Three-dimensional monitoring data time sequence difference compensation and dynamic threshold adaptive adjustment method

Abstract

The invention discloses a three-dimensional monitoring data time sequence difference compensation and dynamic threshold adaptive adjustment method. The method comprises the following steps: collecting three-dimensional point cloud data and environment sensor data; environment feature extraction is carried out by adopting regional adaptive partition PCA, and the problem of spatial heterogeneity is solved; by utilizing a physical information enhanced environment field reconstruction technology, the environment field distribution of the whole monitoring area is accurately reconstructed from limited environment sensor data; constructing a cognitive physical-data dual-drive compensation model, and embedding a physical model into a data drive architecture; calculating expected displacement caused by the environment based on the environment-structure response model and deducting the expected displacement from the actually measured displacement; and dynamically adjusting the change detection threshold. According to the method, the problem of'false change 'caused by environmental factors is solved, and the accuracy and reliability of three-dimensional monitoring data are improved.

Description

Technical Field

[0001] The invention belongs to the field of structural health monitoring, and in particular to a method for compensating for time series differences in three-dimensional monitoring data and adaptively adjusting dynamic thresholds. Background Art

[0002] Three-dimensional monitoring technology has important application value in the fields of construction engineering, ancient building protection, infrastructure management, etc. The point cloud data obtained by high-precision three-dimensional laser scanners can accurately record and monitor changes in the geometric form of the structure, providing a key basis for structural health assessment and safety warning. However, outdoor building structures are exposed to the natural environment for a long time. Environmental factors such as temperature, humidity, and light will cause small displacements such as thermal expansion and contraction of materials and humidity deformation. These "pseudo-changes" caused by the environment are often mixed with the real deformation of the structure, which seriously affects the accuracy and reliability of change detection. Therefore, it is of great theoretical and practical significance to study the compensation mechanism for the impact of environmental factors on three-dimensional monitoring data.

[0003] At present, a series of technical methods have been formed in the field of three-dimensional point cloud change detection. Common point cloud-based change detection methods include point-to-point comparison method, point-to-plane distance method and feature comparison method. In order to reduce the impact of environmental factors, researchers have proposed a specific time period sampling comparison method, that is, collecting data in time periods with similar environmental conditions for comparison; some scholars have also introduced statistical filtering technology to screen changes by setting fixed thresholds or adaptive thresholds based on statistical characteristics; in terms of point cloud registration, the iterative closest point (ICP) algorithm and its improved algorithm are widely used for accurate registration of point clouds at different times. For compensation of environmental factors, some studies have begun to try to establish a simple linear regression model to describe the relationship between a single environmental parameter and structural displacement.

[0004] However, existing technologies still face several key technical bottlenecks in practical applications. First, traditional environmental compensation methods assume that there is a linear relationship between environmental factors and structural responses, and are unable to handle nonlinear thermal expansion of materials and multi-factor coupling effects, resulting in poor compensation effects under extreme environmental conditions. Second, existing methods ignore the spatial heterogeneity of structures. Due to differences in materials, orientations, and stress conditions, different parts of the same structure also have significant differences in their response characteristics to environmental factors. Using a unified model for compensation will produce regional errors. Third, environmental sensors are arranged in discrete positions. How to accurately reconstruct the environmental parameter field distribution of the entire monitoring structure surface from limited environmental monitoring points and achieve accurate environmental compensation for each point cloud point is a difficult problem that has not been well solved by existing technologies. These technical problems lead to a high false positive rate in existing systems under complex structures and changing environmental conditions, affecting monitoring reliability. Summary of the invention

[0005] The purpose of the invention is to provide a method for compensating for time series differences in three-dimensional monitoring data and adaptively adjusting dynamic thresholds, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution is a method for compensating for time series differences in three-dimensional monitoring data and adaptively adjusting dynamic thresholds, including:

[0007] Collect and preprocess point cloud data to obtain registration point cloud data;

[0008] Based on the registered point cloud data and pre-stored material parameter data, PCA environmental features are extracted by region to obtain regional environmental features;

[0009] Using regional environmental features and pre-stored environmental sensor data, perform physical information enhanced environmental field reconstruction to obtain the final environmental field and build a cognitive physical-data dual-driven compensation model based on it to obtain a training model;

[0010] When in use, the dynamic threshold adaptive adjustment and timing difference compensation are performed through the training model, and the environmental displacement field, the compensated displacement field and the adaptive threshold are calculated to obtain the change detection result.

[0011] Beneficial effects: The present invention enables more accurate modeling of the complex coupling relationship of environmental factors, realizes the fine reconstruction of the local environmental field, and significantly reduces the regional error caused by the overall unified model; at the same time, it not only solves the problem of insufficient sensor coverage, but also provides an accurate environmental compensation basis for each point cloud point; it solves the "pseudo-change" problem caused by environmental factors, and improves the accuracy and reliability of three-dimensional monitoring data. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A flowchart of the steps of a method for compensating for timing differences in three-dimensional monitoring data and adaptively adjusting dynamic thresholds provided in an embodiment of the present application.

[0013] Figure 2 A flowchart of the steps for extracting PCA environmental features by region provided in an embodiment of the present application.

[0014] Figure 3 A flowchart of the steps for performing response similarity evaluation provided in an embodiment of the present application.

[0015] Figure 4 A flowchart of the steps of applying time-series weighted PCA feature extraction provided in an embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0017] It should be noted that in order to clearly show the steps of this application, serial numbers are marked for each step in the specification. These serial numbers are only used for the convenience of explanation and do not limit the order of execution of the steps. In actual operation, according to the technical requirements of the specific implementation scenario, the steps can be executed in a different order than that shown in the specification, and in some cases, parallel processing between steps can be achieved.

[0018] like Figure 1 As shown, the method for compensating for time series differences in three-dimensional monitoring data and adaptively adjusting dynamic thresholds includes the following steps:

[0019] S1, collect point cloud data and preprocess it to obtain registration point cloud data;

[0020] Specifically, a 3D laser scanner is used to obtain point cloud data on the surface of the target structure, including digital representation of information such as shape and position. At the same time, environmental information such as temperature and humidity can also be collected, and the specific time of collection can be recorded.

[0021] S2, based on the registered point cloud data and the pre-stored material parameter data, extracting PCA environmental features by region to obtain regional environmental features;

[0022] Specifically, material parameter data includes physical properties such as thermal expansion coefficient and elastic modulus. The point cloud data is divided into several regions. In each region, PCA (principal component analysis) is used to extract key characteristics related to environmental changes. Finally, each region will get a feature data set that describes how this region responds to environmental factors.

[0023] S3, using regional environmental features and pre-stored environmental sensor data, performing physical information enhanced environmental field reconstruction to obtain a final environmental field;

[0024] Specifically, environmental sensor data is real-time environmental data recorded by sensors, such as temperature changes, humidity changes, etc. Using the laws of physics (such as heat conduction, air flow, etc.), constraints are added to the model to ensure that the reconstructed environmental field conforms to physical phenomena, rather than just based on sensor data.

[0025] S4. Based on the final environmental field, a cognitive physical-data dual-driven compensation model is constructed to obtain a training model;

[0026] Specifically, the cognitive physics-data dual-driven compensation model combines physical laws and data-driven methods for compensation and optimization. That is, the basic framework is established through physical theory, while the model is improved through data analysis and learning to make it more accurate and practical.

[0027] S5. When in use, the dynamic threshold adaptive adjustment and timing difference compensation are performed through the training model, and the environmental displacement field, the compensated displacement field and the adaptive threshold are calculated to obtain the change detection result.

[0028] Specifically, when in use, the threshold is dynamically adjusted according to the current situation, like an adjustable standard line, so as to adapt to different environmental changes. Then, the time difference is taken into account and compensation is performed, that is, when comparing data in different time periods, the error caused by time changes is eliminated to ensure the accuracy of the results.

[0029] This embodiment can accurately identify and monitor dynamic changes in the environment and effectively reduce data dimensions, thereby improving computing speed and analysis efficiency; ensuring that the model not only follows the theoretical basis, but can also adaptively learn from data to improve accuracy and enhance practicality; at the same time, it not only improves the overall reliability of the model, but also optimizes the accuracy of the analysis, achieving high-precision, strong robustness and high-efficiency data monitoring and change detection.

[0030] According to one aspect of the present application, the pre-processing step comprises:

[0031] S11. Multi-source data acquisition. Use a 3D laser scanner to collect point cloud data of the target structure. t , where t represents the collection time point; deploy an environmental sensor array to collect environmental parameter data E t , including environmental factors such as temperature, humidity, and light, and synchronously record the acquisition time t; obtain the material parameter data M of the structure, including physical properties such as thermal expansion coefficient and elastic modulus, and record the scanning position information L and scanning parameters S for subsequent point cloud registration and data correction.

[0032] S12, point cloud data preprocessing. Read the original point cloud data, perform noise reduction processing, and obtain a noise-reduced point cloud; downsample the noise-reduced point cloud to control the uniformity of point density and obtain a uniform point cloud; extract the geometric features of the uniform point cloud, including normal vectors, curvature, etc., to generate a feature point cloud; based on the feature point cloud and the scanning position information L, register the point clouds acquired at different times to obtain a registered point cloud.

[0033] S13, environmental data preprocessing. Read environmental parameter data E t, detect outliers and obtain filtered environmental data; perform time series smoothing on the filtered environmental data to obtain smoothed environmental data; calculate the change in environmental parameters at adjacent time points to generate environmental change data; associate the smoothed environmental data with the sensor position coordinates C to establish an environment-space mapping.

