top floor plane of the building, extracting the top floor plane, and calculating its normal vector. Using the normal vector as the new coordinate system The axial direction is determined by the right-hand rule. shaft and A unified architectural coordinate system is established along the axial direction; according to Coordinate transformation was performed on each batch of point cloud data, among which... The original point cloud coordinates, The center point of the top floor plan, Let be a rotation matrix, such that axis and the normal vector Alignment; Extracting point clouds from building wall surfaces in point cloud data, identifying point cloud regions of target components from the wall surface point clouds, and extracting the geometric parameters of the target components include: performing multi-plane RANSAC segmentation on the coordinate-remapped building point clouds, extracting each wall plane, and separating the wall surface point clouds into main plane parts and non-main plane parts by plane normal vector directions; performing DBSCAN density clustering on the non-main plane parts to obtain candidate component regions; projecting the 3D point cloud of the candidate component region onto the corresponding wall plane, calculating the 2D convex hull of the projected point cloud, finding the largest inscribed rectangle within the convex hull, and extracting the 3D dimensions of the axis-aligned bounding box of the candidate component region as the length. ,Width ,thick Threshold filters are set for various parameters to identify the point cloud region of the target component; geometric parameters are extracted into seven-dimensional parameters, including: length ,Width ,thick , points ,density Tightness Area ratio ,in This represents the area of ​​the largest inscribed rectangle. The area of ​​the two-dimensional convex hull; A fusion autoencoder model is constructed and trained using target component data under normal conditions. The fusion autoencoder model includes: a point cloud feature encoder for extracting geometric feature vectors from the point cloud region of the target component; a parametric feature encoder for extracting parametric feature vectors from the geometric parameters; a fusion module for mapping the geometric feature vectors and the parametric feature vectors to a latent space to obtain a joint feature representation; and a decoder for reconstructing the point cloud and geometric parameters based on the joint feature representation. The newly scanned target component data is input into the trained fusion autoencoder model to obtain the reconstruction result and calculate the reconstruction error. Based on the degree to which the reconstruction error deviates from the normal state reconstruction error distribution, it is determined whether the target component has undergone deformation.

2. The intelligent monitoring method for temporal deformation of building components according to claim 1, characterized in that, The point cloud feature encoder is a PointNet encoder, which has a three-layer shared MLP structure and outputs geometric feature vectors through global max pooling; the parametric feature encoder has two fully connected layers and outputs parametric feature vectors; the fusion module concatenates the geometric feature vectors and parametric feature vectors and maps them to the latent space through the MLP to obtain a joint feature representation; the decoder includes a point cloud decoding branch and a CSV decoding branch, which reconstruct the point cloud and geometric parameters from the joint feature representation, respectively.

3. The intelligent monitoring method for time-series deformation of building components according to claim 1, characterized in that, The fusion autoencoder model is trained using target component data under normal conditions, employing a joint loss function: ,in For point cloud reconstruction loss, The mean square error of the geometric parameter reconstruction loss. This is the balance coefficient.

4. The intelligent monitoring method for time-series deformation of building components according to claim 1, characterized in that, The determination of whether the target component has deformed based on the degree to which the reconstruction error deviates from the normal reconstruction error distribution includes: Calculate the reconstruction error from the training data under normal conditions and calculate its mean. and standard deviation Set deformation threshold , This is the sensitivity coefficient; When the reconstruction error corresponds to the new scanned data When the time is right, it is determined that the component has deformed.

5. The intelligent monitoring method for temporal deformation of building components according to claim 1 or 4, characterized in that, When determining that the target component has undergone deformation, the relative rate of change of each geometric parameter is calculated. ， In the formula, The parameter values ​​corresponding to the new scan data. These are the parameter values ​​corresponding to the same component under normal historical conditions; The deformation level is output by combining the degree of deviation of the reconstruction error with the relative change rate of each geometric parameter.

6. An intelligent monitoring device for the temporal deformation of building components, characterized in that, include: The data acquisition module is used to acquire multiple batches of 3D point cloud data of buildings at different time stages; The coordinate remapping module is used to remap the coordinates of point cloud data from different batches to establish a unified building coordinate system, thus aligning the spatial dimensions of the point cloud data from different batches. This includes: performing RANSAC plane fitting on the top floor plane of the building, extracting the top floor plane, and calculating its normal vector. Using the normal vector as the new coordinate system The axial direction is determined by the right-hand rule. shaft and A unified architectural coordinate system is established along the axial direction; according to Coordinate transformation was performed on each batch of point cloud data, among which... The original point cloud coordinates, The center point of the top floor plan, Let be a rotation matrix, such that axis and the normal vector Alignment; The component recognition module is used to extract building wall point clouds from point cloud data, identify the point cloud regions of target components from the wall point clouds, and extract the geometric parameters of the target components. This includes: performing multi-plane RANSAC segmentation on the coordinate-remapped building point clouds to extract each wall plane, and separating the wall point clouds into primary and non-primary plane parts based on the plane normal vector direction; performing DBSCAN density clustering on the non-primary plane parts to obtain candidate component regions; projecting the 3D point cloud of the candidate component region onto the corresponding wall plane, calculating the 2D convex hull of the projected point cloud, finding the largest inscribed rectangle within the convex hull, and extracting the 3D dimensions of the axis-aligned bounding box of the candidate component region as its length. ,Width ,thick Threshold filters are set for various parameters to identify the point cloud region of the target component; geometric parameters are extracted into seven-dimensional parameters, including: length ,Width ,thick , points ,density Tightness Area ratio ,in This represents the area of ​​the largest inscribed rectangle. The area of ​​the two-dimensional convex hull; The fusion autoencoder model includes: a point cloud feature encoder for extracting geometric feature vectors from the point cloud region of the target component; a parametric feature encoder for extracting parametric feature vectors from the geometric parameters; a fusion module for mapping the geometric feature vectors and the parametric feature vectors to a latent space to obtain a joint feature representation; and a decoder for reconstructing the point cloud and geometric parameters based on the joint feature representation. The deformation detection module is used to input the newly scanned target component data into the trained fusion autoencoder model, obtain the reconstruction result and calculate the reconstruction error, and determine whether the target component has deformed based on the degree to which the reconstruction error deviates from the normal state reconstruction error distribution.

7. The intelligent monitoring device for time-series deformation of building components according to claim 6, characterized in that, It also includes a visualization output module, which generates 3D visualization models and inspection reports that use different colors to distinguish deformed components from normal components.

Images