Building component deformation analysis method and device based on three-dimensional point cloud data, equipment and medium

Abstract

The application provides a building component deformation analysis method and device based on three-dimensional point cloud data, equipment and medium, wherein the method comprises the following steps: collecting component point clouds, images and calibration data by using a laser scanning and RGB imaging device in multiple time sequences, obtaining pretreated point clouds by SAM segmentation, constructing fusion features, obtaining an initial deformation field by semantic coarse registration and gradient weighted ICP, eliminating abnormal points to output a corrected deformation field, generating a predicted deformation field and giving an early warning by means of anchor point coupling time sequence prediction, superimposing layers to output visualized data; the application deduces multi-period deformation data by means of anchor point coupling time sequence prediction, sets layered early warning to realize trend prediction, converts rigid body coordinates to unify time sequence data, carries out partition operation to eliminate data crosstalk, superimposes deformation, early warning layers on a three-dimensional model by layered layer rendering, supports interactive labeling and multi-format export, intuitively displays real-time deformation and predicts trends, and provides traceable quantitative visualized support for building structure safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of building structural health monitoring technology, specifically to a method, apparatus, equipment, and medium for analyzing the deformation of building components based on three-dimensional point cloud data. Background Technology

[0002] Structural health monitoring of building engineering is a core technical means to ensure the safety of buildings throughout their entire life cycle. Various building components such as concrete beams, columns, floor slabs, steel structures, and curtain wall panels are subjected to external forces such as loads, temperature, and settlement over a long period of time, resulting in continuous and slow deformation. The accumulation of small deformations over a long period of time can lead to major safety hazards such as structural cracking, joint failure, and curtain wall detachment. Therefore, it is necessary to continuously and accurately acquire spatial deformation data of components. Three-dimensional laser scanning technology can quickly collect three-dimensional point clouds of the entire component, which can completely restore the overall spatial shape of the component. RGB imaging equipment can simultaneously record the visual texture of the component surface. The combination of the two can simultaneously acquire two types of basic monitoring data: geometric and image data. At present, the industry generally uses the method of repeatedly collecting data in multiple time series to track component deformation. It relies on point cloud registration algorithms to compare the data differences of different collection periods to quantify the amount of component deformation. As the requirements for building monitoring become increasingly refined, the method of simply relying on single geometric feature registration is difficult to adapt to complex construction site environments. The industry urgently needs an integrated deformation analysis solution that integrates visual semantics, geometric features, and time series prediction to support the long-term automated monitoring of various types of building components.

[0003] Existing building component deformation monitoring and processing solutions have several inherent defects. Traditional acquisition methods lack standardized synchronous acquisition specifications, and images and point cloud data are acquired separately, making it impossible to establish a stable correspondence between pixels and 3D spatial coordinates. Construction sites generate a large number of invalid point clouds due to attachments such as scaffolding, dust, and temporary signs. Manually removing these attachments is time-consuming and prone to omissions. Conventional registration algorithms rely solely on point cloud geometric features for matching, failing to combine image semantic information to distinguish between different components and debris. Cross-temporal registration is prone to cross-regional mismatches, with significant errors in point cloud matching at edge areas. The generated deformation field contains numerous abnormal deviation points and lacks an automated correction mechanism. Traditional monitoring methods can only calculate the deformation value for the current period and cannot predict subsequent deformation trends based on historical time-series data. Unified deformation judgment standards cannot adapt to the differentiated deformation characteristics of steel structures, concrete components, and curtain wall panels. Risk warnings are limited in form, and data output only provides simple numerical tables, lacking 3D visualization and interactive display capabilities. The monitoring results are not intuitive enough, making it difficult to quickly locate high-risk component areas. Summary of the Invention

[0004] This invention addresses the aforementioned shortcomings of existing technologies by providing a method, apparatus, equipment, and medium for analyzing the deformation of building components based on 3D point cloud data. It simultaneously acquires multi-temporal component data through laser scanning and RGB imaging, establishes a pixel-to-3D point mapping relationship based on joint calibration parameters, distinguishes components from attachments using semantic segmentation, and completes point cloud preprocessing. A semantically constrained coarse registration algorithm is constructed by fusing visual and geometric features, and a high-precision deformation field is solved and corrected using edge gradient weighted registration. Anchor points are used to align temporal data, and a coupled temporal prediction algorithm is used to predict subsequent component deformation and provide tiered early warnings. Finally, multiple deformation and warning layers are superimposed on a textured 3D model to achieve multi-dimensional visualization output. The entire process is automated, adaptable to various types of building components, and balances accurate deformation calculation, trend prediction, and intuitive display. It effectively solves the problems of low accuracy, difficulty in removing interference, lack of predictive ability, and weak visualization in traditional monitoring methods, and is suitable for structural health monitoring during building construction and operation and maintenance phases.

[0005] According to a first aspect of the present invention, a method for deformation analysis of building components based on three-dimensional point cloud data is provided, the specific steps of which are as follows: S1. Using a device that integrates laser scanning and RGB imaging, simultaneously acquire three-dimensional point cloud and RGB images of the same component in no less than three time-series cycles, pre-store camera and laser joint calibration parameters, and store point cloud, image, timestamp, calibration parameter copy and acquisition position record in each cycle. S2. Read the calibration parameters to establish a pixel-to-3D point mapping index; use SAM segmentation to obtain component masks and attachment masks, identify anchor point pixels and 3D coordinates; use the index to remove attachment point clouds, downsample and normalize the component point cloud to obtain a preprocessed point cloud; S3. Based on the preprocessed point cloud and image, extract SIFT key points and FPFH features to construct fused features; under SAM semantic constraints, use the semantic fusion coarse registration loss algorithm to solve the rotation and translation matrix in the coarse registration stage; use Canny edge gradient to weight the point cloud, take the coarse registration matrix as the initial value, run the gradient weighted ICP convergence algorithm to calculate the deformation displacement field, and obtain the initial deformation field; traverse the initial deformation field, mark the abnormal points according to the neighborhood deviation and backtrack to verify, locally re-register and interpolate the abnormal area to fuse, and output the corrected deformation field; S4. Extract anchor point coordinates and align them according to the first time sequence; extract point deformation sequences by partition, use the anchor point coupled time sequence prediction algorithm to calculate the predicted displacement of the next period to generate the predicted deformation field, and mark the over-limit warning. S5. Generate chromatographic and warning layers based on the deformation field and early warning, overlay them onto the 3D model, attach semantic information, deformation values ​​and time series curves, and output visualized data.