[0034] S14, structural segmentation and material labeling. Based on the feature point cloud and material parameter data M, the point cloud is segmented into different material regions to obtain segmented point clouds; corresponding material parameters are assigned to each region in the segmented point cloud to generate material labeling point clouds; the geometric properties (area, volume, etc.) of each segmented region are calculated to obtain regional characteristic data R; the segmented region adjacency relationship G is established to represent the spatial connection relationship between different material regions.

[0035] S15, initial change detection. Select the registered point cloud at the reference time point as a reference; compare the registered point cloud at the current time point t with the reference point cloud, calculate the point-to-point distance or point-to-surface distance, and generate an original change map; filter the original change map (for example, by applying a traditional threshold method) to obtain an initial change detection result.

[0036] like Figure 2 As shown, according to one aspect of the present application, the step of extracting PCA environmental features by region and obtaining regional environmental features includes:

[0037] S21, performing initial division of the structural response region based on the registered point cloud data and the material parameter data to obtain an initial partition;

[0038] S22, evaluating the response similarity of the initial partitions, and obtaining the optimized partitions by constructing a response correlation matrix and performing spectral clustering analysis;

[0039] S23, applying time-series weighted PCA feature extraction to the optimized partitions, and obtaining a PCA model by constructing a time-attenuated weight function and a weighted covariance matrix;

[0040] S24, based on the PCA model, perform robust environmental feature extraction to obtain an environmental feature vector;

[0041] S25. Based on the environmental feature vector and the PCA model, a nonlinear kernel function mapping is constructed to obtain regional environmental characteristics.

[0042] Specifically, based on the material annotation point cloud and material parameter data M, the potential environmental sensitivity of each point is calculated to obtain a sensitivity map; according to the sensitivity map and regional characteristic data, the region growing algorithm is used to perform initial partitioning to obtain the initial partitioning result; for each initial partition, the environmental response statistical characteristics are calculated to generate a partition characteristic matrix; based on the partition characteristic matrix and the adjacency relationship of the segmented area, the boundaries of adjacent partitions are adjusted to obtain adjusted partitions.

[0043] Environmental change data of multiple historical time points and the corresponding registered point clouds are selected. For each adjusted partition, the correlation matrix between point displacement and environmental change is calculated to obtain the response correlation matrix. Based on the response correlation matrix, the spectral clustering algorithm is used to evaluate the response similarity of points within the partition to obtain the similarity index. According to the similarity index, the partition boundaries are optimized, highly similar adjacent partitions are merged, and partitions with high response heterogeneity are split to generate optimized partitions.

[0044] For each optimized partition, an environment-displacement data matrix is ​​constructed, which contains the environmental parameters and corresponding displacements at historical time points; the time decay weight function W(t-τ) = e -λ(t-τ) , calculate the time weight vector; where t is the current time point, τ is the historical time point, λ is the time decay factor, and e is the base of the natural logarithm. Construct a weighted covariance matrix based on the time weight vector to obtain the weighted covariance; perform eigendecomposition on the weighted covariance, extract the principal component vector and eigenvalue, and generate a PCA model.

[0045] The M-estimator is applied to the historical environment-displacement data of each partition to identify and reduce the influence of outliers and obtain robust data; based on the robust data and the PCA model, the projection of environmental characteristics in the principal component space is calculated to obtain the environmental feature vector; a feature reconstruction mapping is constructed to project the environmental feature vector back to the original space to obtain the reconstructed environmental data; the reconstruction error between the reconstructed environmental data and the original environmental data is calculated to generate a feature quality index.

[0046] Based on the environmental feature vector and PCA model, a nonlinear kernel function mapping is constructed for each partition to obtain the kernel PCA model; according to the material characteristics of the partition, an appropriate kernel function (RBF kernel, polynomial kernel, etc.) is selected for each kernel PCA model; the influence weight of environmental factors is calculated, and the importance of each environmental parameter in the model is adjusted to obtain the parameter weight; the kernel PCA model, parameter weight and optimized partition are integrated to construct a regional environmental feature library.

[0047] like Figure 3 As shown, according to one aspect of the present application, the steps of evaluating response similarity and obtaining optimized partitions include:

[0048] Extract environmental change data and registration point cloud data at historical time points to establish an environment-response mapping table;

[0049] Based on the environment-response mapping table, multivariate regression analysis is performed on the point cloud within the initial partition to obtain the sensitivity coefficient matrix;

[0050] Based on the sensitivity coefficient matrix, the point-to-point response pattern similarity is calculated and the response correlation matrix is ​​constructed;

[0051] Apply Laplace matrix transformation and eigenvalue decomposition to the response correlation matrix and perform spectral clustering analysis to obtain the subpartition consistency matrix;

[0052] Based on the sub-partition consistency matrix, the partition consistency index and the similarity between partitions are calculated, and the partition merging or splitting operation is performed to obtain the optimized partition.

[0053] Specifically, read the historical environmental change data ΔE within a predetermined time window τ (τ = t-1, t-2, ..., tn), organize it into a time series matrix, and obtain the environmental change time series matrix ΔE t ; Extract the registration point cloud of the corresponding time point, calculate the point displacement vector between each pair of adjacent time points, and obtain the displacement vector field. Perform time series connection on the displacement vector of each point to form a displacement time series and obtain the point displacement time series matrix. Establish the corresponding relationship table between the environment change time series matrix and the point displacement time series matrix to generate the environment-response mapping table.

[0054] For each point cloud in the adjustment partition, extract the data in its environment-response mapping table, build a multivariate linear regression model, use the environment change time series matrix as the independent variable, and the point displacement time series matrix as the dependent variable to obtain a linear response model. Calculate the residual sum of squares to evaluate the degree of fit of the linear model and generate a linear fitting error; apply a nonlinear kernel regression model, also using the environment change time series matrix and the point displacement time series matrix as input to obtain a nonlinear response model; calculate the residual sum of squares of the nonlinear model to generate a nonlinear fitting error.

[0055] Compare the linear fitting error and nonlinear fitting error of each point to determine the best fitting model type and record it as the optimal model type; based on the optimal model type, extract the environmental factor sensitivity coefficient of each point and generate a sensitivity coefficient matrix; calculate the point-to-point response pattern similarity within the partition and construct the point-to-point similarity matrix PS z :PS z (i,j) = exp(-||S z coef (i) - S z coef (j)|| 2 / σ 2 );where S z coef is the sensitivity coefficient matrix, i and j are point indexes, and σ is the scale parameter. Based on the point pair similarity matrix PS z , construct a weighted adjacency graph and obtain the response correlation matrix R z : R z (i,j) = PS z(i,j) · W(d(i,j)), where W is the spatial distance weighting function and d(i,j) is the Euclidean distance between point i and point j.

[0056] Apply the Laplace matrix transform to each response correlation matrix and calculate the normalized Laplace matrix L z norm :L z norm = D z -1 / 2 · (D z -R z ) · D z -1 / 2 , where D z is the degree matrix, and the diagonal elements are the response correlation matrix R z The sum of each row. Calculate the eigenvalues ​​and eigenvectors of the normalized Laplace matrix, select the eigenvectors corresponding to the k smallest non-zero eigenvalues, and form an eigenvector matrix; apply the K-means clustering algorithm to the row vectors of the eigenvector matrix, divide the points in the partition into several sub-partitions, and obtain the spectral clustering results; calculate the response consistency index of the points in each sub-partition, and construct the sub-partition consistency matrix.

[0057] Based on the sub-partition consistency matrix SC z , calculate the internal response consistency score of each adjustment partition and generate the partition consistency index IC z :IC z = 1 - ∑ i (var(SC z (i)) / |SC z (i)|) / k; where var is the variance function, |SC z (i) | is the sub-partition consistency matrix SC z The size of the i-th sub-partition in , k is the number of sub-partitions. Calculate the similarity of response patterns between adjacent partitions and construct the inter-partition similarity matrix ZS: ZS(i,j) = cos(S z coef i , S z coef j ), where cos is the cosine similarity function. Based on the partition consistency index IC z And the partition similarity matrix ZS, calculate the comprehensive similarity index, and get the similarity index S z :S z = α · IC z + (1-α) max jZS(z,j), where α is a balancing factor used to adjust the weights of internal consistency and external similarity. z , set the merge threshold θ merge and the splitting threshold θ split , perform the following optimization operations: For S z < θ split The partition of is further subdivided by recursive bisection to obtain subdivided partitions; for adjacent partitions i and j, if ZS(i,j) > θ merge , then merge the two partitions to obtain a merged partition; integrate the subdivided partitions and the merged partitions to generate the final optimized partition.

[0058] like Figure 4 As shown, according to one aspect of the present application, the step of applying time-series weighted PCA feature extraction to obtain a PCA model includes:

[0059] Based on the optimized partition and environment-response mapping table, the environment-displacement joint matrix is ​​constructed;

[0060] Based on the preset time decay coefficient, the exponential decay function is applied to calculate the weight of the historical time point to obtain the normalized time weight;

[0061] Based on the normalized time weights, the weighted mean vector and weighted outer product of the environment-displacement joint matrix are calculated to construct the regularized covariance matrix;

[0062] Perform eigenvalue decomposition on the regularized covariance matrix, sort by contribution rate and select the main eigenvectors to obtain the principal component matrix;

[0063] Based on the principal component matrix, the projection transformation matrix and the reconstruction transformation matrix are constructed, and combined with the weighted mean vector to generate the PCA model.

[0064] Specifically, for each point in the optimized partition, historical data is extracted from the environment-response mapping table; the environmental parameters are organized in chronological order to construct a partition environment matrix, where rows represent time points and columns represent different environmental parameters; the average displacement vector of the points in the partition at each time point is calculated to construct a partition displacement matrix, which also uses time as the row index; the partition environment matrix and the partition displacement matrix are concatenated to form an environment-displacement joint matrix J. z :J z = [E z T | D z T ], where E z T is the partition environment matrix, D z T is the partition displacement matrix.

