Prediction method and prediction device for building deformation point cloud time series

Abstract

The present invention provides a method and device for predicting building deformation point cloud time series, relating to the field of deformation monitoring technology. The method comprises: analyzing aligned point cloud sets to obtain point cloud sets for abnormal regions; extracting topological features of preset connection structures in abnormal regions based on the multi-resolution features of the point cloud sets in abnormal regions, and constructing a topological map of the abnormal regions; applying stress constraints and material property constraints to the topological map of the abnormal regions to determine the latent space features of the topological map; interpolating and extrapolating the latent space features based on a latent diffusion model to determine a predicted vector field of the latent space features and suspected damaged regions; performing multi-scale regression analysis on the suspected damaged regions to obtain local parameters; fusing and evaluating the consistency of the three-dimensional deformation prediction field and local parameters corresponding to the predicted vector field of the latent space features to obtain a fused global deformation prediction field, and finally obtaining a high-risk deformation distribution. The present invention can accurately predict deformation.

Description

Technical Field

[0001] The present invention relates to the technical field of building deformation monitoring, and in particular to a method and device for predicting a time series of building deformation point clouds. Background Art

[0002] With the accelerating pace of urbanization and the increasing number of large and complex buildings, the safety and stability of buildings have received widespread attention. 3D laser scanning technology, due to its high precision, high efficiency, and non-contact characteristics, is being used in building deformation monitoring.

[0003] 3D laser scanning technology typically begins by collecting spatial information about building surfaces using hardware devices such as lidar, structured light, or photogrammetry. Feature extraction and coordinate calibration are then performed to align and merge point clouds from different measurement time periods. Subsequently, registration methods based on iterative closest point algorithms or graph optimization are employed to ensure high accuracy and comparability across multiple scans, based on a unified benchmark.

[0004] However, while current 3D laser scanning technologies offer capabilities for registration, noise reduction, and structural crack identification, they are often limited to static or single-dimensional analysis of geometric information. When faced with large-scale data generated by multi-period acquisition, these methods struggle to simultaneously characterize material properties, stress field distribution, and microcrack characteristics during the alignment process. Furthermore, to capture dynamic deformation trends, most existing methods lack hierarchical modeling of stress concentration areas during regression or time series prediction, preventing flexible integration between local and global approaches and resulting in inaccurate predictions. Summary of the Invention

[0005] The embodiments of the present invention provide a method and device for predicting the time series of building deformation point clouds, so as to solve the problem that the current prediction methods of deformation point clouds cannot accurately predict.

[0006] In a first aspect, an embodiment of the present invention provides a method for predicting a time series of building deformation point clouds, comprising:

[0007] Performing coordinate system transformation and noise reduction on the acquired original point cloud data of the target building to obtain an aligned point cloud set; the original point cloud data includes point cloud data collected at multiple time periods and / or by multiple devices;

[0008] Perform local difference analysis on the aligned point cloud set to obtain the point cloud set of the abnormal area;

[0009] Based on the multi-resolution features of the point cloud set of the abnormal area, the topological features of the preset connection structure of the abnormal area are extracted, and a topological map of the abnormal area is constructed; stress constraints and material property constraints are applied to the topological map of the abnormal area to determine the latent space characteristics of the topological map of the abnormal area;

[0010] Based on the latent diffusion model, the latent space features are interpolated and extrapolated to determine the predicted vector field of the latent space features and the suspected damage area;

[0011] Perform multi-scale regression analysis on the suspected damage area to obtain the local deformation prediction value and local uncertainty distribution;

[0012] The three-dimensional deformation prediction field corresponding to the prediction vector field of the latent space feature, the local deformation prediction value and the local uncertainty distribution are fused and evaluated for consistency to obtain the fused global deformation prediction field;

[0013] The fused global deformation prediction field is analyzed to obtain the high-risk deformation distribution.

[0014] In one possible implementation, based on the multi-resolution features of the point cloud set of the abnormal area, the topological features of the preset connection structure of the abnormal area are extracted, and a topological map of the abnormal area is constructed, including:

[0015] The multi-resolution feature extraction of the point cloud set in the abnormal area is performed based on the multi-resolution encoder, wherein the multi-resolution encoder is based on the introduction of the gradient vector in the multi-resolution encoder to perform downsampling and feature extraction on the multi-resolution point cloud set in the abnormal area;

[0016] Extracting the topological features of the preset connection structure of the abnormal area based on the topological feature extraction module in the multi-resolution encoder to obtain the topological features of the preset connection structure of the abnormal area; wherein the preset connection structure includes beams, columns, trusses and facades of the target building;

[0017] Based on the topological features of the preset connection structure of the abnormal area, a topological map of the abnormal area is constructed.

[0018] In one possible implementation, stress constraints and material property constraints are applied to the topology of the abnormal region to determine the latent space characteristics of the topology of the abnormal region, including:

[0019] Apply stress constraints to the diffusion process of the latent space field corresponding to the topological map of the abnormal area to obtain the first latent space feature;

[0020] Based on the material property database, material property constraints are added to each diffusion process of the latent space field corresponding to the topological map of the abnormal area to obtain the second latent space feature;

[0021] Based on the deformation amplitude and gradient, the deformation area of the latent space field corresponding to the topological map of the abnormal area is constrained to obtain the third latent space feature;

[0022] Based on the first latent space feature, the second latent space feature, and the third latent space feature, a latent space feature of a topological map of the abnormal region is determined.

[0023] In one possible implementation, based on the latent diffusion model, latent space features are interpolated and extrapolated to determine the predicted vector field of the latent space features and the suspected damage area, including:

[0024] Based on the trained latent diffusion model and the stress and material property constraints in the latent space, adjacent time segments in the latent space feature vector field are interpolated and extrapolated to obtain the predicted deformation vector field of the latent space feature.

[0025] The deformation intensity of the topological features of the preset connection structure is calculated, and the suspected damage area is determined based on the preset deformation threshold and the predicted deformation vector field of the latent space features.

[0026] In one possible implementation, a multi-scale regression analysis is performed on the suspected damage area to obtain the local deformation prediction value and the local uncertainty distribution, including:

[0027] Based on the positioning information of the point set in the suspected damage area and the spatial filtering operator, high-risk suspected areas are screened; the positioning information of the point set includes the coordinates of each point and the predicted deformation vector field of each point;

[0028] Based on a pre-constructed composite kernel function, multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain local deformation prediction values and local uncertainty distribution; the composite kernel function includes geometric correlation, material property similarity kernel function and topological coupling of preset connection structures.

[0029] In one possible implementation, a multi-scale regression analysis is performed on each scale level in the high-risk suspected area based on a pre-built composite kernel function to obtain local deformation prediction values and local uncertainty distributions, including:

[0030] Based on the pre-built composite kernel function, a multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain the first regression result;

[0031] Dynamically updating the preset weights in the composite kernel function to obtain the final weight of the composite kernel function; wherein the geometric correlation, the material property similarity kernel function, and the topological coupling of the preset connection structure in the composite kernel function are each provided with a corresponding preset weight;

[0032] Based on the final weight of the composite kernel function, a multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain the second regression result;

[0033] The first regression result and the second regression result are integrated to obtain the local deformation prediction value and the local uncertainty distribution.