[0006] Preferably, in step S1, the types of components include concrete beams, concrete columns, floor slabs, steel structures, and curtain wall panels; when collecting data for different components, the equipment's field of view completely covers the single-panel segment of the curtain wall, the intersection area of ​​the flange web of the steel structure, or the beam-column node area; the time interval is fixed at 7 days, 30 days, or 90 days; before each collection, the equipment's installation height, horizontal coordinates, pitch angle, and horizontal rotation angle are fixed; each period's data packet independently stores point cloud, RGB image, timestamp, calibration parameter copy, and installation position record, and stores them in partitions according to time sequence number; point cloud cropping, image filtering, and coordinate conversion calculations are prohibited throughout the collection process.

[0007] Preferably, in step S2, the pixel-to-3D point mapping index is generated through perspective coordinate conversion, and each index record contains the pixel's horizontal and vertical coordinates, the matching 3D X / Y / Z coordinates, and the time sequence number; SAM segmentation uses a fixed 0.85 confidence threshold, and the output component mask corresponds to the imaging area of ​​each component, while the attachment mask corresponds to the scaffolding, dust, temporary markers, and sticker areas; the anchor point types include beam-column nodes, embedded steel plates, curtain wall grid reference angles, and structural embedded parts, and each anchor point record corresponds to the image file name, 2D pixel coordinates, and 3D coordinates; point cloud downsampling uses a fixed-side-length cubic grid for simplification, and coordinate normalization converts all point clouds in a single time series to the same local spatial coordinate system.

[0008] Preferably, in the S3 step neighborhood deviation determination process, the neighborhood search radius is set to an arbitrary fixed value within the range of 3mm to 8mm, and the continuity tolerance threshold is set to an arbitrary fixed value within the range of 0.2mm to 0.5mm. When performing backtracking verification for marked outliers, the SAM mask confidence value corresponding to the matched pixel of the outlier, the Canny edge gradient range of the corresponding pixel, and the mapping relationship between the pixel and the 3D point are retrieved sequentially. The radius of the local point cloud intercepted in the local reregistration operation is set to an arbitrary fixed value within the range of 10mm to 20mm. The interpolation operation selects either radial basis function interpolation or moving least squares method. Among them, the radial basis function interpolation uses the Gaussian kernel function, and its mathematical expression is: In the formula The three-dimensional Euclidean distance between the point to be interpolated and the locally known correction point is given. The shape parameter is fixed at 0.8 (pre-calibrated based on the ratio of the average spacing of the local point cloud to the truncation radius of 15 mm). The interpolation support radius is consistent with the local reregistration truncation radius (i.e., a fixed value of 15 mm within the range of 10 mm to 20 mm). The local correction deformation vector is diffused and fused into the global deformation field by solving the radial baseline linear equations. The moving least squares method uses linear basis functions, and its basis function vector is... (Applicable to points in three-dimensional space), the weighting function is a cubic spline weighting function, whose mathematical expression is: (when hour), (when (time), in the formula To fit the Euclidean distance between the point and its neighbors, The weight function supports the radius, which is fixed at 15mm. Based on the above basis functions and weight functions, a weighted least squares fitting polynomial is established in the local support domain to calculate the correction deformation vector at each point in the global domain. The local reregistration fully executes the entire operation process of SIFT keypoint and FPFH feature fusion construction, semantic fusion coarse registration loss algorithm iteration, and gradient weighted ICP convergence algorithm iteration.

[0009] Preferably, the timing alignment in step S4 uses rigid body coordinate transformation to process all correction deformation fields from the second to the nth timing sequence; in the rigid body coordinate transformation matching stage, only anchor points of the same type of SAM segmentation markers are matched; the component partition boundaries are divided according to the component mask pixel boundaries output by SAM segmentation; the timing point deformation sequences corresponding to different component partitions are stored, read, and calculated independently; the deformation threshold range for steel structure partitions is set to 0.6mm to 0.9mm, the deformation threshold range for concrete beams and concrete columns is set to 1.0mm to 1.5mm, and the deformation threshold range for curtain wall panel partitions is set to 1.0mm to 1.3mm.

[0010] Preferably, the chromatographic layers generated in step S5 are divided into two categories: the real deformation chromatographic layer and the predicted deformation chromatographic layer. The two types of chromatographic layers each use two sets of non-overlapping color ranges. The first-level warning points are matched with fixed pure color highlight blocks, and the second-level warning points are matched with dynamically switched color flashing layers. The generation boundaries of the chromatographic layers and warning layers strictly match the component mask pixel boundaries of the SAM segmentation output, and the rendering range of the layers does not exceed the corresponding component partition boundaries.

[0011] Preferably, in step S5, the surface of the 3D model is completely fitted with the texture map of the original RGB image collected in the current period; static text annotations are synchronously attached to the 3D model screen, and the text annotations include component partition classification text, the actual deformation vector magnitude value of the mouse selected point, the predicted deformation vector magnitude value of the mouse selected point, the total time series deformation value curve of the mouse selected point, and the warning level text corresponding to the mouse selected point; the visualization data supports retrieving the screen of a single component partition layer, the screen of a single time series period layer, and the screen of a single warning level layer separately; after the calculation is completed, the rendered screenshot in image format, the deformation value file in table format, and the original coordinate file of the time series curve in text format are output synchronously.

[0012] According to another aspect of the present invention, a deformation analysis device for building components based on three-dimensional point cloud data is provided, the device comprising: The acquisition module integrates a laser scanning unit and an RGB camera unit, pre-stores camera-laser joint calibration parameters, and acquires 3D point cloud, RGB image, timestamp, and installation position data of the component according to a fixed time sequence period and a fixed equipment posture, and stores the data packets in time sequence. The preprocessing segmentation module is connected to the acquisition module by signal. It reads the calibration parameters to construct a pixel-3D point mapping index, uses the SAM segmentation model to generate component masks and attachment masks, identifies anchor point coordinates and removes attachment point clouds, completes point cloud downsampling and coordinate normalization, and outputs preprocessed point clouds. The registration deformation solution module is connected to the preprocessing segmentation module. It extracts SIFT and FPFH fusion features, runs the semantic fusion coarse registration loss algorithm to solve the coarse registration matrix, constructs the gradient weighted ICP algorithm based on the Canny gradient to calculate the initial deformation field, marks and corrects outliers, and outputs the corrected deformation field. The temporal prediction and early warning module is connected to the registration deformation solution module. It completes rigid body alignment across time-series anchor points, constructs temporal deformation sequences by component and partition, generates predicted deformation fields through anchor point coupled temporal prediction algorithms, completes first and second-level early warning markings based on partition deformation thresholds, and stores all temporal calculation data. The 3D visualization output module connects with the time-series prediction and early warning module to generate two types of color chromatogram layers (real / predictive) and two levels of early warning layers. These layers are then overlaid onto a 3D component model with RGB texture maps, and interactive text annotations are attached. The module supports multi-dimensional image retrieval and output of visualization files in image, table, and text formats. According to a third aspect of the present invention, a computer terminal is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, can be used to perform the method described in any one of the above inventions.