[0065] Set the time decay coefficient λ, which controls the rate at which the influence of historical data decays over time; calculate the time interval Δt = t – τ from each historical time point τ to the current time t; apply the exponential decay function to calculate the weight of each time point and obtain the time weight vector W t :W t (τ) = exp(-λ·(t-τ)); for the time weight vector W t Perform normalization to ensure that the sum of weights is 1, and obtain the normalized time weight W t norm :W t norm (τ) = W t (τ) / ∑ i W t (i).

[0066] Using the normalized temporal weights, calculate the weighted mean vector μ of the joint environment-displacement matrix w :μ w = ∑ τ W t norm (τ) · J z (τ); for the environment-displacement joint matrix J z Subtract the weighted mean vector μ from each row of w , get the centralization matrix; based on the normalized time weight, calculate the weighted outer product of the centralization matrix and construct the weighted covariance matrix C z w :C z w = ∑ τ W t norm (τ) · J z c (τ) T · J z c (τ), where J z c is the centralization matrix, T Represents transpose. Regularize the weighted covariance matrix to increase numerical stability and obtain the regularized covariance matrix C z reg :C z reg = C z w + ε·I, where ε is a small positive number and I is the identity matrix.

[0067] Perform eigenvalue decomposition on the regularized covariance matrix to obtain eigenvalue vectors and eigenvector matrices; arrange the eigenvalues ​​in descending order and adjust the order of the eigenvectors accordingly to obtain sorted eigenvalues ​​and sorted eigenvectors; calculate the cumulative variance contribution rate and determine the number of principal components k to be retained to ensure that the cumulative contribution rate reaches the preset threshold η (such as 95%): k = min{j |∑ i =1 j λ s i / ∑ i λ s i ≥ η}, where λ s is the sorting eigenvalue; select the first k columns from the sorting eigenvector to form the principal component matrix; calculate the variance contribution rate corresponding to each principal component and generate the contribution rate vector CR z :CR z (i) = λ s i / ∑ j λ s j , i = 1,2,...,k.

[0068] Based on the principal component matrix, a data projection transformation matrix is ​​constructed to realize the mapping from the original space to the principal component space, and the projection matrix is ​​obtained; a reconstruction transformation matrix from the principal component space to the original space is established to obtain the reconstruction matrix; the environment-displacement joint matrix is ​​projected to the principal component space to obtain the principal component coefficient matrix PC z : PC z = J z c · P z , where P z is the principal component matrix, J z c is the centralization matrix; the weighted mean vector, principal component matrix, contribution rate vector, projection matrix and reconstruction matrix are integrated to build a complete PCA model.

[0069] According to one aspect of the present application, the steps of extracting robust environmental features to obtain an environmental feature vector include:

[0070] Based on the environment-displacement joint matrix, the Huber loss function is used as a robust M-estimator to obtain the weight function;

[0071] The weight function is used to calculate the weight of data points, identify outliers and make corrections to obtain a robust data matrix;

[0072] Project the robust data matrix through the projection matrix of the PCA model to generate a robust coefficient matrix;

[0073] The corresponding relationship between environmental parameters and principal components is analyzed based on the robust coefficient matrix, key environmental features are extracted, and environmental feature vectors are obtained.

[0074] Specifically, based on the environment-displacement joint matrix, the Mahalanobis distance of each data point to its temporal neighbor is calculated to obtain a distance matrix; a robust loss function ρ(x) is designed, such as the Huber loss or the Tukey double-weighted loss function: ρ Huber (x) = {x 2 / 2, if |x| ≤ c; c·|x| - c 2 / 2, if |x| > c}, where c is a tuning parameter that controls the sensitivity to outliers. Based on the outliers in the historical data, optimize the M-estimator parameters, such as the parameter c of the Huber loss, to obtain the optimized M-estimator. Design the corresponding weight function w(x) = ρ'(x) / x to weight the data points and obtain the weight function. Where x is the data point and ρ'(x) is the derivative of the loss function ρ(x) with respect to x.

[0075] For each data point in the environment-displacement joint matrix, the distance matrix and weight function are applied to calculate the weight to obtain the data point weight matrix; the outlier recognition threshold is set, and the data points with weights lower than the threshold are identified as outliers to generate an outlier labeling matrix; the identified outliers are corrected by local smoothing or interpolation methods to obtain corrected data points; the original data points and the corrected data points are merged according to the weights to form a robust data matrix J z rob :J z rob = W z point Θ J z + (1- W z point ) Θ J z corr , where Θ represents the element-wise multiplication operation, W z point is the data point weight matrix, J z is the environment-displacement joint matrix, J z corr To correct the data points.

[0076] The robust principal component coefficients are calculated using the robust data matrix and the projection matrix in the PCA model to obtain a robust coefficient matrix; the statistical characteristics of the robust coefficient matrix, such as the mean, variance, distribution shape, etc., are analyzed to generate coefficient statistical characteristics; the correspondence between environmental parameters and principal components is identified, the physical meaning captured by each principal component is determined, and a principal component interpretation table is generated; based on the principal component interpretation table, the main environmental-related features are extracted from the robust coefficient matrix to form an environmental feature vector.

[0077] Use the environmental feature vector and the reconstruction matrix in the PCA model to reconstruct the environmental parameters and obtain the reconstructed environmental data; calculate the mean square error between the reconstructed environmental data and the original environmental data to generate the reconstruction error; use the leave-one-out cross-validation method to evaluate the generalization ability of the feature extraction model and obtain the cross-validation error; combine the reconstruction error and the cross-validation error to calculate the comprehensive score of the feature quality and generate the feature quality index Q z :Q z = β · (1 - norm(MSE z )) + (1-β) · (1 -norm(CVE z )), where norm is the normalization function, β is the trade-off parameter, and MSE z is the reconstruction error, CVE z is the cross validation error.

[0078] Based on the statistical characteristics of the robust data matrix, an adaptive anomaly detection model is constructed to obtain an anomaly detector; the spatiotemporal distribution characteristics of anomaly patterns in historical data are analyzed to build an anomaly pattern library; an adaptive threshold adjustment strategy is designed to dynamically adjust the threshold according to the current environmental conditions and historical anomaly patterns to obtain an adaptive threshold function AT z :AT z (x) = θ b ase ·(1 + γ · sim(x, AM z )), where θ b ase is the basic threshold, γ is the adjustment parameter, sim is the similarity function, AM z It is an abnormal pattern library; the anomaly detector and adaptive threshold function are integrated to form a complete robust feature extraction system.

[0079] In one embodiment of the present application, robust environmental features are extracted. The Mahalanobis distance is calculated for the environment-displacement joint matrix J to identify outliers: Mahalanobis distance formula: d(J(τ)) = sqrt((J(τ)-μ) T ·C -1(J(τ)-μ)), for example: d(J(t-5)) = 3.82 > c(=1.345); w(J(t-5)) = 1.345 / 3.82 = 0.352, 4 outliers are identified, and the robust data matrix J is constructed z rob :J z rob (t-5) = 0.352×J(t-5) + 0.648×J corr (t-5), where J corr (t-5) is calculated by weighted average of the five neighboring points; the robust data matrix is ​​projected through the PCA model: PC z = (J z rob - μ w )×Projection Matrix= (J z rob - μ w )×P T , and obtain the robust coefficient matrix, where μ w is the weighted mean vector.

[0080] Select radial basis function (RBF) to construct nonlinear kernel function mapping: k(x,y) = exp(-γ||xy|| 2 ), where γ = 0.1 is the kernel parameter. Example of calculating the kernel matrix K: K(1,2) = exp(-0.1×||PC z (1)-PC z (2)|| 2 ); Apply kernel PCA to calculate the new feature projection: α = kernel principal component feature vector; kernel space projection = K×α, integrate to get regional environmental features: regional environmental features (region 1A) = {v env , PCA model , kernel function parameter, α}, PCA model The PCA model is used. Through the above steps, the environmental feature extraction of region 1A is completed. The same process is repeated for the other 6 partitions to obtain a complete set of regional environmental features.

[0081] According to one aspect of the present application, the steps of performing physical information enhanced environment field reconstruction to obtain the final environment field include:

[0082] S31. Based on regional environmental characteristics, registration point cloud data and material parameter data, a three-dimensional computational domain geometric model and physical boundary conditions are constructed to obtain a computational domain model and boundary condition function;

[0083] S32, based on the computational domain model, calculating the spatial gradient of the environmental parameter and generating a multi-resolution grid; wherein a fine grid is used in an area where the spatial gradient of the environmental parameter exceeds a preset threshold;

[0084] S33. For the temperature field, humidity field and light field, a heterogeneous physical model is constructed in combination with material parameter data, including a heat conduction equation, a convection-diffusion equation and a radiation transfer equation, to obtain an environmental field physical model set;

[0085] S34, fusing the environmental field physical model set with the environmental sensor data through the observation operator and the ensemble Kalman filter, performing data assimilation, and obtaining the posterior environmental field;

[0086] S35. Calculate the theoretical structural response based on the posterior environmental field, compare it with the measured displacement, and adjust the environmental field parameters through feedback optimization to obtain the final environmental field.

[0087] Specifically, based on the registration point cloud and the segmentation point cloud, a three-dimensional computational domain geometric model is constructed to obtain the computational domain model; the boundary surface of the computational domain model is identified to generate the boundary condition area; according to the location and environmental characteristics of the boundary condition area, the physical boundary conditions (such as temperature and humidity boundary values) are set to obtain the boundary condition function; according to the material annotation point cloud, material physical parameters are assigned to different areas in the computational domain to generate regional physical parameters.