[0034] In one possible implementation, the three-dimensional deformation prediction field corresponding to the prediction vector field of the latent space features, the local deformation prediction value, and the local uncertainty distribution are fused and evaluated for consistency to obtain a fused global prediction field, including:

[0035] Based on the decoder, the predicted vector field of the latent space features is mapped to the physical coordinate system to obtain the three-dimensional deformation prediction field of the latent space features;

[0036] Construct a damage map based on the predicted vector field and suspected damage areas of latent space features;

[0037] Based on the 3D deformation prediction field and damage map, a global prediction field of latent space features is constructed;

[0038] Determining a global skeleton of the global prediction field; wherein a function of the global skeleton is determined by a double integral of a correlation weight of the target position and a first deformation difference, where the first deformation difference is a difference between the three-dimensional deformation prediction field and a preset average reference;

[0039] Construct the weights of local parameters and global skeleton, and fuse the local parameters and global skeleton to obtain the fused prediction field; where the local parameters include the local deformation prediction value and the local uncertainty distribution;

[0040] Based on the differences between the local parameters and the global skeleton in the high-risk suspected area, as well as the differences between the gradient of the fused prediction field and the gradient of the global skeleton, the fused prediction field is verified to obtain the fused global deformation prediction field and damage distribution.

[0041] In one possible implementation, the fused global deformation prediction field is analyzed to obtain a high-risk deformation distribution, including:

[0042] Based on the fused global deformation prediction field and high-risk suspected areas, the deformation degree and uncertainty in the high-risk suspected areas are determined;

[0043] The deformation degree and uncertainty in the high-risk suspected area and the high-risk suspected area are input into the weight update module of the latent diffusion model, and an iterative process is performed to obtain the updated weight of the latent diffusion model;

[0044] Based on the composite kernel function, multi-scale regression analysis is performed on the high-risk suspected areas to obtain the high-risk deformation distribution.

[0045] In one possible implementation, local difference analysis is performed on the aligned point cloud set to obtain a point cloud set in the abnormal area, including:

[0046] Determine the strength of each point cloud in the aligned point cloud set based on a pre-built anomaly evaluation model;

[0047] Based on the intensities of all point clouds in the aligned point cloud set, the point cloud set of the abnormal area is determined.

[0048] In a second aspect, an embodiment of the present invention provides a prediction device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method in the first aspect or any possible implementation of the first aspect is implemented.

[0049] In an embodiment of the present invention, to comprehensively consider the impact of various factors on building deformation, the acquired raw point cloud data of the target building is first subjected to coordinate system transformation and noise reduction to obtain an aligned point cloud set. After obtaining the aligned point cloud set, local dissimilarity analysis is performed on the aligned point cloud set to narrow the study area, resulting in a point cloud set for the abnormal region. This allows for more accurate study of the point cloud set in the abnormal region. Next, based on the multi-resolution features of the point cloud set in the abnormal region, the topological features of the pre-set connection structure in the abnormal region are extracted, and a topological map of the abnormal region is constructed. Stress and material property constraints are applied to the topological map of the abnormal region, and the latent space features of the topological map of the abnormal region are determined. This allows for capturing the local details of stress concentrations in the pre-set connection structure. Combined with the stress and material property constraints, material and mechanical constraints are reflected in the diffusion process, enabling more accurate deformation prediction. Next, based on the latent diffusion model, the latent space features are interpolated and extrapolated to determine the predicted vector field of the latent space features and suspected damage regions. This allows for capturing areas of potential damage by combining stress and material information at the latent space level. After identifying suspected damage areas, multi-scale regression analysis is performed on these areas to obtain local deformation predictions and local uncertainty distributions, providing a basis for refined compensation in subsequent fusion with the global framework. To ultimately form a global deformation field and record the damage distribution within a unified framework, the three-dimensional deformation prediction field corresponding to the predicted vector field of latent space features, the local deformation predictions, and the local uncertainty distributions must be fused and evaluated for consistency to obtain a fused global deformation prediction field. Finally, the fused global deformation prediction field is analyzed to obtain a high-risk deformation distribution. This allows for a more accurate generation of a high-risk deformation distribution over a continuous time series. During the prediction process, a coordinated analysis of stress distribution, material properties, and local structural differences is implemented, particularly for fine cracks, complex component connections, and the combined effects of multiple materials. This improves the integration of global and local information, enabling more accurate deformation prediction within the time series dimension. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1This is a flowchart of an implementation method for predicting building deformation point cloud time series provided by an embodiment of the present invention;

[0051] Figure 2 3 is a schematic diagram of the structure of a device for predicting building deformation point cloud time series provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] The inventors discovered that while current 3D point cloud monitoring and data fusion methods offer the capabilities of registration, noise reduction, and structural crack identification, they are often limited to static or single-dimensional analysis of geometric information. When faced with large-scale data generated by multi-period acquisition, these methods struggle to simultaneously characterize material properties, stress field distribution, and microscopic crack characteristics during the alignment process. Existing studies typically utilize traditional filtering and matching algorithms to focus on data cleaning or macroscopic deformation detection, but fail to achieve deeper dynamic tracking of local, detailed structural and material differences. For complex building components, such as beam-column connections and irregular facades, different materials may exhibit inconsistent deformation behavior. However, existing methods typically rely on additional offline simulations or empirical parameters to determine these local differences, failing to address them in a timely and unified manner during multi-period monitoring and prediction. Furthermore, to capture dynamic deformation trends, most existing methods lack hierarchical modeling for stress concentration areas during regression or time series prediction, and struggle to achieve flexible integration between the local and global domains. This results in insufficient insight into potential crack evolution and stress distribution in heterogeneous materials.

[0054] In order to solve the above problems, the present invention provides a method and device for predicting the time series of building deformation point clouds.

[0055] See also Figure 1 , which shows a flow chart for implementing a method for predicting building deformation point cloud time series provided by an embodiment of the present invention, and is described in detail as follows:

[0056] S110 , performing coordinate system transformation and noise reduction processing on the acquired original point cloud data of the target building to obtain an aligned point cloud set.

[0057] Among them, the original point cloud data includes point cloud data collected from multiple time periods and multiple devices.

[0058] In some embodiments, coordinate system transformation and noise reduction processing may be performed based on environmental parameters and hardware properties of each device.

[0059] S1110. Based on the recorded environmental parameters and hardware properties of each device, quantitatively analyze the noise and density distribution.

[0060] Define the sampling data matrix , Indicates the time index With device index The corresponding raw scan value. Construct function , used to characterize the same device Under different time periods and The noise difference between the accumulation:

[0061] ;

[0062] in, Indicates the time index Under device index k, the corresponding coordinates The scanning intensity value at and Respectively represent the index range of the point cloud in the horizontal and vertical directions. The peak distribution law of the noise and the environmental parameters in the measurement scene are combined to complete the noise feature prediction of the target area.

[0063] S1120, coordinate system transformation

[0064] Indexing devices The corresponding original point cloud coordinate set , let the rotation matrix With translation vector Based on the initial external parameter calibration, the conversion between the hybrid coordinate system and the standard coordinate system is performed:

[0065] ;

[0066] in, Represents the corrected coordinate vector in the standard coordinate system. By preserving the sensor's intrinsic and extrinsic parameters, subsequent batch calibration steps can synchronize multi-source data based on the same benchmark.