[0013] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can be used to perform the method described in any one of the above inventions.

[0014] By adopting the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art: I. This invention achieves automatic differentiation between building component targets and external attachments by simultaneously acquiring data through laser scanning and RGB imaging, combined with the SAM semantic segmentation mechanism. It utilizes a pixel-based 3D point mapping index to perform batch filtering of stray point clouds and simultaneously performs lightweight point cloud regularization processing, reducing the interference of irrelevant information on deformation calculations from the data source. It integrates image SIFT visual features and point cloud FPFH geometric features to construct a semantically constrained registration loss function, and uses image edge gradients to weight the point cloud registration process. Combined with deformation field anomaly point backtracking verification and local re-registration interpolation correction processes, it significantly improves the alignment accuracy of multi-cycle point cloud registration, weakens the calculation bias caused by uneven surface textures and missing point clouds in the components, and unifies the entire data processing chain by using a semantic mask to define the computational boundaries. It automates data cleaning, feature matching, and deformation solving throughout the entire process, eliminating the need for manual component delineation and effective point cloud selection, reducing manual intervention costs, and ensuring stable and consistent deformation calculation results for different types of concrete, steel structures, and curtain wall components.

[0015] II. This invention utilizes an anchor-coupled time-series prediction model combined with zoned differentiated deformation judgment rules to perform linked extrapolation of historical multi-cycle deformation data. It distinguishes different deformation degrees and sets layered early warning indicators to achieve component deformation trend prediction and risk classification prompts. Rigid body coordinate transformation is used to achieve unified alignment of full-time-series data. The deformation sequences of each component are independently calculated and stored to avoid crosstalk between deformation data of different structural components. It adapts to the differentiated deformation control needs of various building components, builds a layered layer rendering system, generates independent color chromatographic layers of actual deformation and predicted deformation, and overlays early warning layers with different display effects onto a 3D model with realistic textures. It is equipped with an interactive time-series numerical annotation function, supports the separate retrieval and export of multi-dimensional data in multiple format result files, and can not only intuitively present the current deformation state of the component, but also predict the subsequent deformation development trend in advance, providing quantitative, visual, and traceable complete data support for long-term safety monitoring of building structures. Attached Figure Description

[0016] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart of a method for analyzing the deformation of building components based on 3D point cloud data; Figure 2 This is a framework diagram of a building component deformation analysis device based on 3D point cloud data. Detailed Implementation

[0017] The embodiments of the present invention are described in detail below: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Example

[0018] Reference Figure 1 One embodiment of the present invention proposes a method for analyzing the deformation of building components based on three-dimensional point cloud data. It adopts a data processing mechanism of multi-cycle laser-RGB synchronous acquisition, SAM semantic segmentation constraint fusion registration, anchor point coupling temporal prediction and layered visualization rendering. It can accurately calculate the deformation displacement of concrete beams and columns, steel structures and curtain wall panels without relying on a large number of manually embedded deformation sensors, and provide graded early warning and three-dimensional visualization of points that exceed the deformation threshold.

[0019] The method described in this embodiment specifically includes: S1. Using a device that integrates laser scanning and RGB imaging, simultaneously acquire 3D point cloud and RGB images of the same component in no less than three time-series cycles. Pre-store the joint calibration parameters of the camera and laser. Store the point cloud, image, timestamp, a copy of the calibration parameters, and the acquisition position record for each cycle. Before acquisition, fix the device's setup height, horizontal coordinates, pitch angle, and horizontal rotation angle. The time-series cycle interval is fixed at 7 days, 30 days, or 90 days. Point cloud cropping, image filtering, and coordinate conversion calculations are prohibited throughout the acquisition process. Store independent data packets for each cycle in partitions according to the time sequence number. S2. Read the pre-stored camera-laser joint calibration parameters and generate a pixel-to-3D point mapping index through perspective coordinate conversion; use the SAM segmentation model with a 0.85 confidence threshold to output component masks and attachment masks, identify the pixels and 3D coordinates of beam-column nodes, embedded steel plates, and curtain wall grid reference corner anchor points; rely on the mapping index to remove the attachment point clouds corresponding to scaffolding, dust, and temporary labels, use a fixed-side-length cubic mesh to downsample the point cloud, and convert all point clouds in a single time series to the same local spatial coordinate system to complete normalization, and output the preprocessed point cloud; S3. Based on preprocessed point clouds and synchronously acquired RGB images, extract key points from SIFT images and construct fusion features with FPFH 3D geometric features; under the constraint of SAM semantic mask, substitute semantic fusion coarse registration loss algorithm to solve the coarse registration rotation and translation matrix; rely on the Canny operator to extract image edge gradients and assign gradient weights to the point cloud, use the coarse registration matrix as the initial value of iteration to run gradient weighted ICP convergence algorithm to calculate the initial deformation displacement field; set a neighborhood search radius of 3mm-8mm and a continuity tolerance threshold of 0.2mm-0.5mm to mark abnormal points in the initial deformation field, retrieve the SAM mask confidence, Canny gradient interval, and pixel-3D point mapping relationship of the matched pixels of the abnormal points to complete backtracking verification; cut out a local point cloud with a radius of 10mm-20mm, and perform the complete set of operations of feature fusion, coarse registration, and gradient weighted ICP to achieve local reregistration; use radial basis function interpolation or moving least squares method to complete global interpolation fusion, and output the corrected deformation field; S4. Extract all time-series anchor point coordinates, and perform rigid body coordinate transformation only on anchor points of the same type marked by SAM to complete cross-time-series coordinate alignment; divide the component into partitions according to the component mask pixel boundaries output by SAM, and store, read and calculate the deformation sequence of time-series points in each partition independently; substitute historical multi-cycle deformation data into the anchor point coupling time-series prediction algorithm to calculate the predicted displacement of the next cycle and generate the predicted deformation field; use a deformation threshold of 0.6mm-0.9mm for steel structure partitions, 1.0mm-1.5mm for concrete beam and column partitions, and 1.0mm-1.3mm for curtain wall panel partitions, and complete two-level early warning marking based on the comparison between the predicted deformation vector magnitude and the partition threshold; after each cycle calculation, store the full file of the corrected deformation field, anchor point coordinate table, partition deformation sequence, global predicted deformation field, and early warning point coordinates, and transmit them to the visualization stage in a unified manner; S5. Generate real deformation chromatogram layers and predicted deformation chromatogram layers respectively. The two types of chromatograms use non-overlapping color ranges. First-level warning points are matched with solid color highlight blocks, and second-level warning points are matched with dynamic color flashing layers. The rendering boundaries of all layers strictly fit the SAM component mask. The layers are superimposed on the 3D component model that fully fits the original RGB texture map. Component classification text, real / predicted deformation vector magnitude of selected points, full-time deformation curves, and static text labels of warning levels are attached. Visualized data supports individual retrieval of single component, single time series, and single warning level screens, and synchronously outputs rendered screenshots, table deformation value files, and original coordinate text files of time series curves.