[0088] Based on the computational domain model, an initial uniform grid is constructed to obtain an initial grid; the spatial gradient of environmental parameters is calculated, the area with larger gradient is identified, and a gradient map is generated; the grid refinement index is constructed by combining the gradient map and the sensitivity map to obtain a refinement index; according to the refinement index, the initial grid is adaptively refined to generate a multi-resolution grid; a fine grid is used in the area where the spatial gradient of the environmental parameters is larger than the preset threshold.

[0089] A heterogeneous heat conduction equation is constructed for the temperature field, and the position-dependent heat conduction coefficient is set based on the regional physical parameters to obtain the temperature field model. A convection-diffusion equation considering air flow is constructed for the humidity field to obtain the humidity field model. A model combining radiation transmission and ray tracing is constructed for the illumination field to obtain the illumination field model. The temperature field model, humidity field model and illumination field model are integrated to construct an environmental field physical model set.

[0090] Construct an observation operator to associate the state variables of the environmental field physical model set with the smoothed environmental data; initialize the model state variables, use the physical model for forward calculation, and obtain the initial prediction field; calculate the difference between the initial prediction field and the smoothed environmental data to generate the observation residual; apply the ensemble Kalman filter algorithm, combine the observation residual to update the model state, and obtain the posterior environmental field.

[0091] Based on the material parameter data and the posterior environmental field, the theoretical structural response is calculated to obtain the theoretical response; the displacement difference between the theoretical response and the actual measurement is compared to generate the response residual; based on the response residual, the environmental field parameters are adjusted, and the posterior environmental field is updated using the feedback optimization algorithm; global consistency constraints are applied to ensure that the updated posterior environmental field conforms to the laws of physics to obtain the final environmental field.

[0092] According to one aspect of the present application, the step of obtaining an environmental field physical model set includes:

[0093] Based on the material parameter data, the inhomogeneous heat conduction equation is established through the position-dependent heat conductivity field and heat source function to obtain the temperature field model;

[0094] Based on the material parameter data and pre-stored spatial geometric characteristics, the humidity convection-diffusion equation is established through the humidity diffusion coefficient field and the air flow field to obtain the humidity field model;

[0095] Based on the optical characteristic parameters in the material parameter data, the radiation transfer equation is constructed through the bidirectional reflection distribution function to obtain the illumination field model;

[0096] Based on multi-resolution grids, design spatial and temporal discrete formats suitable for the characteristics of each physical process and build a numerical solution framework;

[0097] Analyze the coupling relationship between physical fields, construct coupling solution strategies and calculation scheduling algorithms, integrate temperature field, humidity field and light field models, and obtain a set of environmental field physical models.

[0098] Specifically, the material annotation point cloud and material parameter data are read, the heat conduction related parameters of different materials are extracted, and the heat conduction parameter mapping is constructed; based on the material distribution in the regional physical parameters, the position-dependent heat conduction coefficient field is constructed to obtain the heat conduction coefficient field κ t :κ t (x) = H_κ(material(x)), where material(x) is the material type at position x, and H_κ is the heat conduction parameter mapping; design the heat source term function s(x,t), consider external heat sources (such as sunlight radiation) and internal heat sources (such as equipment heat generation), and obtain the heat source function S t ; Construct a complete heterogeneous heat conduction equation, define the temperature field evolution law, and form a temperature field model T_m: ΨT / Ψt - ▽·(κ t (x)▽T) = S t (x,t) in Ω; T = g t (x,t) on ΨΩ D κ t (x)ΨT / Ψn = h t (x,t) on ΨΩN , where g t and h t are the Dirichlet and Neumann boundary condition functions, Ω is the computational domain model, Ψ is the symbol of the partial derivative, T is the temperature field variable, ▽ is the gradient operator, S t (x, t) is the heat source function, which represents the external or internal heat source at position x and time t; ΨΩ D is the Dirichlet boundary region, n is the unit normal vector, ΨΩ N is the Neumann boundary region.

[0099] Based on the material parameter data M, the humidity diffusion coefficient and hygroscopic performance parameters of different materials are extracted to construct the humidity diffusion parameter mapping; according to the spatial geometric characteristics and material distribution, the position-related humidity diffusion coefficient field is constructed to obtain the humidity diffusion coefficient field; based on the building opening position and ambient wind speed data, the air flow field is constructed to obtain the wind field model; the humidity source term function is designed, and the humidity sources such as humidification equipment and human breathing are considered to obtain the humidity source function. Construct a complete humidity convection-diffusion equation to form a humidity field model: ΨH / Ψt + V t ·▽H - ▽·(D t (x)▽H) = S H (x,t) in Ω; H = g H (x,t) on ΨΩ D ;D t (x)ΨH / Ψn = h H (x,t) on ΨΩ N , where g H and h H is the boundary condition function of the humidity field, D t is the humidity diffusion coefficient field, V t is the wind field model; S H (x, t) is the humidity source function, and H is the humidity field variable.

[0100] Based on the material parameter data M, the optical characteristic parameters of different materials, such as reflectivity, absorptivity, etc., are extracted to construct an optical parameter mapping, determine the position and characteristics of the light source, including natural light sources (sun) and artificial light sources, and build a light source model; define a bidirectional reflectance distribution function (BRDF) for each surface material to describe the interaction between light and the surface and obtain a BRDF set; construct a physics-based radiation transfer equation to simulate the propagation of light in the environment and form a light field model: L(x,ω) = L e (x,ω) + ∫ Ω f r (x,ω',ω)L i(x,ω')(n·ω')dω', where L represents the radiant brightness, L e is the self-luminous term, f r is the BRDF, ω and ω' are the outgoing and incident directions respectively, and n is the surface normal vector.

[0101] Based on multi-resolution grids, the physical model is spatially discretized to construct a spatial discrete operator. A time discrete format suitable for the characteristics of each physical process is designed, such as implicit Euler, Crank-Nicolson or WENO format, to obtain a time discrete operator. For nonlinear terms, a suitable linearization strategy or iterative solution scheme is designed to obtain a nonlinear processing operator. The spatial discrete operator, time discrete operator and nonlinear processing operator are integrated to build a complete numerical solution framework.

[0102] Analyze the coupling relationship between physical fields, such as the impact of temperature on humidity diffusion and light on temperature distribution, and construct a physical field coupling matrix; design numerical processing strategies for coupling terms, such as explicit / implicit coupling, separate solution or overall solution, to obtain a coupling solution strategy; design a scheduling algorithm for multi-physical field calculations, determine the order and iteration method of calculations of different physical fields, and obtain a calculation scheduling algorithm; integrate temperature field models, humidity field models, and light field models, consider the physical field coupling matrix, and construct a complete set of environmental field physical models.

[0103] According to one aspect of the present application, the step of obtaining the a posteriori environment field includes:

[0104] Determine the location of environmental sensors in the multi-resolution grid, design observation operators for each environmental parameter, and obtain a unified observation operator through the observation error model;

[0105] The unified observation operator is combined with the environmental sensor data to construct the initial state vector, and the forward time integration is performed through the environmental field physical model set to obtain the model prediction state and observation residual;

[0106] Based on the model prediction state, a predetermined set of members is generated to represent different possible states, and the prior mean and covariance matrix are calculated to construct the state-observation cross-covariance matrix;

[0107] Combine the state-observation cross-covariance matrix and the observation residual to calculate the Kalman gain, update the model prediction state and set members, and obtain the posterior set;

[0108] The global state vector of the posterior set is decomposed into local region state vectors, and the ensemble Kalman filter algorithm is applied independently to each local region. The posterior environmental field is obtained by fusing the overlapping region information.

[0109] Specifically, for each environmental sensor, determine its position in the multi-resolution grid, establish a spatial mapping between the sensor and the grid nodes, and obtain the sensor position mapping; design a corresponding observation operator for each environmental parameter (temperature, humidity, light), map the model state variables to physical quantities measurable by the sensor, and obtain the parameter observation operator; construct an observation error model for sensor characteristics (such as response time, measurement error), and obtain the observation error distribution; integrate the parameter observation operators, construct a complete observation system equation, and form a unified observation operator H, y = H(x) + ε; where y is the sensor observation value vector, x is the model state vector, and ε is the observation error vector, which obeys the observation error distribution.

[0110] Based on the smoothed environmental data and sensor position mapping, an initial estimate of the environmental field is constructed to obtain the initial state vector; the physical equations in the environmental field physical model set are applied, and the forward time integration is performed using the numerical solution framework to obtain the model prediction state; the model prediction state is mapped to the observation space using a unified observation operator to obtain the model prediction observation y f :y f =H(x f ), where H is the unified observation operator, x f is the model prediction state; compare the model prediction observation with the actual smoothed environmental data, calculate the observation residual, and generate the observation residual O t :O t = E t s -y f , where E t s It is the actual smoothing environment data.

[0111] Based on the model prediction state, N set members are generated to represent different possible states, forming the state set X f :X f = {x f 1 , x f 2 , ..., x f N}; Perturb the observation error distribution, generate N groups of observation samples, and obtain the observation set; calculate the mean and covariance matrix of the state set, and obtain the prior mean and prior covariance; apply the unified observation operator H to the state set, obtain the predicted observation set, and calculate the state-observation cross covariance matrix.