[0067] S1130, based on the architectural structure layout of the target building, select several stable and easily identifiable reference points as key reference points for accurate positioning in subsequent time sequence alignment. Each key reference point is stored as , where r is the reference point number. The coordinate information of these key reference points is mapped to the unified coordinate system obtained in step S1120.

[0068] S1140, combined with step S1110 The result of the distribution is recorded as the noise threshold , define the judgment function , for each 3D point The distance value to its neighborhood center is monitored. If it exceeds the noise threshold , then remove the point and count the remaining point clouds:

[0069] ;

[0070] in Represents the centroid vector of the points in the neighborhood.

[0071] By executing batch by batch This step continuously narrows the anomaly range, reduces misjudgment of normal structure points, and obtains a point cloud with less noise.

[0072] S1150, the remaining point cloud is unified by step size and scale factor Perform downsampling. Let the remaining point cloud be ,go through Function Get , so that multi-time point clouds can be represented within the same resolution scale, reducing the mismatch between different densities and taking into account the computational efficiency of the subsequent latent space model.

[0073] S1160 , perform registration time period by time period.

[0074] Given adjacent time periods and Corresponding point cloud collection and , find the rotation matrix and translation vectors , minimize the sum of squared Euclidean distances between corresponding points:

[0075] ;

[0076] in and Respectively indicate the time period and time period The coordinates of the corresponding points in .

[0077] Therefore, based on the results output from the above steps, a multi-time period aligned point cloud set can be formed.

[0078] S120 , performing local difference analysis on the aligned point cloud set to obtain a point cloud set in the abnormal area.

[0079] In order to provide candidate areas for subsequent processing and correction and to process abnormal point clouds more specifically, the present invention performs local difference analysis on the point cloud sequence, thereby quickly locking the abnormal area.

[0080] In some embodiments, the intensity of each point cloud in the aligned point cloud set can be first determined based on a pre-built anomaly assessment model. The anomaly assessment model is constructed based on the difference between the grayscale or intensity information of each point cloud and a preset average intensity value. Then, based on the intensities of all point clouds in the aligned point cloud set, the point cloud set in the abnormal region can be determined.

[0081] In this embodiment, the aligned point cloud set is denoted as , and for each Perform local difference analysis. Set the time index Down, The coordinate index is , whose grayscale or intensity information is .

[0082] Constructed abnormal evaluation model The distribution of noise or deformation anomalies in each point cloud frame can be quantified:

[0083] ;

[0084] in, Indicates the visible area in the two-dimensional index The integral domain on , Indicates the average intensity value of Xᵗ in the area. Calculate , it is possible to locate the noise concentration area and the candidate area with large displacement gradient. The larger the value, the more significant the noise or geometric changes in the time slice. , you can quickly lock the "abnormal" area.

[0085] S130 , extracting topological features of a preset connection structure of the abnormal area based on the multi-resolution features of the point cloud set of the abnormal area, and constructing a topological map of the abnormal area.

[0086] In order to preserve the overall outline of the target building while extracting the local details of the preset connection structure of key parts such as beams and columns, an encoder is used to perform multi-resolution feature extraction.

[0087] In some embodiments, a multi-resolution encoder can be used to extract multiple resolutions of the point cloud set of the abnormal region to obtain multi-resolution features of the point cloud set of the abnormal region. The multi-resolution encoder downsamples and extracts features from the multi-resolution point cloud set of the abnormal region based on the introduction of gradient vectors into the multi-resolution encoder.

[0088] Then, the topological features of the preset connection structure of the abnormal area are extracted based on the topological feature extraction module in the multi-resolution encoder to obtain the topological features of the preset connection structure of the abnormal area. The preset connection structure includes beams, columns, trusses and facades of the target building.

[0089] Finally, based on the topological features of the preset connection structure of the abnormal area, a topological map of the abnormal area is constructed.

[0090] In this embodiment, after the abnormal area is determined, the point cloud needs to be characterized at multiple scales to take into account both the overall structural morphology and local details.

[0091] Let the encoder be , whose input is , the output is ,in Indicates the resolution level. To ensure geometric consistency at multiple scales, the vector Indicates the resolution Corresponding position below Feature representation, adaptive downsampling and feature extraction:

[0092] ;

[0093] in, Indicates the resolution The channel dimension under Indicates that In position The local geometric gradient vector, is the weight set, Resolution The bias term below.

[0094] In addition, in order to strengthen the topological information of the preset connection structure and the topological information of complex spatial hierarchical parts such as beams, columns, trusses and facades, a topological feature extraction module is added. , whose output is recorded as .

[0095] ;

[0096] Among them, suppose each local connection structure contains construction units, and the topological connection matrix of each construction unit is , corresponding to the node set , cc is the corresponding node set The multi-scale feature channel matrix of each node in the network can accurately capture local details and preset connection structures.

[0097] Finally, construct a topological map of the abnormal area .

[0098] S140 , performing stress constraints and material property constraints on the topological map of the abnormal region to determine the latent space characteristics of the topological map of the abnormal region.

[0099] Because the target building's structure also involves stress constraints and material property constraints, relying solely on geometric denoising can disrupt the originally satisfied mechanical balance. Therefore, this application explicitly incorporates the "stress field" into the iterative process of latent space denoising, ensuring that the denoised deformation results are not only smooth but also consistent with architectural mechanics.

[0100] In some embodiments, stress constraints may be first applied to the diffusion process of the latent space field corresponding to the topological map of the abnormal region to obtain a first latent space feature.

[0101] Then, based on the material property database, material property constraints are added to each diffusion process of the latent space field corresponding to the topological map of the abnormal area to obtain the second latent space feature.

[0102] Then, based on the deformation amplitude and gradient, the deformation area of the latent space field corresponding to the topological map of the abnormal area is constrained to obtain the third latent space feature.

[0103] Finally, based on the first latent space feature, the second latent space feature and the third latent space feature, the latent space feature of the topological map of the abnormal region is determined.

[0104] In this embodiment, an energy functional can be constructed based on adding a gravitational field to the topological map of the abnormal region to impose stress constraints on the diffusion process. Indicates time ,coordinate The estimated stress magnitude under Incorporating latent space diffusion to construct energy functionals Used to guide the latent space diffusion process,

[0105] ;

[0106] in is the volume area of the building target structure, represents the stress estimate at the corresponding position after latent space mapping, It is the weighted coefficient for adjusting the smoothness of the latent space and the matching degree of the stress field. , which can be used in the denoising iteration Consistent with the real structural mechanical characteristics. On the basis of conventional geometric denoising, the stress field is explicitly introduced as a physical prior, and the diffusion / denoising process of the latent space field is guided by constructing and minimizing the energy functional. This variational framework of mechanical constraints + smooth regularization can make the final deformation field It not only gets rid of noise interference, but also does not violate the mechanical characteristics of the building, thus providing a more reliable basis for subsequent material, crack and damage assessments.

[0107] In this embodiment, the material property database is used , in order to further distinguish the mechanical properties of different building materials in the latent space. Assume that the material type is in the coordinate The table lookup result is , which is in The parameter vector in is denoted as The diffusion process introduces material influence factors at each iteration step. :

[0108] ;

[0109] in, represents the elastic modulus, represents Poisson's ratio, and the rest are other material mechanical parameters, is an adjustable coefficient.