[0020] Specifically, the execution logic of this invention lies in establishing a method for analyzing the deformation of building components based on 3D point cloud data. This method first uses a laser-RGB integrated device to complete the non-destructive acquisition of raw data over multiple cycles, fully preserving the original point cloud, images, and calibration parameters without preprocessing. Then, relying on SAM semantic segmentation, it achieves pixel-level differentiation of components and attachments, eliminates interfering point clouds, identifies stable anchor points, and completes point cloud standardization preprocessing. Subsequently, it fuses image SIFT and 3D FPFH features, using a semantic fusion loss function to constrain coarse registration accuracy, and combines edge gradient weighted ICP to solve the global deformation of the component. It also performs local re-registration correction on abnormal points to eliminate deformation errors caused by registration deviations. Next, it uses global stable anchor points to unify multi-temporal coordinates, establishes independent deformation time series for each component and region, and predicts the deformation of the next cycle based on anchor point spatial distance and temporal decay weights. It divides the data into two levels of warning according to the deformation tolerance threshold of different component materials. Finally, it constructs two sets of independent color gamuts and two-level warning layers, integrates them with a 3D texture model to achieve interactive visualization, and outputs multi-format engineering data files. By employing semantic segmentation, geometric-image fusion registration, and multi-layer constraint linkage of anchor point temporal prediction, automated high-precision calculation and risk warning of minute deformations of building components can be achieved without human intervention.

[0021] Optionally, the execution logic for component classification, equipment setup, and data packet storage in step S1 includes: The components to be collected are categorized as concrete beams, concrete columns, floor slabs, steel structure flanges and webs, and curtain wall panels. During data collection, the equipment position should be adjusted to ensure complete coverage of individual curtain wall panels, intersecting areas of steel structure flanges and webs, or beam-column joints. The time interval can be fixed at 7 days, 30 days, or 90 days. In engineering monitoring, a 7-day interval is used for short-term construction deformation monitoring, a 30-day interval for medium- and long-term settlement monitoring of the main structure, and a 90-day interval for annual curtain wall deformation inspection. Before each data collection operation, a tripod and level should be used to fix the equipment's height and horizontal XY coordinates. The camera position remains unchanged throughout the entire acquisition process, including the target, lens tilt angle, and horizontal rotation angle, until the end of each acquisition cycle. Each time-series cycle generates an independent data packet, which stores the original uncropped 3D point cloud, the original unfiltered RGB image, the acquisition timestamp, a copy of the camera-laser joint calibration parameters, and equipment setup coordinates and attitude records. All data packets are stored hierarchically in folders organized by component and partition according to their time sequence number. Throughout the entire acquisition process, the onboard acquisition program does not perform point cloud cropping, image Gaussian / median filtering, or global coordinate conversion calculations, thus preserving the original acquisition data for unified backend processing.

[0022] Optionally, the execution logic for pixel-to-3D point mapping index generation, SAM segmentation output file, and anchor point data storage in step S2 includes: The pre-stored camera-laser joint calibration intrinsic and extrinsic parameter matrix in S1 is read. A one-to-one mapping index between pixels and 3D points is established through camera perspective projection coordinate conversion. Each index record includes the pixel's horizontal coordinate u, vertical coordinate v, matching 3D space X / Y / Z coordinates, and corresponding time sequence number. The SAM segmentation model inference confidence threshold is fixed at 0.85. Inference outputs two types of uncompressed grayscale image mask files. Pixel values ​​of 0 in the mask image represent the background area, and values ​​of 1 represent the target area. The component mask corresponds to the imaging pixels of all concrete, steel, and curtain wall components. The object mask corresponds to the pixels of scaffolding, wall dust, temporary stickers, and construction protective boards; the anchor point selection is limited to beam-column cross nodes, four corners of embedded steel plates, curtain wall grid reference angles, and structural embedded bolt positions. For each anchor point, the RGB image file name, two-dimensional UV pixel coordinates, and three-dimensional X / Y / Z spatial coordinates are extracted and matched; point cloud downsampling is performed using a cube grid with uniform side length. The grid side length is 5mm in engineering. All component point clouds in a single time series are uniformly translated and rotated to a local spatial coordinate system with the component partition center as the origin to complete coordinate normalization.

[0023] The pixel-to-3D point mapping index is stored as a structured CSV table text file, with the table columns being time sequence number, pixel u, pixel v, 3D X, 3D Y, and 3D Z; the component masks and attachment masks output by SAM are stored as lossless PNG grayscale images; the anchor point dataset is generated as a separate TXT table file, with each anchor point record having fields including a unique numeric ID, the text of the component partition to which it belongs, the filename of the matching RGB image, the 2D horizontal pixel coordinates, the 2D vertical pixel coordinates, the 3D X coordinate, the 3D Y coordinate, and the 3D Z coordinate.