[0112] Based on the prior covariance, cross covariance and observation error distribution, calculate the Kalman gain matrix K: K = P xy ·(H·P f ·H T + R)-1 , where P f is the prior covariance, P xy is the cross covariance, R is the observation error covariance matrix; using the observation residual and Kalman gain K, the model state is updated to obtain the posterior state x a :x a = x f + K·(E t s -y f ); Update each set member to generate the posterior set X a :x a i = x f i + K·(E t s + ε i -H(x f i )), where ε i is the perturbed observation error; calculate the mean and covariance of the posterior set to obtain the posterior mean and posterior covariance.

[0113] Decompose the global state vector into local region state vectors and construct the local state set {x l 1 , x l 2 , ..., x l _M}; for each local area, determine the observation data within its influence range and establish a local observation mapping; independently apply the ensemble Kalman filter algorithm to each local area to obtain the local posterior state set {x a,l 1 , x a,l 2 , ..., x a,l _M}; Using overlapping area information fusion technology, local posterior states are combined into a global posterior state to generate an integrated posterior environment field.

[0114] According to one aspect of the present application, the step of obtaining a training model includes:

[0115] S41. Based on the environmental field physical model set, construct the physically constrained neural ordinary differential equation in the form of dx / dt =f θ (x,t) + g(x,t), we get the mixed model; where f θ is the parameterized part of the neural network, and g is the physical residual term;

[0116] S42, construct environmental field reconstruction, structural response prediction and physical consistency loss functions, and comprehensively form a multi-task loss function;

[0117] S43, converting the hybrid model into a Bayesian model, calculating the posterior distribution of the parameters (e.g., by using a variational inference method), generating environmental field samples in combination with the final environmental field, and constructing a displacement distribution;

[0118] S44. Based on the displacement distribution, a multi-scale model from the material microscopic to the structural macroscopic is constructed, and a cross-scale fusion model is obtained through scale bridging;

[0119] S45. Integrate the hybrid model, Bayesian model and cross-scale fusion model, optimize the parameters through the multi-task loss function, and obtain the training model.

[0120] Specifically, based on the set of physical models of the environmental field, a model architecture with physical constraints is designed to obtain a physical framework; the physical framework is embedded in the neural network architecture, and the neural ordinary differential equations with physical constraints are constructed to generate a hybrid model; a set of feature extraction functions based on physical principles is designed to generate a physical feature extractor; the physical feature extractor is applied to process the final environmental field to extract features with clear physical meanings to obtain physical features.

[0121] Design the environmental field reconstruction loss function, measure the difference between the predicted environmental field and the measured environmental data, and obtain the environmental reconstruction loss; design the structural response prediction loss function, evaluate the difference between the predicted displacement and the measured displacement, and obtain the response prediction loss; construct the physical consistency loss function to ensure that the model prediction conforms to the physical law, and obtain the physical consistency loss; integrate the environmental reconstruction loss, response prediction loss and physical consistency loss to construct a multi-task loss function. Based on the hybrid model, construct a Bayesian neural network to model the model parameter distribution and obtain the Bayesian model; use the variational inference method to approximate the parameter posterior distribution and obtain the parameter distribution; based on the parameter distribution and the final environmental field, generate multiple environmental field samples and obtain the environmental field distribution; propagate the uncertainty of the environmental field to the structural response, calculate the uncertainty of the structural displacement, and obtain the displacement distribution.

[0122] Construct a multi-scale physical model, from the microscopic response of materials to the macroscopic behavior of structures, and obtain a multi-scale model set; design a scale bridging method to realize information transfer between models of different scales, and obtain a scale bridging operator; develop a scale adaptive selection algorithm, select the appropriate model scale according to the computing needs and accuracy requirements, and obtain a scale selection strategy; integrate the multi-scale model set, scale bridging operator and scale selection strategy to construct a cross-scale fusion model. Combine the hybrid model, Bayesian model and cross-scale fusion model to construct a comprehensive model for environmental-structural response prediction, and obtain a comprehensive prediction model; use historical data to train the comprehensive prediction model, optimize the model parameters, and obtain a training model; design a model self-evaluation mechanism, continuously monitor the model performance, and obtain a performance evaluation function; build an anomaly detection and response system to handle extreme environmental conditions, and obtain an anomaly handling strategy.

[0123] According to one aspect of the present application, the step of constructing a displacement distribution comprises:

[0124] The deterministic network parameters of the hybrid model are converted into random variables, and appropriate prior distributions are constructed for different types of parameters to obtain a Bayesian model.

[0125] Construct the evidence lower bound objective function, and obtain the ELBO function and SGVI optimizer through stochastic gradient variational inference and reparameterization;

[0126] Use the SGVI optimizer to maximize the ELBO function, obtain the optimized variational parameters, construct the parameter distribution and extract the parameter sample set;

[0127] For each set of parameter samples, a hybrid model is used for forward prediction to generate an environmental field sample set, calculate statistical moments, and construct an environmental field distribution;

[0128] Based on the environmental field sample set, the corresponding structural response is calculated, the displacement sample set and its probability distribution are constructed, and the displacement distribution is obtained.

[0129] Specifically, based on the hybrid model, the deterministic network parameters are converted into random variables, and the parameter prior distribution is constructed to obtain the parameter prior; suitable prior distributions are designed for different types of parameters, such as Gaussian priors for weights and inverse gamma priors for variance parameters, to construct a hierarchical prior structure; the posterior distribution form of the parameters is designed, such as the mean field Gaussian distribution family, to construct a posterior distribution family; the hybrid model, parameter prior and posterior distribution family are integrated to construct a complete Bayesian model.

[0130] Construct the evidence lower bound (ELBO) objective function as an approximation of the KL divergence between the posterior distribution and the true posterior, and obtain the ELBO function L eLBO :L eLBO = E_q[log P(D|θ)] - KL[q(θ)||p(θ)], where q(θ) is the approximate posterior, p(θ) is the prior, P(D|θ) is the likelihood function, and E_q represents the expectation function; design the stochastic gradient variational inference (SGVI) algorithm, optimize ELBO to obtain the best approximate posterior, and construct the SGVI optimizer; implement the reparameterization technique to reduce the variance of the gradient estimate, improve the optimization stability, and obtain the reparameterization module; design the inference convergence criterion to determine when to stop the variational inference iteration, and construct a convergence detector.

[0131] Use the SGVI optimizer to maximize the ELBO function, obtain the optimized variational parameters, and construct the optimized variational parameters; construct an approximate posterior distribution based on the optimized variational parameters to obtain the parameter distribution; extract multiple groups of parameter samples from the parameter distribution to form a parameter set and obtain a parameter sample set; analyze the statistical characteristics of the parameter sample set, such as mean, variance, distribution shape, etc., and generate a parameter statistical report.

[0132] For each group of parameter samples in the parameter statistical report, a hybrid model is used for forward prediction to generate environmental field samples and obtain an environmental field sample set; the statistical moments of the environmental field sample set, such as the mean field, variance field, etc., are calculated to construct the environmental field distribution; visualization methods are designed to display the uncertainty distribution of the environmental field, such as uncertainty heat maps, confidence intervals, etc., to obtain uncertainty visualization tools; based on the predicted uncertainty of the environmental field, high uncertainty areas and locations that require additional attention or sampling are identified to generate uncertainty hotspot maps.

[0133] Based on the environmental field sample set, the corresponding structural response is calculated to generate a displacement sample set; the statistical characteristics of the displacement sample set are calculated, the probability distribution of the displacement is constructed, and the displacement distribution is obtained; the spatial distribution of displacement uncertainty is analyzed, the structural areas with high uncertainty are identified, and the displacement uncertainty map is generated; the probability distribution of the change detection results is calculated, the reliability of the detection results is evaluated, and the detection reliability evaluation is obtained.

[0134] According to one aspect of the present application, the step of obtaining a change detection result includes:

[0135] S51, input the final environment field into the training model, predict the displacement field caused by the environment, extract the displacement vector corresponding to each point cloud point, and obtain the environment displacement field and the point cloud environment displacement;

[0136] S52, calculating the measured displacement field between the registration point cloud data and the reference point cloud, subtracting the point cloud environment displacement therefrom, and obtaining the compensated displacement field and compensation uncertainty;

[0137] S53, based on regional response statistics and compensation uncertainty, calculating a basic threshold for each region, building an adaptive threshold calculation model in combination with measurement uncertainty, and obtaining an adaptive threshold;

[0138] S54, applying an adaptive threshold to screen the displacement field after compensation, calculating the change confidence through spatial continuity and temporal consistency constraints, and obtaining a change detection result;

[0139] S55. Record the corresponding relationship between the detection threshold and the result, establish a threshold history database, optimize the threshold adjustment strategy through reinforcement learning, and obtain an optimized threshold system.

[0140] Specifically, the current final environmental field is input into the training model to predict the structural displacement field caused by the environment and obtain the environmental displacement field; the uncertainty of the predicted displacement is calculated in combination with the displacement distribution to obtain the displacement uncertainty; the expected displacement vector corresponding to each point cloud point is extracted from the environmental displacement field to obtain the point cloud environmental displacement; the environmental response statistical characteristics of different material areas are calculated to generate regional response statistics.

[0141] The current registration point cloud and the registration point cloud at the reference time point are read, the actual displacement field is calculated, and the measured displacement field is obtained; the point cloud environmental displacement is subtracted from the measured displacement field to obtain the compensated displacement field; based on the displacement uncertainty, the uncertainty propagation of the compensation process is calculated to obtain the compensated uncertainty; the compensated displacement field is applied to the registration point cloud to generate a point cloud after environmental impact compensation to obtain a compensated point cloud.