[0110] By dynamically adjusting the noise injection amount and diffusion stability range during the latent space update process, invalid white noise interference is reduced and the deformation characteristics caused by material differences are highlighted.

[0111] In this embodiment, the easily deformed or cracked areas are processed in a concentrated manner to enhance the damage information fusion in the diffusion denoising process. Let the easily deformed area set be denoted as , which is in Space corresponding coordinate notation ). In order to directionally amplify the deformation characteristics of these areas, a directional strengthening damage function is constructed based on the deformation amplitude and gradient.

[0112] ;

[0113] in and Respectively control the strength of deformation amplitude and local gradient. Incorporate optimization objectives to reveal potential damage signs such as tiny cracks and stress concentration areas.

[0114] Finally, the latent space features of the topological map of the abnormal region not only incorporate multi-resolution geometric features but also incorporate material and stress characteristics. However, since this is only distributed over discrete time intervals t = 1, …, T, there is still some noise or discontinuity. Therefore, temporal interpolation of the latent space representation is necessary to further smooth, diffuse, or infer deformation evolution trends.

[0115] S150. Based on the latent diffusion model, interpolation and extrapolation processing are performed on the latent space features to determine the predicted vector field of the latent space features and the suspected damage area.

[0116] In some embodiments, first, based on the trained latent diffusion model and the stress constraints and material property constraints in the latent space, adjacent time segments in the vector field of the latent space features can be interpolated and the vector field of the latent space features can be extrapolated to obtain a predicted deformation vector field of the latent space features.

[0117] Finally, the deformation intensity of the topological features of the preset connection structure is calculated, and the suspected damage area is determined based on the preset deformation threshold and the predicted deformation vector field of the latent space features.

[0118] In this embodiment, the latent space features of the topological map of the abnormal region are recorded as ,in Indicates the In order to accurately preserve key structural information in time series interpolation and subsequent extrapolation operations, a latent diffusion model with a guided mechanism is constructed. network , embedding physical priors and material layering information into the network weight initialization process. The core parameter set is , define the initial boot function As a network guide item, the latent space features mentioned above After superposition, we get The initial parameters are determined by minimizing the following objective function :

[0119] ;

[0120] in represents the latent space coordinate domain, is the weighting coefficient for adjusting the smoothness of the model parameters.

[0121] Incorporating stress and material property constraints into initial training makes the network more sensitive to vulnerable areas during subsequent interpolation and denoising. By introducing the LDM as a "neural network model with physical / material priors," it learns the dynamic evolution of latent space representations, making it particularly sensitive to weak / vulnerable structural areas during interpolation and denoising.

[0122] After the LDM training is completed, The network is applied to the interpolation of adjacent time slices. Given adjacent time slices ,exist and Insert between Intermediate deformation frames to generate an interpolation sequence . Define the dynamic diffusion operator ,make In combination with guidance mechanisms during diffusion:

[0123] ;

[0124] in is the latent space vector of the previous interpolated frame, represents the latent space vector for the next time slice, Indicates in The physical fields and noise suppression factors used in the step interpolation.

[0125] The distortion of uneven temporal sequence is reduced by step-by-step diffusion composite interpolation, making the overall deformation trajectory of the target building continuous and smooth in the time dimension.

[0126] Then, at the current known moment On, for the subsequent future moments Deformation extrapolation is performed to support early detection of potential anomalies. Define the extrapolation operation , referring to the generated interpolation sequence and mechanical constraints, estimate the time Latent space representation of .

[0127] Let the extrapolation order be , comprehensive historical multi-frame interpolation results ,…, , The following recursive formula is obtained with the structural material prior:

[0128]

[0129] in represents the extrapolation coefficient, and represent the stress constraint field and material property field in the latent space respectively, The boundary of the building target area.

[0130] By synergistically considering multi-frame interpolation and material mechanics layering information, the prediction Potential deformation trend.

[0131] Therefore, the predicted deformation vector field of the latent space features can be obtained based on the results of the above steps.

[0132] Next, it is necessary to determine the suspected damage area based on the predicted deformation vector field of the latent space characteristics, the deformation intensity of the topological characteristics of the preset connection structure, and the preset deformation threshold.

[0133] Deformation assessment is performed based on the topological map of the abnormal area. A representative component unit , and calculate its deformation strength in combination with the topological features extracted in step S2.

[0134] Defining a deformation strength measurement system :

[0135] ;

[0136] in In the latent space Down The gradient vector of the position, represents the structural stress value, Material characteristic factor is the weighting coefficient. By determining which components have the highest degree of deformation, we can focus on damage screening of these high-strength areas.

[0137] based on The result and the preset deformation threshold are used to determine whether the deformation exceeds the threshold. suspected damaged area.

[0138] Select judgment function :

[0139] ;

[0140] in, (x) indicates Topology fragment deformation strength.

[0141] This indicates abnormal deformation. These points are added to a list of suspected damage, preserving their coordinates and adjacent reference markers. Additionally, additional diffusion denoising cycles are performed on suspected damage areas to suppress sudden noise errors.

[0142] Summarize all suspected damage areas and match them with the continuous time-series deformation trajectory generated by the interpolation and extrapolation sequence to form a preliminary damage map . In time index The spatial coordinates and deformation amplitude of the suspected damage points are recorded. By comparing the position changes of the suspicious areas after multiple iterations across time slices, local anomalies are combined with the overall deformation trend.

[0143] Finally, the final interpolation is compared with the extrapolated sequence and damage map Output: Maintain consistency across both adjacent and future time periods, and consolidate suspected cracks or concentrated beam-column connection locations into the damage demarcation list.

[0144] S160. Perform multi-scale regression analysis on the suspected damage area to obtain a local deformation prediction value and a local uncertainty distribution.

[0145] In some embodiments, high-risk suspected areas may be screened based on the positioning information of a point set in the suspected damage area and a spatial filtering operator, wherein the positioning information of the point set includes the coordinates of each point and the predicted deformation vector field of each point.

[0146] Then, based on a pre-built composite kernel function, a multi-scale regression analysis is performed at each scale level within the high-risk suspected area to obtain local deformation predictions and local uncertainty distributions. The composite kernel function includes geometric correlation, material property similarity kernel functions, and topological coupling of pre-set connection structures.

[0147] In this embodiment, first, based on a pre-constructed composite kernel function, a multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain a first regression result.

[0148] Then, the preset weights in the composite kernel function are dynamically updated to obtain the final weight of the composite kernel function. Among them, the geometric correlation, material property similarity kernel function, and topological coupling of the preset connection structure in the composite kernel function are each corresponding to a preset weight.

[0149] Then, based on the final weight of the composite kernel function, a multiscale regression analysis is performed on each scale level in the high-risk suspected area to obtain the second regression result.

[0150] Finally, the first regression result and the second regression result are integrated to obtain the local deformation prediction value and the local uncertainty distribution.

[0151] Specifically, the positioning information of the point set in the suspected damage area is formed into a target data set , containing coordinates and its predicted deformation vector field .

[0152] Define the target domain Represents a local area containing potential cracks or component defects, using the following spatial filtering operator (·) Screen out high-risk suspected areas of building structures :

[0153] ;

[0154] in Pass threshold Determine whether the point belongs to a high-risk suspected area. If it meets , then the point is determined to be the input that needs to be focused on.