[0024] Optionally, the execution logic of the semantic fusion coarse registration loss algorithm in step S3 includes: The mathematical expression for the semantic fusion coarse registration loss function is:

[0025] in The normalization calculation formula is: , The normalization calculation formula is: In the formula Represents the total loss from integration. Represents fixed weighting coefficients for three-dimensional geometric features, with fixed values. , The fixed weight coefficients represent the grayscale features of the image, with fixed values. , Representing three-dimensional feature points , The FPFH geometric distance is normalized to the maximum and minimum values. The value after the interval, Representative feature points and Original FPFH geometric distance, , Representing three-dimensional feature points, , These represent the minimum and maximum FPFH geometric distances within the current matched feature point group, respectively. The pixel grayscale difference representing SIFT keypoint matching is normalized to the maximum and minimum values. The value after the interval, Represents SIFT keypoints in the image. This represents the original grayscale difference between two sets of SIFT keypoints. , These represent the minimum and maximum SIFT grayscale differences within the current matched feature point group, respectively. Represents the semantic masking coefficient, which is used when two points are within the same component mask pixel area. When two points belong to the pixel regions of the component mask and the attached mask respectively, or when two points belong to any two different component mask pixel regions. ; and The normalization operation is performed in real time before each loss iteration calculation.

[0026] The algorithm execution flow is as follows: extract all SIFT+FPFH fused feature points from the two temporal point clouds to be registered; calculate the FPFH geometric distance and SIFT grayscale difference for each feature point pair and normalize them in real time; retrieve the SAM mask labels of the corresponding pixel positions of the two points to determine the semantic masking coefficient; and substitute the fixed weight coefficients. , Calculate the feature fusion loss of a single set, iteratively minimize the global total loss, and output the optimal coarse registration rotation transformation matrix R and translation matrix T.

[0027] Optionally, the execution logic of the gradient-weighted ICP convergence algorithm in step S3 includes: the mathematical expression of the gradient-weighted ICP objective function is:

[0028] The gradient normalization calculation formula is: The formula for calculating gradient weights is: In the formula Represents the ICP objective function. Represents the measured coordinates of a three-dimensional point. This represents the three-dimensional point coordinates matching the standard mold cavity shape. , To iteratively optimize the rotation and translation matrices, the initial values ​​for iteration are directly derived from the semantic fusion coarse registration output. , , The pixel Canny edge gradient values ​​are normalized to the maximum and minimum values. The value after the interval, Representing the Each 3D point corresponds to the original Canny edge gradient value of the pixel. Represents the three-dimensional point index number. , These represent the minimum and maximum Canny gradient values ​​of all pixels within the current component partition, respectively. Representing the Gradient weighting coefficients for each three-dimensional point This represents a fixed coefficient for gradient amplification, with a fixed value. , The fixed weight coefficients representing the deformation continuity regularization term are fixed at a certain value. , This represents the total number of valid 3D points in a single component partition. This represents the continuity regularization term of the deformation field, used to constrain the deformation difference between neighboring points to not exceed the continuity tolerance threshold.

[0029] The algorithm execution flow is as follows: Read the coarse registration output. , Starting from the iteration point, all valid 3D points within the partition are traversed, and the corresponding pixel Canny edge gradient is extracted and normalized to obtain... Calculate the gradient weights at each point. Substitute each point into the objective function to calculate the residual, and iterate to update. , Minimize the weighted sum of residuals while introducing a continuity regularization term to constrain deformation smoothness; after iterative convergence, output the three-dimensional deformation displacement vectors at all points in the global domain to generate the initial deformation field.

[0030] Optionally, the execution logic for anomaly marking, backtracking verification, and local reregistration in step S3 includes: The parameters for anomaly point determination are fixed within the following range: neighborhood search radius 3mm-8mm, continuity tolerance threshold 0.2mm-0.5mm. For engineering monitoring of beam and column components, a neighborhood radius of 5mm and a continuity threshold of 0.3mm are selected. Each three-dimensional point in the initial deformation field is traversed, and a neighborhood search radius is defined with that point as the center. The deformation displacement vectors of all points in the neighborhood are extracted, and the difference between the current point and the mean deformation of the neighborhood is calculated. When the absolute value of the difference is greater than the continuity tolerance threshold, it is marked as an anomaly point. For all marked outliers, a backtracking verification is performed: First, the SAM mask inference confidence value corresponding to the RGB pixels of the 3D point is retrieved. If the confidence value is lower than 0.85, it is determined to be a mask segmentation interference point. Second, the normalized interval of the pixel Canny edge gradient is retrieved. Outliers in smooth planar regions with gradients close to 0 are determined to be scanning noise points. Third, the pixel-3D point mapping index matching relationship is retrieved. Points without a complete one-to-one corresponding index are determined to be acquisition misalignment points. After verification, local re-registration is performed on the confirmed real deformation anomaly areas: local point clouds with a radius of 10mm-20mm are extracted centered on the anomaly points, and a 15mm extraction radius is selected for engineering; the entire set of operations, including SIFT+FPFH feature fusion construction, semantic fusion coarse registration loss iteration, and gradient weighted ICP convergence iteration, is repeated for the local point cloud to solve for the local optimal deformation vector; the radial basis function interpolation method is used for global interpolation to diffuse and fuse the local corrected deformation vector into the global deformation field to generate a corrected deformation field without abrupt deviation.

[0031] Optionally, the execution logic of the anchor-coupled timing prediction algorithm in step S4 includes: the mathematical expression for anchor-coupled timing prediction is:

[0032] The formula for calculating the time-decay weight is: In the formula Representing spatial points exist Predicted deformation displacement vector for time series period, Representing the The time-series decay weights corresponding to each historical time series Represents the global timing decay constant, with a value range of... Medium- and long-term structural monitoring uses , represent The corrected deformation displacement vector output by time-series gradient-weighted ICP. This represents the natural exponent operation function. The value range represents the attenuation constant of the component's dedicated space. Steel structure selection Concrete zoning selection Curtain wall partition selection , Representing spatial points With the most recent stable anchor point The three-dimensional Euclidean distance between them This represents the latest historical time sequence number that has been collected. Representing the A historical time sequence number, This represents the number of periods between the current predicted time series and the historical time series. This represents the nearest stable anchor point to spatial point l, where l represents the spatial point index. Represents the total number of historical time series cycles. This represents the historical time-series index. The algorithm execution flow is as follows: after completing the rigid body alignment of all historical time-series anchor points, extract the multi-period deformation sequence of each point by component and region; calculate the three-dimensional distance of each point to the nearest anchor point. Substitute the corresponding component Values ​​and fixed attenuation constants are used to calculate the attenuation weights for each historical time series. The product of all historical time-series deformation vectors and the time-series and spatial distance attenuation exponents is superimposed to output the predicted deformation vector for the next period, generating a global predicted deformation field. The horizontal length of the predicted deformation vector is compared with the corresponding partition deformation threshold. If the magnitude slightly exceeds the threshold, a first-level warning is marked; if the magnitude significantly exceeds the threshold, a second-level warning is marked.