[0142] Based on regional response statistics and compensation uncertainty, the basic threshold is calculated for each optimized partition to obtain the basic threshold; considering factors such as point cloud acquisition accuracy and registration error, the measurement uncertainty is calculated to obtain the measurement uncertainty; combining the basic threshold and measurement uncertainty, an adaptive threshold calculation model is constructed to obtain the regional threshold model; according to the historical change detection performance, the threshold coefficient is dynamically adjusted to generate the final adaptive threshold.

[0143] An adaptive threshold is applied to perform threshold screening on the compensated displacement field to obtain the initial screening changes; the initial screening changes are further screened using spatial continuity and temporal consistency constraints to obtain the screening changes; the confidence of each detected change is calculated, and the change confidence is generated based on the compensation uncertainty and detection indicators; the screening changes and change confidence are combined to generate the final change detection results.

[0144] Record the change detection results and the corresponding adaptive thresholds, establish a historical threshold database, and obtain the threshold history; based on the threshold history and actual detection results, use the reinforcement learning method to optimize the threshold adjustment strategy and obtain the threshold learning model; for areas with poor detection performance, analyze the failure causes, adjust the corresponding threshold calculation parameters, and obtain the parameter adjustment strategy; integrate the threshold learning model and parameter adjustment strategy into the threshold calculation process to obtain an optimized threshold system.

[0145] According to one aspect of the present application, the step of obtaining an optimized threshold system includes:

[0146] Record adaptive thresholds and corresponding change detection results, calculate detection performance indicators, and build a threshold history database;

[0147] The threshold adjustment is modeled as a reinforcement learning task, the state is defined as the current threshold and historical performance, the action is the direction and amplitude of the threshold adjustment, the reward is the degree of improvement in detection performance, and a threshold learning model is constructed;

[0148] Analyze the spatial distribution of change detection results, identify areas with poor detection performance, study their characteristics, and obtain defect factor analysis;

[0149] Based on defect factor analysis, parameter adjustment rules are designed, automatic adjustment modules and expert intervention interfaces are implemented, and parameter adjustment strategies are constructed;

[0150] By integrating the threshold learning model and parameter adjustment strategy, a unified threshold optimization framework and status monitoring mechanism are constructed to obtain an optimized threshold system.

[0151] Specifically, record the adaptive threshold used for each change detection and its corresponding detection result change detection result; calculate the performance indicators of each detection, such as precision, recall rate, F1 score, etc., and build a performance record; associate the detection threshold, detection results and performance indicators, establish a complete historical data item, and obtain a historical item set; store the historical items by time index to build a complete threshold history.

[0152] Model the threshold adjustment problem as a reinforcement learning task, define the state, action, and reward functions, and construct the RL problem definition: state: current threshold and historical performance; action: threshold adjustment direction and magnitude; reward: degree of improvement in detection performance. Design RL algorithms suitable for threshold optimization, such as Q-learning, policy gradient, etc., and build RL algorithm implementation. Train the RL model based on the threshold history, learn the optimal threshold adjustment strategy, and obtain the threshold policy model. Implement the online learning mechanism so that the model can continuously learn and optimize from new detection results, and build an online learning module. Integrate the RL problem definition, RL algorithm implementation, threshold policy model, and online learning module to build a complete threshold learning model.

[0153] Analyze the spatial distribution of change detection results, identify areas with poor detection performance, and construct a performance defect map; study the characteristics of performance defect areas, such as material properties, environmental conditions, etc., find out the influencing factors, and obtain defect factor analysis; design corresponding improvement strategies for different types of performance defects and construct an improvement strategy mapping; develop performance evaluation automation tools, regularly evaluate detection performance and identify problem areas, and obtain automatic evaluation tools.

[0154] Based on defect factor analysis and improved strategy mapping, customized parameter adjustment strategies are designed for different regions to obtain parameter adjustment rules; an automated mechanism for parameter adjustment is implemented to automatically update threshold parameters based on detection feedback and build an automatic adjustment module; a manual intervention interface is designed to allow experts to manually adjust threshold parameters based on experience and build an expert intervention interface; an audit tracking system for parameter adjustment is established to record all parameter changes and their effects and build an adjustment audit system. Integrate parameter adjustment rules, automatic adjustment modules, expert intervention interfaces and adjustment audit systems to build a complete parameter adjustment strategy. Integrate threshold learning models and parameter adjustment strategies to build a unified threshold optimization framework and get an optimization framework.

[0155] In a specific embodiment of the present application, in the historical building monitoring scene, three-dimensional point cloud data containing 5000 points was collected. The building is composed of three different materials: masonry structure (area 1), wooden structure (area 2) and metal components (area 3). Environmental monitoring includes three parameters: temperature, humidity and light. Data was collected once a day in the past 30 days. The steps of the method for compensating the time series difference of three-dimensional monitoring data and adaptively adjusting the dynamic threshold are as follows:

[0156] Step 1: Regional adaptive partitioning PCA environment feature extraction.

[0157] 1.1. Initial division of structural response area.

[0158] First, the initial division is performed based on the material parameter data: Material parameter data M: masonry structure: thermal expansion coefficient α = 8×10 -6 / °C, elastic modulus E = 10 GPa; wooden structure: thermal expansion coefficient α = 5×10 -6 / °C(radial), 50×10 -6 / °C (tangential), elastic modulus E = 12 GPa; metal components: thermal expansion coefficient α = 23×10 -6 / °C, elastic modulus E = 200 GPa. Three initial partitions are formed according to the material properties, each containing the number of point clouds: Region 1 (brick): 2500 points; Region 2 (wood): 1800 points; Region 3 (metal): 700 points.

[0159] 1.2. Response similarity evaluation and optimized partitioning.

[0160] Method for constructing environment-response mapping table: mapping relationship between environmental factor changes and displacement t = [ΔT t , ΔH t , ΔL t ]; displacement vector D t = [Δx t, Δy t , Δz t ]; where ΔT t is the temperature change (°C); ΔH t is the humidity change (%); ΔL t is the illumination change (lux); Δx t , Δy t , Δz t are the displacements in the x, y, and z directions (mm) respectively; t is the time index. The sensitivity coefficient of a point in area 1 is calculated by taking the data of the last five days (t = 1 to 5): Environmental change data: Date 1: ΔT = 5.2°C, ΔH = 12%, ΔL = 1500 lux; Date 2: ΔT = -3.1°C, ΔH = 5%, ΔL = -800 lux; Date 3: ΔT = 1.8°C, ΔH = -7%, ΔL = 500 lux; Date 4: ΔT = 0.5°C, ΔH = 3%, ΔL =1200 lux; Date 5: ΔT = -2.3°C, ΔH = -8%, ΔL = -1300 lux. Corresponding displacement data: Date 1: Δx = 0.12mm, Δy = 0.08mm, Δz = 0.15mm; Date 2: Δx = -0.07mm, Δy = -0.05mm, Δz = -0.09mm; Date 3: Δx = 0.04mm, Δy = -0.02mm, Δz = 0.05mm; Date 4: Δx =0.01mm, Δy = 0.03mm, Δz = 0.03mm; Date 5: Δx = -0.06mm, Δy = -0.07mm, Δz =-0.08mm.

[0161] Multivariate regression analysis method: Δd = β t ·ΔT + β H ·ΔH + β_L·ΔL + ε; where Δd is the displacement component (mm); β t , β H , β_L is the environmental factor sensitivity coefficient; ΔT is the temperature change (°C); ΔH is the humidity change (%); ΔL is the light change (lux); ε is the residual term. Through multivariate regression analysis, the sensitivity coefficient matrix S of this point is obtained: S = [β t x, β H x, β_Lx;β t y, β H y, β_Ly;β t z, β Hz, β_Lz]= [0.0223, 0.0031, 0.00005; 0.0157, 0.0042, 0.00002; 0.0278, 0.0026, 0.00004].

[0162] Point-to-point response mode similarity calculation method: point-to-point similarity PS z (i,j) = exp(-||S z (i)-S z (j)|| 2 / σ 2 );where S z (i) S z (j) is the sensitivity coefficient matrix of point i and point j; ||·|| is the matrix Frobenius norm; σ is the scale parameter, which is 0.05; i and j are the point indices in the point cloud. Calculate the response similarity of two adjacent points (i=15, j=16) in region 1: S(15) = [0.0223, 0.0031, 0.00005;0.0157, 0.0042, 0.00002;0.0278,0.0026, 0.00004]; S(16) = [0.0235, 0.0029, 0.00006;0.0149, 0.0045, 0.00002;0.0283, 0.0024, 0.00005]; ||S(15)-S(16)|| 2 = 0.0000267; PS z (15,16) = exp(-0.0000267 / 0.05 2 ) = 0.9989, indicating that the environmental response patterns of these two points are very similar.

[0163] Construct the response correlation matrix R z , R z =PS z (i,j)× W(d), where W(d) is the spatial distance weighting function; W(d) = exp(-d 2 / 0.1 2 ) is the spatial distance weighting function, d is the distance between two points; the Laplace matrix is ​​calculated for the response correlation matrix: the diagonal matrix element D z (i,i) = Σj R z (i,j); Laplace matrix L z norm (i,j) = -R z (i,j) / sqrt(D z (i,i)×D z (j,j)); calculate Lz norm The eigenvalues ​​of the region 1 are taken, and the first three eigenvectors are taken to construct the spectral embedding. K-means (k=2) clustering is applied to obtain the sub-partitioning. After applying spectral clustering analysis to region 1, it is found that there are differences in the point response patterns within the region, and it is further subdivided into two sub-regions: region 1A (south-facing masonry): 1400 points; region 1B (north-facing masonry): 1100 points. The final optimized partitioning results are: region 1A (south-facing masonry): 1400 points; region 1B (north-facing masonry): 1100 points; region 2 (wooden structure): 1800 points; region 3 (metal components): 700 points.