[0155] In addition, based on the traditional Gaussian process regression kernel function, the present invention combines the geometric characteristics of buildings and material differences to propose a composite kernel function The spatial proximity, topological coupling and material properties, i.e. geometric correlation, material property similarity kernel function and topological coupling of the preset connection structure are constructed as a comprehensive metric:

[0156] ;

[0157] in, represents the geometric correlation based on Euclidean distance,

[0158] represents the kernel function of the similarity of material properties between material parameters,

[0159] for The material property vector at location , Indicates the topological coupling degree of local preset connection structures such as beam-column connection or facade connection, is an adjustable weighting coefficient.

[0160] Multi-scale regression is performed on high-risk suspected areas, and regional deformation is fitted step by step from macroscopic scale to crack details. Assume that each area has scale level , call each layer Inference Operator , output the prediction results :

[0161] ;

[0162] in Indicated on scale The prior mean function under , represents the kernel vector, is the kernel matrix, represents the observation data vector processed by the scaling layer, is the prior mean on the corresponding dimension. To the micro crack layer The fitting is refined step by step, and the prediction uncertainty interval is shrunk according to the error metric function after each layer of regression is completed, until a layered deformation prediction that meets the accuracy requirements is obtained.

[0163] In addition, in order to perform regression analysis more accurately, a weighted update mechanism can also be defined (·), for the composite kernel function Perform dynamic tuning:

[0164] ;

[0165] in, is the number of iterations, Indicates in The mean square error between the regression and the observed value is, is the learning rate.

[0166] After determining the final weights based on the composite kernel function, multi-scale regression analysis was performed again on each scale level in the high-risk suspected area to obtain the second regression results.

[0167] The results of the two regression analyses are integrated to generate accurate deformation prediction values and local uncertainty distributions for the local area.

[0168] Defining a result set Includes every location in all regions The predicted value and variance information of:

[0169]

[0170] in Indicates that based on the improved The final shape prediction of represents the uncertainty measure, which is estimated by regressing the variance of the posterior distribution.

[0171] This can provide a high-precision local supplement for the deformation assessment of the target building, and is suitable for high-risk areas such as beam-column junctions or crack ends.

[0172] S170 , fusing and evaluating the consistency of the three-dimensional deformation prediction field, the local deformation prediction value, and the local uncertainty distribution corresponding to the prediction vector field of the latent space feature to obtain a fused global deformation prediction field.

[0173] In some embodiments, first, the prediction vector field of the latent space feature may be mapped to a physical coordinate system based on the decoder to obtain a three-dimensional deformation prediction field of the latent space feature.

[0174] Then, a damage map is constructed based on the predicted vector field of latent space features and suspected damage areas.

[0175] Secondly, based on the three-dimensional deformation prediction field and damage map, a global prediction field of latent space features is constructed.

[0176] Then, a global skeleton of the global prediction field is determined, wherein the function of the global skeleton is determined by the double integral of the correlation weight of the target position and the first deformation difference, which is the difference between the three-dimensional deformation prediction field and a preset average benchmark.

[0177] Next, the weights of local parameters and global skeleton are constructed and fused to obtain the fused prediction field. The local parameters include the local deformation prediction value and the local uncertainty distribution.

[0178] Finally, based on the differences between the local parameters and the global skeleton in the high-risk suspected area, as well as the differences between the gradient of the fused prediction field and the gradient of the global skeleton, the fused prediction field is verified to obtain the fused global deformation prediction field and damage distribution.

[0179] In this embodiment, the predicted vector field of the latent space features is:

[0180] ;

[0181] in,

[0182] To facilitate global-local fusion, the decoder of the same network is called , maps the predicted vector field of the latent space feature back to the physical coordinate system to form a three-dimensional deformation prediction field

[0183] ;

[0184] in, Indicates at time , spatial location The displacement vector.

[0185] Based on the 3D deformation prediction field and damage map, the global prediction field of the latent space feature is constructed as follows:

[0186] ;

[0187] in, Represents the time index, The spatial coordinates representing the overall scope of the building are used to maintain the macro outline and trend in the subsequent fusion process.

[0188] Further build global field average benchmark

[0189] ;

[0190] Describe the entire building at all times The average deformation of is used as the reference zero-datum for subsequent main ridge extraction.

[0191] The relevance weight with the target position can be expressed as, for any target position , the time-varying correlation weight is:

[0192] ;

[0193] in Controls the decay scale of spatial and deformation amplitudes.

[0194] In this embodiment, the process of determining the global skeleton of the global prediction field is as follows:

[0195] Global Skeleton Functions Represents the overall deformation reference skeleton.

[0196] Select the main ridge line and important topological node information of the deformation, and filter out local high-frequency noise:

[0197] ;

[0198] in, Represents the target position The relevance weight of is the average benchmark of the global field, For the overall building space area, is the three-dimensional deformation prediction field.

[0199] The global skeleton function provides a global coordinate frame in subsequent steps to ensure that small-scale correction actions do not destroy the overall deformation trend.

[0200] In this embodiment, after determining the global skeleton, local deformation prediction values, and local uncertainty distribution, fusion processing can be performed to integrate local details with global information. Multiple evaluations and cyclic verifications are required before and after fusion to improve the overall accuracy and stability of the deformation field. The specific fusion and verification process is as follows:

[0201] Before fusion, evaluate the local and global error ranges and set the evaluation function Measure the difference between the local prediction and the global skeleton:

[0202] ;

[0203] in, is the local uncertainty, A set of coordinates representing high-risk areas.

[0204] For high-risk areas, a weight mapping function is constructed to prevent small-scale damage from being ignored in the macro skeleton. , in high-risk areas Take a larger value near the edge and gradually decay in other areas:

[0205] ;

[0206] in, is the prediction field obtained after fusion, Adjust the weight of global skeleton and local high-precision regression results. When the local strain is large or the uncertainty is high, Trend 1 emphasizes reliance on local corrections to ensure a balance between low-frequency and high-frequency features in the global field.

[0207] After completing the initial integration, you can also A secondary comparison is performed on the crack endpoints or strain sudden increase positions in the image to ensure that the correction will not cause local distortion and will be consistent with the global skeleton trend.

[0208] Assume local verification function Combining the crack edge coordinates with the fused field gradient, perform the following checks:

[0209] ;

[0210] like Exceeding the threshold , it means that the local gradient changes too much after fusion. Do extreme value limiting to avoid overcorrection.

[0211] Then you can also loop check, and The two items are simultaneously included in the fusion result evaluation criteria, and the following fusion scoring function is set :

[0212] ;

[0213] in, and is the adjustment coefficient, by minimizing , for the fused prediction field Verification eliminates local overfitting and global deviation, and the differences between sub-scales are continuously leveled, ultimately improving the overall accuracy and stability of the deformation field.

[0214] After completing all the loop checks, the final fused prediction field and damage distribution information Output together, Covering architectural detail areas.

[0215] Define output carrier The deformation vector and risk mark recorded at each time t:

[0216] ;

[0217] The output can truly reflect the coexistence of macrostructural deformation and local damage, providing an intuitive and visual basis for building maintenance and monitoring strategies.

[0218] S180. Analyze the fused global deformation prediction field to obtain a high-risk deformation distribution.

[0219] In some embodiments, first, based on the fused global deformation prediction field and the high-risk suspected area, the deformation degree and uncertainty in the high-risk suspected area may be determined.