[0033] Optionally, the S4 step of graded early warning determination and time-series data storage execution logic includes: Fixed ranges for deformation thresholds in each component zone: 0.6mm-0.9mm for steel structure zone, with a lower limit of 0.6mm for engineering monitoring and an upper limit of 0.9mm for severe warning; 1.0mm-1.5mm for concrete beams and columns zone, with a lower limit of 1.0mm for warning and an upper limit of 1.5mm for severe warning; and 1.0mm-1.3mm for curtain wall panel zone, with a lower limit of 1.0mm for warning and an upper limit of 1.3mm for severe warning. Two-level warning determination rules: When the predicted deformation vector magnitude is greater than the lower limit threshold of the partition and less than the upper limit threshold of the partition, a level one warning is marked; when the predicted deformation vector magnitude is greater than or equal to the upper limit threshold of the partition, a level two warning is marked. After each time-series cycle of full-process calculation is completed, five types of files are uniformly stored: current period correction deformation field point cloud file, full partition anchor point coordinate summary table, independent time-series point deformation sequence table for each component partition, global prediction deformation field point cloud file, and first / secondary warning point three-dimensional coordinate text; all stored files establish a unified data interface and are completely transmitted to the S5 three-dimensional visualization output module as the data source for layer rendering and interactive annotation.

[0034] Optionally, the execution logic for the S5 step—layer rendering, 3D model texturing, interactive annotation, and multi-format file output—includes: The layers are divided into two types of color spectrum layers and two types of warning layers: the actual deformation color spectrum layer and the predicted deformation color spectrum layer are configured with two sets of completely non-overlapping color gradient ranges. The actual deformation uses a blue-green-orange gradient, and the predicted deformation uses a cyan-purple-red gradient. The first-level warning point layer is a fixed pure yellow highlight color block, and the second-level warning point layer is a red-orange dynamically flashing color block. The rendering boundaries of the color spectrum layer and the warning layer are strictly limited to the SAM component mask pixel range and do not spread outward to the background and attached object areas. 3D model texture processing: The original uncompressed RGB image acquired by S1 is completely mapped and pasted onto the surface of the corresponding time-series 3D component model, preserving the original texture, construction marks, and surface markings of the component; Interactive static text annotation content includes: component partition category text labels, actual deformation vector magnitude value of the selected point, predicted deformation vector magnitude value of the selected point, full-time deformation value polyline curve of the selected point, and text corresponding to the first-level / second-level warning level of the selected point. The image retrieval function supports three independent view outputs: loading only the image of a single component partition layer, loading only the image of all layers in a single time series, and loading only the image of the first-level or only the second-level warning point layer. After the calculation is completed, three types of project files are output simultaneously: high-resolution rendered screenshot images, an Excel spreadsheet storing the deformation values ​​of all points, and a plain text file storing the coordinates of the full-time curve.

[0035] Optionally, the method further includes an equivalent acquisition and correction step corresponding to inverter nonlinear distortion compensation (adapted to the power supply conditions of the laser equipment in the field): Obtain the nonlinear distortion parameters of the power supply for the laser-RGB integrated scanning device; The nonlinear distortion parameters are used to compensate the original laser ranging values ​​collected in real time by the device to obtain compensated ranging point cloud data, which is then used for subsequent S2 preprocessing operations.

[0036] Specifically, the laser ranging offset error of the equipment under different power supply voltages is measured in advance through an offline calibration program to generate a ranging compensation amount that matches the power supply voltage; during the real-time acquisition stage, the power supply voltage of the equipment is read, and the corresponding compensation correction amount is retrieved to perform algebraic correction on the original point cloud 3D coordinates to eliminate the systematic deviation of the ranging system caused by the power supply voltage drop and improve the accuracy of the original point cloud coordinates.

[0037] Optionally, the multi-source data synchronous acquisition execution logic in step S1 includes: Hardware synchronization is achieved by using the built-in clock of the LiDAR and the clock of the RGB camera, so as to trigger the laser ranging sampling and camera analog-to-digital conversion synchronously with a unified master clock signal; the sampling frequency of the laser point cloud 3D coordinates and the acquisition frequency of the RGB image frame are set to a synchronous 10Hz, which is much higher than the natural deformation rate of the building components, avoiding the registration deviation caused by timing misalignment; the device's built-in storage unit partitions and caches the data packets for each cycle to avoid data loss during the acquisition process. Example