[0164] 1.3. Time series weighted PCA feature extraction.

[0165] Time decay weight function calculation method: W t (τ) = exp(-λ·(t-τ)); where t is the current time point; τ is the historical time point; λ is the time decay factor, which is 0.2; exp is the base of the natural logarithm. Using the data from the past 10 days (τ = t-10 to t-1), set the time decay factor λ=0.2, and calculate the weight of each historical time point: W t (t-1) = exp(-0.2×1)= 0.8187; W t (t-2) = exp(-0.2×2) = 0.6703;...;W t (t-10) = exp(-0.2×10) = 0.1353. Normalized weight: sum = 0.8187 + 0.6703 + ... + 0.1353 = 4.311; W t norm (t-1) =0.8187 / 4.311 = 0.1899; W t norm (t-2) = 0.6703 / 4.311 = 0.1555;...;W t norm (t-10) =0.1353 / 4.311 = 0.0314.

[0166] Weighted covariance matrix construction method: C z w = ∑ τ W t norm (τ)·(X τ -μ w ) T ·(X τ -μ w ); where X τis the environment-displacement joint matrix at time point τ; μ w is the weighted mean vector; W t norm (τ) is the normalized time weight; τ is the historical time point. Construct the environment-displacement joint matrix J for area 1A, where each row represents a time point and the columns include temperature, humidity, light, and displacement in the x, y, and z directions: J = [T, H, L, Δx, Δy, Δz]. An example of data at a certain time point is: J(t-1) = [25.3, 65, 5200, 0.12, 0.08, 0.15], and calculate the weighted mean vector μ w :μ w = ∑ τ W t norm (τ)·J(τ)=0.1899×[25.3, 65, 5200, 0.12, 0.08, 0.15] + ... = [23.7, 63.2, 4850, 0.08,0.06, 0.11]; calculate the centering matrix J c :J c (t-1) = J(t-1) - μ w = [1.6, 1.8, 350, 0.04,0.02, 0.04]; construct the weighted covariance matrix C z w , the example only gives part of the calculation: C z w (1,1) = 0.1899×1.6 2 +0.1555×(-1.2) 2 + ... = 3.24; C z w (1,4) = 0.1899×1.6×0.04 + 0.1555×(-1.2)×(-0.03) + ... = 0.062; ... Through eigenvalue decomposition, we get the sorted eigenvalues ​​and corresponding eigenvectors: eigenvalues: λ= [15.6, 8.2, 3.5, 0.9, 0.4, 0.1]; cumulative contribution rate: [54.2%, 82.8%, 95.0%, 98.1%,99.5%, 100%]; select the first three principal components, the contribution rate reaches 95%, and construct the principal component matrix: P = [e1, e2, e3]=[0.15, 0.42, 0.31;0.22, 0.38, -0.53;0.82, -0.29, 0.06;0.31, 0.45, 0.52;0.28,0.51, -0.33;0.29, 0.38, 0.48].

[0167] 1.4. Robust environmental feature extraction.

[0168] Huber loss function method: ρ Huber (x) = {x 2 / 2, |x| ≤ c; c · |x| - c 2 / 2, |x| > c}; weight function w(x) = {1, |x| ≤ c; c / |x|, |x| > c}; where x is the residual; c is the threshold parameter, which is 1.345; |x| is the absolute value. The Huber loss function is applied to identify and reduce the influence of outliers. For the environment-displacement joint matrix, the Mahalanobis distance is calculated, and the weight function is applied: the Mahalanobis distance of a data point: d = 2.6; since d >1.345, the weight w = 1.345 / 2.6 = 0.517; 5 outliers in area 1A are identified, and the robust data matrix is ​​obtained after weight correction, and the robust coefficient matrix is ​​obtained by projection through the PCA model. Analyze the physical meaning of the principal components: the first principal component (54.2%): mainly captures the changes in illumination and the displacement caused by it; the second principal component (28.6%): mainly reflects the changes in temperature and the displacement caused by it; the third principal component (12.2%): mainly represents the influence of humidity changes. Environmental feature vector (region 1A): v env = [0.82, 0.31, 0.28, 0.29], indicating that light has the greatest impact on this area, followed by temperature.

[0169] Step 2: Reconstruction of environmental field enhanced by physical information.

[0170] Through the environmental field reconstruction technology enhanced by physical information, the environmental field distribution of the entire monitoring area is accurately reconstructed from the limited environmental sensor data. Five environmental sensors are arranged in the monitoring area to measure temperature, humidity and light. The sensor locations are as follows: Sensor 1: South side of the building (25.4°C, 58%, 6200lux); Sensor 2: North side of the building (22.1°C, 62%, 1800lux); Sensor 3: Center of the room (23.8°C, 55%, 850lux); Sensor 4: East side of the room (24.2°C, 54%, 1200lux); Sensor 5: West side of the room (23.5°C, 56%, 750lux).

[0171] 2.1. Construction of three-dimensional computational domain geometric model.

[0172] The computational domain geometry model is constructed based on the point cloud data. The building size is 20m×15m×8m, and the wall thickness is about 0.5m. Set boundary conditions: South wall outer surface: T = ambient temperature (sensor 1), convection heat dissipation coefficient h = 25 W / (m 2 ·K); North wall outer surface: T = ambient temperature (sensor 2), convection heat dissipation coefficient h = 20 W / (m 2 ·K); Internal space: initial temperature is set to 23.5°C.

[0173] 2.2. Multi-resolution mesh generation.

[0174] Calculation method of grid refinement index: CI(x) = α·|▽T(x)| + β·|▽H(x)| + γ·|▽L(x)|; where CI(x) is the refinement index at position x; ▽T(x), ▽H(x), ▽L(x) are the gradients of temperature, humidity and light, respectively; α, β, γ are weight coefficients, which are 1.0, 0.6, and 0.8, respectively; |·| is the gradient norm. Calculate the gradient of environmental parameters, taking temperature as an example: initial uniform grid point spacing: 0.5m; temperature gradient between north and south walls: |▽T| = |25.4-22.1| / 20= 0.165°C / m; temperature gradient in window area: |▽T| = |25.4-23.8| / 0.5 = 3.2°C / m. Based on the gradient calculation, refine the index: window area: CI = 1.0×3.2 + 0.6×8 + 0.8×670 = 540.4; wall area: CI =1.0×0.62 + 0.6×1.5 + 0.8×5 = 5.42. Set the threshold CI t hreshold = 10, use finer meshes in areas with large environmental gradients such as windows and doors: regular areas: mesh size 0.5m; high gradient areas: mesh size 0.1m. The resulting multi-resolution mesh contains about 85,000 nodes.

[0175] 2.3. Construction of heterogeneous physical model.

[0176] Heterogeneous heat conduction equation: ΨT / Ψt = ▽·(κ t (x)▽T) + S t (x, t); where: T is the temperature field (°C); t is the time (s); κ t (x) is the thermal conductivity at position x (W / (m·K)); S t (x,t) is the heat source function (W / m 3);▽ is the gradient operator;Ψ is the partial derivative. Set the thermal conductivity for different material areas: masonry structure: κ = 0.6-1.0 W / (m·K); wooden structure: κ = 0.12-0.04 W / (m·K); metal components: κ = 45-55 W / (m·K). Light heat source function (south window area): S t (x,t) = η·I(x,t)·(1-ρ(x)), where η is the photothermal conversion efficiency (0.3) and I is the incident light intensity (W / m 2 ), ρ is the reflectivity (0.2).

[0177] Humidity convection-diffusion equation: ΨH / Ψt + V t ·▽H = ▽·(D t (x)▽H) + S H (x, t); where H is the humidity field (%); t is time (s); V t is the wind field (m / s); D t (x) is the humidity diffusion coefficient (m 2 / s); S H (x, t) is the humidity source function (% / s); ▽ is the gradient operator; Ψ represents the partial derivative. The humidity diffusion coefficient is related to the material and temperature: For porous materials: D t (x) = D_0·exp(b·T(x)), where D_0 is the reference diffusion coefficient (2.5×10 -6 m 2 / s), b is the temperature coefficient (0.025 / °C).

[0178] Radiative transfer equation: L(x,ω) = L e (x,ω) + ∫ Ω f r (x,ω',ω)L i (x,ω')(n·ω')dω'; where L is the radiant brightness (W / (m 2 ·sr));L e is the self-luminous item; f r is the bidirectional reflectance distribution function; ω and ω' are the outgoing and incident directions respectively; n is the surface normal vector; Ω is the hemispherical space; sr is the solid angle unit. The south window uses an anisotropic BRDF model: r (ω i ,ω_o) = k D / π + (n+2) / (2π)·k_s·(ω r ·ω_o) n , where k Dis the diffuse reflection coefficient (0.3), k_s is the specular reflection coefficient (0.7), and n is the glossiness parameter (10).

[0179] 2.4. Environmental field data assimilation.

[0180] Ensemble Kalman filter update equation: x a = x f + K·(y - H(x f )); K = P xy ·(H·P f ·H T +R) -1 ; where: x a is the posterior state vector; x f is the prior state vector; K is the Kalman gain matrix; y is the observation vector; H is the observation operator; P xy is the state-observation cross covariance matrix; P f is the prior error covariance matrix; R is the observation error covariance matrix. Initialize the state vector x based on the physical model f , containing the temperature, humidity and light values ​​of all grid points: x f = [T 1 ,T 2 , ..., T N , H 1 , H 2 , ..., H N , L 1 , L 2 , ..., L N ] T ; The observation vector y contains the measurements of 5 sensors: y = [25.4, 22.1, 23.8, 24.2, 23.5, 58, 62, 55, 54, 56, 6200, 1800, 850, 1200,750] T , generate 50 set members, perturb the initial state to build a priori set: x f i = x f + ε i , i=1,2,...,50; where ε i The perturbation is random, with a standard deviation of 0.5°C for temperature perturbation, 2% for humidity, and 100 lux for illumination. After calculating the prior ensemble mean and covariance matrix, the ensemble Kalman filter is applied to update the posterior state: observation error: temperature ±0.3°C, humidity ±2%, illumination ±50lux; Kalman gain matrix K: dimension (3N×15); posterior state x a: Dimension (3N×1). The global state is decomposed into 5 sub-regions and filters are applied independently. Information fusion is performed in the overlapping regions to finally obtain the complete posterior environment field.