[0220] Then, the deformation degree and uncertainty in the high-risk suspected area and the high-risk suspected area are input into the weight update module of the latent diffusion model for iterative processing to obtain the updated weight of the latent diffusion model.

[0221] Finally, based on the composite kernel function, multi-scale regression analysis is performed on the high-risk suspected areas to obtain the high-risk deformation distribution.

[0222] In this embodiment, the fused global deformation prediction field and high-risk suspected areas are centrally stored to form a data set ,

[0223] in, High-risk suspected area, v i represents the predicted deformation value at this location, Indicates the corresponding uncertainty.

[0224] Constructing a cumulative risk assessment function Used to calculate the comprehensive level of deformation and uncertainty in high-risk areas:

[0225] ;

[0226] in It is a coefficient for adjusting the relative weight of deformation amplitude and uncertainty.

[0227] By quantifying the deformation evolution trend of high-risk areas, a reference is provided for guiding the update of the latent space.

[0228] In this embodiment, the extracted With local high-risk distribution Input to the guided weight update module of the latent space diffusion network.

[0229] Defining the key parameter set of the latent space diffusion network , and in each iteration the guide weights Dynamically adjust to focus more on the structural details of high-risk areas when mapping the latent space. The number of iterations is :

[0230] ;

[0231] in, is the learning rate, Represents the loss function of the diffusion network, including the high-risk suspected areas The weight gain makes the gradient descent pay more attention to the latent space representation at the cracks or strain concentration points.

[0232] After completing the latent space guidance weight update, the high-risk suspected area Perform a new round of improved Gaussian process regression. Based on the composite kernel function , and in this round of iteration, the updated weight factor is brought into the kernel function parameter part. The regression output is recorded as , corresponding to the new predicted value of the high-risk area:

[0233] ;

[0234] in, is the prior mean function of this round of iteration, represents the kernel vector, is the calculation result of the improved kernel function after introducing new weights, is the observation value vector of the previous round of prediction.

[0235] Through re-regression, the latent space correction information is integrated with material differences and topological connection elements to further refine the accuracy of the prediction of cracks or strain concentration points.

[0236] In addition, in order to avoid local overfitting or noise superposition caused by multiple corrections, the global skeleton and The difference between them is statistically analyzed to define the matching metric :

[0237] ;

[0238] in, is the weighting coefficient for high-risk suspected areas or high-noise areas to highlight the difference between these areas and the global skeleton. If it is too large, there is a risk of noise accumulation, and it is necessary to appropriately weaken the latent space update intensity or add a smoothing term in the regression. This ensures that after multiple rounds of correction, the local correction still maintains a robust fit with the overall trend.

[0239] For local updates With the initial deformation field Fusion again, record the fusion field as , to gradually correct the overall deformation distribution.

[0240] Assume the fusion function Mapping local and global deformations:

[0241]

[0242] in is a mapping A function in the interval [0,1] is used to adjust the fusion ratio of local and global deformations. Maintain the updated timing curve in the time dimension | , improve the continuity and accuracy of deformation evolution.

[0243] Therefore, by performing multi-scale regression analysis on high-risk suspected areas, the high-risk deformation distribution can be obtained. The result includes both the global and local fused deformation fields and the local regression predictions of local details.

[0244] In addition, after obtaining the high-risk deformation distribution, it can also be visualized and risk identified. After data is reorganized for each period, 3D rendering is performed to highlight the deformation and crack extension of key components such as beams, columns, facades, and high-risk cracks. Visual reports and classification lists are then recorded for reference in subsequent structural assessments and maintenance strategies. The visualization process is as follows:

[0245] First, load the high-risk deformation distribution.

[0246] To reshape the data before visualization and risk analysis, define the merge function , in the time index Aggregate the combined global and local deformations within the scope, based on the following expression convergence operation for all moments:

[0247] ;

[0248] in, Adjust the convergence ratio of global and local predictions.

[0249] Get the full temporal deformation set , laying a consistent data foundation for visualization rendering and damage analysis.

[0250] Next, for the merged deformation set Perform 3D visualization rendering and highlight beams, columns, facades, and key areas with high risk of cracks, namely pre-set connection structures, using different colors and line thicknesses.

[0251] Introducing the view transformation matrix , project the deformation vector into an interactive visualization space and define the mapping operation :

[0252] ;

[0253] in, represents the plane coordinates, Indicates time The vertical displacement or deformation value at and They are the screen coordinates and depth information after rendering respectively.

[0254] The Gaussian process uncertainty distribution and latent space prediction field , quantify the potential risk level of a component or crack, and define the damage scoring function :

[0255] ;

[0256] in, express The surrounding local neighborhood, represents the deformation increment field in the latent space, represents the uncertainty at the same position, and The coefficient for adjusting the weights of different factors.

[0257] pass The potential damage levels of components and cracks are divided according to the size of the values, and high, medium and low risk intervals are initially formed.

[0258] for Exceeding a certain risk threshold This step includes the coordinate position and its time change information into the classification list. Establish a record matrix , where the row index corresponds to the high-risk coordinate points, and the column index covers the deformation increment, crack expansion and historical trend:

[0259] ;

[0260] like , a new entry is created in the list to record the stress anomaly and crack expansion rate of x to support subsequent structural maintenance and reinforcement strategies.

[0261] Then, based on the results generated in the above steps, a visualization report and data interface are generated. The report includes a time series diagram of the global deformation of the building, a diagram of the expansion of local cracks, and a detailed list of high-risk items. , and output the results in a standardized manner:

[0262] ;

[0263] in, 、 、 They represent the visual encapsulation of charts, risk data, and time series information, respectively, for subsequent calls by external security assessment systems or maintenance strategy modules.

[0264] Finally, the final building deformation and damage information set is output , which includes visual charts, risk level labels and a classification list of high-risk areas:

[0265] ;

[0266] Through this collection, the building deformation trajectory can be accurately displayed in the time dimension and the degree of structural damage can be indicated with quantitative indicators, ultimately providing an intuitive reference for subsequent safety assessment and maintenance decisions.

[0267] In an embodiment of the present invention, to comprehensively consider the impact of various factors on building deformation, the acquired raw point cloud data of the target building is first subjected to coordinate system transformation and noise reduction to obtain an aligned point cloud set. After obtaining the aligned point cloud set, local dissimilarity analysis is performed on the aligned point cloud set to narrow the study area, resulting in a point cloud set for the abnormal region. This allows for more accurate study of the point cloud set in the abnormal region. Next, based on the multi-resolution features of the point cloud set in the abnormal region, the topological features of the pre-set connection structure in the abnormal region are extracted, and a topological map of the abnormal region is constructed. Stress and material property constraints are applied to the topological map of the abnormal region, and the latent space features of the topological map of the abnormal region are determined. This allows for capturing the local details of stress concentrations in the pre-set connection structure. Combined with the stress and material property constraints, material and mechanical constraints are reflected in the diffusion process, enabling more accurate deformation prediction. Next, based on the latent diffusion model, the latent space features are interpolated and extrapolated to determine the predicted vector field of the latent space features and suspected damage regions. This allows for capturing areas of potential damage by combining stress and material information at the latent space level. After identifying suspected damage areas, multi-scale regression analysis is performed on these areas to obtain local deformation predictions and local uncertainty distributions, providing a basis for refined compensation in subsequent fusion with the global framework. To ultimately form a global deformation field and record the damage distribution within a unified framework, the three-dimensional deformation prediction field corresponding to the predicted vector field of latent space features, the local deformation predictions, and the local uncertainty distributions must be fused and evaluated for consistency to obtain a fused global deformation prediction field. Finally, the fused global deformation prediction field is analyzed to obtain a high-risk deformation distribution. This allows for a more accurate generation of a high-risk deformation distribution over a continuous time series. During the prediction process, a coordinated analysis of stress distribution, material properties, and local structural differences is implemented, particularly for fine cracks, complex component connections, and the combined effects of multiple materials. This improves the integration of global and local information, enabling more accurate deformation prediction within the time series dimension.