[0038] Reference Figure 2 This invention provides a deformation analysis device for building components based on three-dimensional point cloud data. The device is used to execute the deformation analysis method for building components based on three-dimensional point cloud data described above. The device includes: The acquisition module integrates a laser scanning unit and an RGB camera unit. It pre-stores camera-laser joint calibration parameters and acquires 3D point clouds, RGB images, millisecond-level timestamps, and setup position and attitude data of the components according to a fixed 7 / 30 / 90-day time series cycle and a fixed equipment setup posture. The complete data package is stored in partitions with time series numbers. The acquisition process does not perform point cloud cropping, image filtering, or coordinate conversion pre-calculation. The preprocessing segmentation module is connected to the acquisition module. It reads the stored camera-laser joint calibration parameters and constructs a pixel-3D point mapping index through perspective coordinate conversion. It uses a SAM segmentation model with a confidence threshold of 0.85 to generate grayscale masks for components and attachments. It identifies beam-column nodes and embedded steel plate-type stable anchor points and extracts 2D pixels and 3D coordinates. Based on the mapping index, it removes the interference point cloud corresponding to the attachments, uses a fixed-side-length cubic mesh to complete the point cloud downsampling, and uniformly transforms all point clouds in a single time series to a local coordinate system to achieve coordinate normalization. It then outputs the standardized preprocessed point cloud to the next module. The registration deformation solving module is connected to the preprocessing segmentation module via a data link. It extracts key points from the SIFT image and 3D geometric features from the FPFH image to construct a fusion feature set. It then substitutes these features into a semantic fusion coarse registration loss algorithm to iteratively solve the coarse registration rotation and translation matrix. Based on the Canny operator, it extracts the image edge gradients and assigns gradient weights to the point cloud. Using the coarse registration matrix as the initial value for iteration, it runs a gradient-weighted ICP convergence algorithm to obtain the initial global deformation displacement field. It marks abnormal deformation points according to a preset neighborhood radius and continuity threshold, and retrieves mask confidence, gradient interval, and pixel mapping relationship to complete backtracking verification. It then extracts a local point cloud with a fixed radius and completely re-executes feature fusion, coarse registration, and gradient-weighted ICP operations. It uses radial basis function or moving least squares method for global interpolation to fuse the local deformation correction and outputs a corrected deformation field that eliminates abrupt deviations. The temporal prediction and early warning module, connected to the registration deformation solving module, extracts the coordinates of anchor points throughout the entire time series. It only matches anchor points of the same type marked by SAM and performs rigid body coordinate transformation to complete cross-temporal coordinate alignment. It divides the computation into independent computation zones according to the SAM component mask boundary, and the temporal deformation sequence of each zone is stored and calculated independently. It substitutes the anchor point coupled temporal prediction algorithm with multi-cycle historical deformation, spatial distance between the point and the anchor point, and temporal attenuation coefficient to calculate the predicted displacement of the next cycle and generate a global predicted deformation field. According to the deformation threshold intervals corresponding to the steel structure, concrete beams and columns, and curtain wall panels, it divides the first-level and second-level early warning points by comparing the predicted deformation vector magnitude. After each cycle of computation, it stores the full file of the corrected deformation field, anchor point coordinates, zone deformation sequence, predicted deformation field, and early warning point coordinates, and pushes it to the visualization module through a standardized data interface. The 3D visualization output module establishes a data interaction link with the time-series prediction and early warning module, generating real deformation color spectrum layers and predicted deformation color spectrum layers with non-overlapping color ranges, as well as solid color high-brightness first-level warning layers and dynamic flashing second-level warning layers. The rendering boundaries of all layers strictly fit the pixel range of the SAM component mask. All layers are superimposed onto a 3D component model that fully fits the original RGB texture map, and component classification, deformation values, time-series curves, and interactive text annotations for warning levels are attached. It supports individual retrieval of single component, single time series, and single warning level view rendering. After the calculation is completed, it synchronously outputs three types of visualization project files: high-resolution screenshots, table deformation value files, and original coordinate text of time-series curves.

[0039] It should be noted that the data connections between the various units described above do not necessarily represent direct hardware connections. Indirect data interaction methods such as wired transmission, wireless LAN transmission, and local disk file reading and writing can be applied to the embodiments of this invention as long as they achieve the purpose of this invention. The above descriptions are merely exemplary embodiments of this invention and should not be construed as limiting the scope of this invention.

[0040] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and practicing this disclosure. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein. Example

[0041] An embodiment of a computer terminal.

[0042] This embodiment provides a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can be used to execute the building component deformation analysis method based on three-dimensional point cloud data described in any of the foregoing embodiments.

[0043] The specific operation process is as follows: the memory stores the original data acquisition data of each time series, the SAM segmentation model weight file, the fixed parameter values ​​of each algorithm, and the deformation threshold configuration file; the processor sequentially calls the data acquisition reading program, the pixel-3D mapping index construction program, the SAM semantic segmentation inference program, the feature fusion coarse registration operation program, the gradient weighted ICP iteration program, the outlier correction interpolation program, the anchor point coupling time series prediction program, the graded early warning judgment program, and the 3D layer rendering visualization program, and executes the entire process of S1 to S5 to complete the operation, and outputs the corrected deformation field, the predicted deformation field, the graded early warning points, and the multi-format visualization engineering files. Example

[0044] An embodiment of a computer-readable storage medium.

[0045] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, can be used to perform the building component deformation analysis method based on three-dimensional point cloud data described in any of the foregoing embodiments.

[0046] The storage medium can be any medium capable of storing program code, such as a USB flash drive, solid-state drive, portable hard drive, read-only memory (ROM), random access memory (RAM), or optical disc. The storage medium contains the executable program code that implements all steps of this method, the SAM segmentation inference model, the mathematical operation functions of the three core algorithms, and the configuration parameters of the visualization rendering layer. Any processor with program reading and operation capabilities can read the program code in the medium and automatically complete the entire process of multi-cycle point cloud deformation calculation, time series prediction, hierarchical early warning, and 3D visualization output of building components.

[0047] Any matters not covered in the above embodiments of the present invention are well-known in the art.

[0048] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for analyzing the deformation of building components based on three-dimensional point cloud data, characterized in that, The specific steps of this method are as follows: S1. Using a device that integrates laser scanning and RGB imaging, simultaneously acquire three-dimensional point cloud and RGB images of the same component in no less than three time-series cycles, pre-store camera and laser joint calibration parameters, and store point cloud, image, timestamp, calibration parameter copy and acquisition position record in each cycle. S2. Read the calibration parameters to establish a pixel-to-3D point mapping index; use SAM segmentation to obtain component masks and attachment masks, identify anchor point pixels and 3D coordinates; use the index to remove attachment point clouds, downsample and normalize the component point cloud to obtain a preprocessed point cloud; S3. Based on the preprocessed point cloud and image, extract SIFT key points and FPFH features to construct fused features; under SAM semantic constraints, use the semantic fusion coarse registration loss algorithm to solve the rotation and translation matrix in the coarse registration stage; use Canny edge gradient to weight the point cloud, take the coarse registration matrix as the initial value, run the gradient weighted ICP convergence algorithm to calculate the deformation displacement field, and obtain the initial deformation field; traverse the initial deformation field, mark the abnormal points according to the neighborhood deviation and backtrack to verify, re-register and interpolate the abnormal area locally, and output the corrected deformation field; S4. Extract anchor point coordinates and align them according to the first time sequence; extract point deformation sequences by partition, use the anchor point coupled time sequence prediction algorithm to calculate the predicted displacement of the next period to generate the predicted deformation field, and mark the over-limit warning. S5. Generate chromatographic and warning layers based on the deformation field and early warning, overlay them onto the 3D model, attach semantic information, deformation values ​​and time series curves, and output visualized data.

2. The method for analyzing the deformation of building components based on three-dimensional point cloud data according to claim 1, characterized in that, In step S1, the types of components include concrete beams, concrete columns, floor slabs, steel structures, and curtain wall panels. When collecting data for different components, the equipment's field of view completely covers the individual curtain wall panels, the intersection area of ​​the flange web of the steel structure, or the beam-column joint area. The time interval is fixed at 7 days, 30 days, or 90 days. Before each collection, the equipment's installation height, horizontal coordinates, pitch angle, and horizontal rotation angle are fixed. Each period's data packet independently stores point clouds, RGB images, timestamps, calibration parameter copies, and installation position records, and stores them in partitions according to time sequence numbers. Point cloud cropping, image filtering, and coordinate conversion calculations are prohibited throughout the collection process.

3. The method for analyzing the deformation of building components based on three-dimensional point cloud data according to claim 1, characterized in that, In step S2, the pixel-to-3D point mapping index is generated through perspective coordinate conversion. Each index record contains the pixel's horizontal and vertical coordinates, the matching 3D X / Y / Z coordinates, and the time sequence number. SAM segmentation uses a fixed 0.85 confidence threshold. The output component mask corresponds to the imaging area of ​​each component, and the attachment mask corresponds to the scaffolding, dust, temporary markers, and sticker areas. Anchor point types include beam-column nodes, embedded steel plates, curtain wall grid reference angles, and structural embedded parts. Each anchor point record corresponds to the image file name, 2D pixel coordinates, and 3D coordinates. Point cloud downsampling uses a fixed-side-length cubic grid for simplification, and coordinate normalization converts all point clouds in a single time series to the same local spatial coordinate system.

4. The method for analyzing the deformation of building components based on three-dimensional point cloud data according to claim 1, characterized in that, The neighborhood search radius set in the S3 step of the neighborhood deviation determination process is selected as any fixed value within the range of 3mm to 8mm, and the continuity tolerance threshold is selected as any fixed value within the range of 0.2mm to 0.5mm. When performing backtracking verification for the marked outliers, the SAM mask confidence value corresponding to the matched pixel of the outlier, the Canny edge gradient range of the corresponding pixel, and the mapping relationship between the pixel and the 3D point are retrieved in sequence. The radius of the local point cloud truncated in the local reregistration operation is selected as any fixed value within the range of 10mm to 20mm. The interpolation operation selects one of the following methods: radial basis function interpolation and moving least squares method. The local reregistration fully executes the entire operation process of SIFT keypoint and FPFH feature fusion construction, semantic fusion coarse registration loss algorithm iteration, and gradient weighted ICP convergence algorithm iteration.

5. The method for analyzing the deformation of building components based on three-dimensional point cloud data according to claim 1, characterized in that, The timing alignment in step S4 uses rigid body coordinate transformation to process all correction deformation fields from the second to the nth time series; in the rigid body coordinate transformation matching stage, only anchor points of the same type as the SAM segmentation markers are matched. The component partition boundaries are defined according to the component mask pixel boundaries output by SAM segmentation; the deformation sequences of time points corresponding to different component partitions are stored, read, and calculated independently; the deformation threshold range is set from 0.6mm to 0.9mm for steel structure partitions, from 1.0mm to 1.5mm for concrete beam and concrete column partitions, and from 1.0mm to 1.3mm for curtain wall panel partitions.

6. The method for analyzing the deformation of building components based on three-dimensional point cloud data according to claim 1, characterized in that, The chromatographic layers generated in step S5 are divided into two categories: the actual deformation chromatographic layer and the predicted deformation chromatographic layer. The two types of chromatographic layers each use two sets of non-overlapping color ranges. Level 1 warning points are matched with a fixed solid color highlight block layer, while level 2 warning points are matched with a dynamically switching color flashing layer. The generation boundaries of the color spectrum layer and warning layer strictly match the pixel boundaries of the component mask output by SAM segmentation, and the rendering range of the layer does not exceed the boundary of the corresponding component partition.

7. The method for analyzing the deformation of building components based on three-dimensional point cloud data according to claim 1, characterized in that, In step S5, the surface of the 3D model is completely fitted with the texture map of the original RGB image collected in the current period; the 3D model screen is synchronously loaded with static text annotations, which include component partition classification text, the actual deformation vector magnitude value of the mouse selected point, the predicted deformation vector magnitude value of the mouse selected point, the curve of all time series deformation values ​​of the mouse selected point, and the warning level text corresponding to the mouse selected point; the visualization data supports the separate retrieval of the single component partition layer screen, the separate retrieval of the single time series period layer screen, and the separate retrieval of the single warning level layer screen. After the calculation is completed, the system synchronously outputs a rendered screenshot in image format, a deformed numerical file in table format, and the original coordinate file of the time-series curve in text format.

8. A deformation analysis device for building components based on three-dimensional point cloud data, the device being applicable to the deformation analysis method for building components based on three-dimensional point cloud data as described in any one of claims 1-7, characterized in that, The device includes: The acquisition module integrates a laser scanning unit and an RGB camera unit, pre-stores camera-laser joint calibration parameters, and acquires 3D point cloud, RGB image, timestamp, and installation position data of the component according to a fixed time period and a fixed equipment posture, and stores the data packets in time sequence. The preprocessing segmentation module is connected to the acquisition module by signal. It reads the calibration parameters to construct a pixel-3D point mapping index, uses the SAM segmentation model to generate component masks and attachment masks, identifies anchor point coordinates and removes attachment point clouds, completes point cloud downsampling and coordinate normalization, and outputs preprocessed point clouds. The registration deformation solution module is connected to the preprocessing segmentation module. It extracts SIFT and FPFH fusion features, runs the semantic fusion coarse registration loss algorithm to solve the coarse registration matrix, constructs the gradient weighted ICP algorithm based on the Canny gradient to calculate the initial deformation field, marks and corrects outliers, and outputs the corrected deformation field. The temporal prediction and early warning module is connected to the registration deformation solution module. It completes rigid body alignment across time-series anchor points, constructs temporal deformation sequences by component and partition, generates predicted deformation fields through anchor point coupled temporal prediction algorithms, completes first and second-level early warning markings based on partition deformation thresholds, and stores all temporal calculation data. The 3D visualization output module connects with the time-series prediction and early warning module to generate two types of color chromatogram layers (real and predicted) and two levels of early warning and alert layers. The layers are overlaid on 3D component models with RGB texture maps, and interactive text annotations are attached. It supports multi-dimensional screen retrieval and output of visualization files in image, table, and text formats.

9. A computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it can be used to perform the method of any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program can be used to perform the method of any one of claims 1-7.

Images