[0268] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0269] The following are device embodiments of the present invention. For details not fully described therein, reference may be made to the corresponding method embodiments described above.

[0270] Figure 2 The following is a schematic diagram showing the structure of a device for predicting a time series of building deformation point clouds provided by an embodiment of the present invention. For ease of explanation, only the parts related to the embodiment of the present invention are shown, which are described in detail as follows:

[0271] like Figure 2 As shown, the device 200 for predicting the time series of building deformation point clouds includes:

[0272] The acquisition and alignment module 210 is used to perform coordinate system transformation and noise reduction on the acquired original point cloud data of the target building to obtain an aligned point cloud set; the original point cloud data includes point cloud data collected at multiple time periods and / or by multiple devices;

[0273] A difference analysis module 220 is used to perform local difference analysis on the aligned point cloud set to obtain a point cloud set in an abnormal area;

[0274] A construction module 230 is configured to extract topological features of a preset connection structure of the abnormal area based on multi-resolution features of a point cloud set of the abnormal area, and to construct a topological map of the abnormal area;

[0275] A feature extraction module 240 is used to perform stress constraints and material property constraints on the topology of the abnormal area to determine the latent space features of the topology of the abnormal area;

[0276] An interpolation and reasoning module 250 is used to perform interpolation and extrapolation processing on the latent space features based on the latent diffusion model to determine the predicted vector field of the latent space features and the suspected damage area;

[0277] Analysis module 260, for performing multi-scale regression analysis on the suspected damage area to obtain a local deformation prediction value and a local uncertainty distribution;

[0278] A fusion module 270 is used to fuse and evaluate the consistency of the three-dimensional deformation prediction field corresponding to the prediction vector field of the latent space features, the local deformation prediction value, and the local uncertainty distribution to obtain a fused global deformation prediction field;

[0279] The fusion analysis module 280 is used to analyze the fused global deformation prediction field to obtain a high-risk deformation distribution.

[0280] In one possible implementation, the construction module 230 is configured to extract multiple resolutions of the point cloud set of the abnormal region based on a multi-resolution encoder to obtain multi-resolution features of the point cloud set of the abnormal region; wherein the multi-resolution encoder is based on introducing a gradient vector into the multi-resolution encoder to perform downsampling and feature extraction on the multiple resolutions of the point cloud set of the abnormal region;

[0281] Extracting the topological features of the preset connection structure of the abnormal area based on the topological feature extraction module in the multi-resolution encoder to obtain the topological features of the preset connection structure of the abnormal area; wherein the preset connection structure includes beams, columns, trusses and facades of the target building;

[0282] Based on the topological features of the preset connection structure of the abnormal area, a topological map of the abnormal area is constructed.

[0283] In one possible implementation, the construction module 230 is configured to perform stress constraints on the diffusion process of the latent space field corresponding to the topological map of the abnormal region to obtain a first latent space feature;

[0284] Based on the material property database, material property constraints are added to each diffusion process of the latent space field corresponding to the topological map of the abnormal area to obtain the second latent space feature;

[0285] Based on the deformation amplitude and gradient, the deformation area of the latent space field corresponding to the topological map of the abnormal area is constrained to obtain the third latent space feature;

[0286] Based on the first latent space feature, the second latent space feature, and the third latent space feature, a latent space feature of a topological map of the abnormal region is determined.

[0287] In one possible implementation, the interpolation inference module 250 is configured to interpolate adjacent time segments in the latent space feature vector field based on the trained latent diffusion model and the stress and material property constraints in the latent space, and to extrapolate the latent space feature vector field to obtain a predicted deformation vector field of the latent space feature.

[0288] The deformation intensity of the topological features of the preset connection structure is calculated, and the suspected damage area is determined based on the preset deformation threshold and the predicted deformation vector field of the latent space features.

[0289] In one possible implementation, the analysis module 260 is configured to screen high-risk suspected areas based on the positioning information of a point set in the suspected damage area and a spatial filtering operator; the positioning information of the point set includes the coordinates of each point and the predicted deformation vector field of each point;

[0290] Based on a pre-constructed composite kernel function, multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain local deformation prediction values and local uncertainty distribution; the composite kernel function includes geometric correlation, material property similarity kernel function and topological coupling of preset connection structures.

[0291] In one possible implementation, the analysis module 260 is configured to perform a multi-scale regression analysis on each scale level in the high-risk suspected area based on a pre-built composite kernel function to obtain a first regression result;

[0292] Dynamically updating the preset weights in the composite kernel function to obtain the final weight of the composite kernel function; wherein the geometric correlation, the material property similarity kernel function, and the topological coupling of the preset connection structure in the composite kernel function are each provided with a corresponding preset weight;

[0293] Based on the final weight of the composite kernel function, a multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain the second regression result;

[0294] The first regression result and the second regression result are integrated to obtain the local deformation prediction value and the local uncertainty distribution.

[0295] In one possible implementation, the fusion module 270 is configured to map the prediction vector field of the latent space feature to a physical coordinate system based on the decoder to obtain a three-dimensional deformation prediction field of the latent space feature;

[0296] Construct a damage map based on the predicted vector field and suspected damage areas of latent space features;

[0297] Based on the 3D deformation prediction field and damage map, a global prediction field of latent space features is constructed;

[0298] Determining a global skeleton of the global prediction field; wherein a function of the global skeleton is determined by a double integral of a correlation weight of the target position and a first deformation difference, where the first deformation difference is a difference between the three-dimensional deformation prediction field and a preset average reference;

[0299] Construct the weights of local parameters and global skeleton, and fuse the local parameters and global skeleton to obtain the fused prediction field; where the local parameters include the local deformation prediction value and the local uncertainty distribution;

[0300] Based on the differences between the local parameters and the global skeleton in the high-risk suspected area, as well as the differences between the gradient of the fused prediction field and the gradient of the global skeleton, the fused prediction field is verified to obtain the fused global deformation prediction field and damage distribution.

[0301] In one possible implementation, the fusion analysis module 280 is configured to determine the deformation degree and uncertainty within the high-risk suspected area based on the fused global deformation prediction field and the high-risk suspected area;

[0302] The deformation degree and uncertainty in the high-risk suspected area and the high-risk suspected area are input into the weight update module of the latent diffusion model, and an iterative process is performed to obtain the updated weight of the latent diffusion model;

[0303] Based on the composite kernel function, multi-scale regression analysis is performed on the high-risk suspected areas to obtain the high-risk deformation distribution.

[0304] In one possible implementation, the difference analysis module 220 is configured to determine the strength of each point cloud in the aligned point cloud set based on a pre-built anomaly evaluation model;

[0305] Based on the intensities of all point clouds in the aligned point cloud set, the point cloud set of the abnormal area is determined.

[0306] An embodiment of the present invention further provides a prediction device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the method in the above method embodiment when executing the computer program.

[0307] In the above embodiments, the descriptions of each embodiment have their own focus. For parts not described or recorded in detail in one embodiment, please refer to the relevant descriptions of other embodiments. Unless otherwise specified or there is a logical conflict, the terms and / or descriptions between different embodiments are consistent and can be referenced to each other. The technical features of different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0308] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A method for predicting building deformation point cloud time series, characterized in that: include: Perform coordinate system transformation and noise reduction on the original point cloud data of the target building to obtain an aligned point cloud set; The original point cloud data includes point cloud data collected in multiple time periods and / or by multiple devices; Performing local difference analysis on the aligned point cloud set to obtain a point cloud set in an abnormal area; Extracting topological features of a preset connection structure of the abnormal area based on multi-resolution features of the point cloud set of the abnormal area, and constructing a topological map of the abnormal area; Performing stress constraints and material property constraints on the topological map of the abnormal region to determine the latent space characteristics of the topological map of the abnormal region; Based on the latent diffusion model, the latent space features are interpolated and extrapolated to determine the predicted vector field of the latent space features and the suspected damage area; Perform multi-scale regression analysis on the suspected damage area to obtain the local deformation prediction value and local uncertainty distribution; Mapping the predicted vector field of the latent space feature to a physical coordinate system based on the decoder to obtain a three-dimensional deformation prediction field of the latent space feature; constructing a damage map based on the predicted vector field of the latent space features and the suspected damage area; constructing a global prediction field of the latent space features based on the three-dimensional deformation prediction field and the damage map; Determining a global skeleton of the global prediction field; wherein a function of the global skeleton is determined by a double integral of a correlation weight of a target position and a first deformation difference, wherein the first deformation difference is a difference between the three-dimensional deformation prediction field and a preset average reference; Constructing weights of local parameters and the global skeleton, and fusing the local parameters with the global skeleton to obtain a fused prediction field; wherein the local parameters include a local deformation prediction value and a local uncertainty distribution; Screening high-risk suspected areas based on the positioning information of the point set in the suspected damage area and a spatial filtering operator; the positioning information of the point set includes the coordinates of each point and the predicted vector field of each point; Based on the difference between the local parameters in the high-risk suspected area and the global skeleton, and the difference between the gradient of the fused prediction field and the gradient of the global skeleton, the fused prediction field is verified to obtain a fused global deformation prediction field and damage distribution; The fused global deformation prediction field is analyzed to obtain a high-risk deformation distribution.

2. The method for predicting building deformation point cloud time series according to claim 1, characterized in that: The step of extracting the topological features of the preset connection structure of the abnormal area based on the multi-resolution features of the point cloud set of the abnormal area and constructing the topological map of the abnormal area includes: Extracting multiple resolutions of the point cloud set of the abnormal area based on a multi-resolution encoder to obtain multi-resolution features of the point cloud set of the abnormal area; wherein the multi-resolution encoder downsamples and extracts features of the multi-resolution of the point cloud set of the abnormal area based on introducing a gradient vector in the multi-resolution encoder; Extracting the topological features of the preset connection structure of the abnormal area based on the topological feature extraction module in the multi-resolution encoder to obtain the topological features of the preset connection structure of the abnormal area; wherein the preset connection structure includes beams, columns, trusses and facades of the target building; A topological map of the abnormal area is constructed based on the topological features of the preset connection structure of the abnormal area.

3. The method for predicting building deformation point cloud time series according to claim 1, characterized in that: The performing stress constraints and material property constraints on the topological map of the abnormal region to determine the latent space characteristics of the topological map of the abnormal region includes: Performing stress constraints on the diffusion process of the latent space field corresponding to the topological map of the abnormal area to obtain a first latent space feature; Based on the material property database, material property constraints are added to each diffusion process of the latent space field corresponding to the topological map of the abnormal area to obtain a second latent space feature; constraining the deformation region of the latent space field corresponding to the topological map of the abnormal region based on the deformation amplitude and gradient to obtain a third latent space feature; Based on the first latent space feature, the second latent space feature, and the third latent space feature, a latent space feature of the topological map of the abnormal region is determined.

4. The method for predicting building deformation point cloud time series according to claim 1, characterized in that: The latent diffusion model is based on which latent space features are interpolated and extrapolated to determine the predicted vector field of the latent space features and the suspected damage area, including: Based on the trained latent diffusion model and the stress constraints and material property constraints in the latent space, adjacent time segments in the vector field of the latent space features are interpolated and the vector field of the latent space features is extrapolated to obtain a predicted vector field of the latent space features; The deformation intensity of the topological feature of the preset connection structure is calculated, and the suspected damage area is determined based on a preset deformation threshold and the predicted vector field of the latent space feature.

5. The method for predicting building deformation point cloud time series according to claim 1, characterized in that: The multi-scale regression analysis of the suspected damage area is performed to obtain the local deformation prediction value and the local uncertainty distribution, including: Based on a pre-constructed composite kernel function, multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain local deformation prediction values and local uncertainty distribution; the composite kernel function includes geometric correlation, material property similarity kernel function and topological coupling of the preset connection structure.

6. The method for predicting building deformation point cloud time series according to claim 5, characterized in that: Based on the pre-constructed composite kernel function, a multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain a local deformation prediction value and a local uncertainty distribution, including: Based on a pre-constructed composite kernel function, a multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain a first regression result; Dynamically updating the preset weights in the composite kernel function to obtain a final weight of the composite kernel function; wherein the geometric correlation, the material property similarity kernel function, and the topological coupling of the preset connection structure in the composite kernel function are each corresponding to a preset weight; Based on the final weight of the composite kernel function, a multi-scale regression analysis is performed on each scale level in the high-risk suspected area to obtain a second regression result; The first regression result and the second regression result are integrated to obtain a local deformation prediction value and a local uncertainty distribution.

7. The method for predicting building deformation point cloud time series according to claim 5 or 6, characterized in that: The analyzing the fused global deformation prediction field to obtain a high-risk deformation distribution includes: Determining the deformation degree and uncertainty within the high-risk suspected area based on the fused global deformation prediction field and the high-risk suspected area; Inputting the deformation degree and uncertainty within the high-risk suspected area and the high-risk suspected area into the weight updating module of the latent diffusion model, performing iterative processing, and obtaining the updated weight of the latent diffusion model; Based on the composite kernel function, a multi-scale regression analysis is performed on the high-risk suspected area to obtain the high-risk deformation distribution.

8. The method for predicting building deformation point cloud time series according to claim 1, characterized in that: The performing of local difference analysis on the aligned point cloud set to obtain a point cloud set in an abnormal area includes: Determining the intensity of each point cloud in the aligned point cloud set based on a pre-built anomaly evaluation model; wherein the anomaly evaluation model is constructed based on the difference between the grayscale or intensity information of each point cloud and a preset average intensity value; The point cloud set of the abnormal area is determined based on the intensities of all point clouds in the aligned point cloud set.

9. A prediction device, characterized in that The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 8 is implemented.

Images