[0181] 2.5. Environmental field feedback optimization.

[0182] Environmental field parameter adjustment method: α N ew = α old ·(1 + η·e r ), where α N ew、α old are the model parameters before and after the update; η is the learning rate (0.1); e r is the response residual, defined as (d meas -d pred ) / d meas ;d meas is the measured displacement; d pred To predict displacement. Calculate the difference between the theoretical structural response and the measured displacement: Theoretical temperature response of a point in area 1A: 0.107mm; measured displacement: 0.127mm; response residual: e r = (0.127-0.107) / 0.127 = 0.157. Adjusted thermal conductivity: Before adjustment: κ old = 0.8 W / (m·K); after adjustment: κ New = 0.8×(1+0.1×0.157) = 0.813W / (m·K). After feedback optimization and adjustment, the final environmental field can accurately reflect the temperature, humidity and light distribution of the entire building.

[0183] Step 3: Cognitive physics-data dual-driven compensation model: By building a cognitive physics-data dual-driven compensation model, the physical model is embedded in the data-driven architecture while taking the uncertainty of the model into account.

[0184] 3.1. Construction of Neural Ordinary Differential Equations with Physical Constraints.

[0185] Physical Constraints God Ordinary Differential Equation: dx / dt = f θ (x,t) + g(x,t); x is the state vector; t is time; f θ is a parameterized neural network; θ is a network parameter; g is a physical constraint. The environmental field physical model is transformed into a neural ordinary differential equation form: For the temperature field: dT / dt = f θt (T,t) + ▽·(κ(x)▽T); neural network f θt A three-layer structure is adopted (4 input nodes, 16 hidden nodes, and 1 output node), and the physical constraints are discretized using the finite difference method.

[0186] 3.2. Construction of multi-task loss function.

[0187] Multi-task loss function: L_mt = α·L e + β·L r + γ·L_p; where L e Reconstruction of losses for the environment; L r is the response prediction loss; L_p is the physical consistency loss; α, β, and γ are weight coefficients, which are 0.4, 0.4, and 0.2 respectively. Environmental reconstruction loss (MSE): L e = 1 / N·∑ i=1 N (E predi -E measi ) 2 ; where E predi is the predicted environmental variable value of the i-th sample; E measi is the true measured environmental variable value of the i-th sample, N is the number of samples used to calculate the environmental reconstruction loss; response prediction loss (MAE): L r = 1 / M ∑ j=1 M |D predj -D measj |; where D predj is the predicted response variable value of the jth sample; D measj is the true measured response variable value of the jth sample, M is the number of samples used to calculate the response prediction loss; Physical consistency loss (energy conservation): L_p = |∫ Ω (ΨT / Ψt - ▽·(κ▽T) - S) 2 dΩ|; Loss value of a batch in model training: L e = 0.052; L r = 0.018; L_p = 0.031; L_mt = 0.4×0.052 + 0.4×0.018 + 0.2×0.031 = 0.034.

[0188] 3.3. Bayesian model and displacement distribution.

[0189] ELBO function: L e LBO = E_q[log P(D|θ)] - KL[q(θ)||p(θ)]; where E_q represents the expectation about q(θ); log P(D|θ) is the log likelihood; KL[q(θ)||p(θ)] is the KL divergence; q(θ) is the approximate posterior distribution; p(θ) is the parameter prior distribution. Convert the mixed model to a Bayesian neural network and set the prior distribution for the network weights: weight w ~ N(0, σ2 ); bias b ~ N(0, 0.1). Using mean field variational inference, approximate the posterior distribution: q(w) ~ N(μ w , σ w 2 );q(b) ~ N(μ b , σ b 2 ); Maximize the ELBO function to optimize the variational parameters. The ELBO value after 500 iterations is: L e LBO = 256.3. 100 sets of parameter samples are drawn from the posterior distribution to predict the temperature change in area 1A: the expected displacement mean for a 10°C temperature rise: 0.241mm; standard deviation: 0.032mm; 95% confidence interval: [0.178mm, 0.304mm].

[0190] 3.4. Multi-scale model construction.

[0191] Establish a multi-scale model from material micro to structural macro: Micro level (10 -6 m): Crystal thermal vibration model; mesoscopic level (10 -3 m): Material volume thermal expansion model; macroscopic level (10 0 m): Overall structural deformation model. Example of cross-scale parameter relationship: Mesoscopic-macroscopic bridging: The thermal expansion coefficient α affects the macroscopic displacement d: d = α·ΔT·L.

[0192] 3.5. Training model integration.

[0193] Integrate hybrid model, Bayesian model and cross-scale fusion model: Comprehensive prediction model: C_m = {H_m, B_m, F_ms}; Use 30 days of historical data to train the model, and the performance indicators on the validation set are: Environmental field reconstruction RMSE: temperature 0.38°C, humidity 1.7%, light 86lux; Displacement prediction MAE: 0.023mm; Displacement prediction R 2 : 0.924.

[0194] Step 4: Dynamic threshold adaptive adjustment: Dynamic threshold adaptive adjustment based on the environment-structure response model is implemented to improve the accuracy of change detection.

[0195] 4.1. Environmental displacement field prediction.

[0196] Environmental displacement prediction method: D t env (x) = C_m train (E t final (x)); where D tenv (x) is the predicted environmental displacement field at position x; C_m train is the trained comprehensive model; E t final (x) is the final environmental field at position x. Input the current environmental field data into the training model: Middle point of the south wall (x0): temperature: 27.3°C; humidity: 52%; illumination: 5800lux. Predicted displacement caused by the environment: x direction: 0.145mm; y direction: 0.063mm; z direction: 0.118mm; total displacement: 0.197mm; displacement uncertainty: ±0.032mm.

[0197] 4.2. Timing difference compensation.

[0198] Timing difference compensation method: D t comp (x) = D t meas (x) - D t env (x); where D t comp (x) is the displacement field after compensation; D t meas (x) is the measured displacement field; D t env (x) is the environmental displacement field. Calculate the difference between the measured displacement and the reference point cloud: Measured displacement of the middle point (x0) of the south wall: x direction: 0.183mm; y direction: 0.071mm; z direction: 0.157mm; total displacement: 0.250mm. Apply environmental displacement compensation: after compensation in the x direction: 0.183-0.145 = 0.038mm; after compensation in the y direction: 0.071-0.063 = 0.008mm; after compensation in the z direction: 0.157-0.118 = 0.039mm; total displacement after compensation: 0.055mm; compensation uncertainty: ±0.045mm.

[0199] 4.3. Regional adaptive threshold calculation.

[0200] Adaptive threshold function: T a (x) = T b ·(1 + γ·U t comp (x) / σ r ef); where T a (x) is the adaptive threshold at position x; T b is the basic threshold (0.05 mm); γ is the adjustment coefficient (0.8); U t comp(x) is the compensation uncertainty at position x; σ ref is the reference standard deviation (0.03mm). Calculate the basic threshold for area 1A: Based on historical response statistics: displacement residual standard deviation: 0.032mm; point cloud measurement error: 0.018mm; basic threshold T b = 0.05 mm. Calculate the adaptive threshold at the middle point of the south wall (x0): Compensate for the uncertainty U t comp = 0.045mm; Adaptive threshold T a = 0.05×(1 + 0.8×0.045 / 0.03) = 0.06mm. Compared with the fixed threshold method: under the fixed threshold (0.1mm): 0.055mm < 0.1mm, it is judged as no change; under the adaptive threshold: 0.055mm < 0.06mm, it is judged as no change, but close to the threshold.

[0201] 4.4. Change detection and result screening.

[0202] Change confidence calculation method: V t (x) = (D t comp (x) - T a (x)) / (2·U t comp (x)); if V t (x) > 0, then normalize to the interval [0,1]; where V t (x) is the change confidence at position x; D t comp (x) is the displacement field after compensation; T a (x) is the adaptive threshold; U t comp (x) is the compensation uncertainty. A set of potential change points (exceeding the adaptive threshold) are detected on the west side of the building: displacement of position y0 after compensation: 0.078mm; adaptive threshold: 0.058mm; compensation uncertainty: 0.042mm. Change confidence V t = (0.078-0.058) / (2×0.042) = 0.238; after normalization: 0.643; after applying the spatial continuity constraint, a continuous change area (10cm 2 ), contains 27 points.

[0203] 4.5. Dynamic threshold learning and optimization.

[0204] Use reinforcement learning to optimize the threshold adjustment strategy: state: current threshold, historical detection performance; action: threshold adjustment amount (-0.01mm, 0, +0.01mm); reward: detection accuracy improvement. After 100 monitoring cycles, the threshold system reached a stable state: Region 1A: 0.058mm±0.012mm; Region 1B: 0.062mm±0.015mm; Region 2: 0.072mm±0.018mm; Region 3: 0.045mm±0.010mm.

[0205] In the test scenario, the false positive rate of the traditional fixed threshold method was 18.7%, which was reduced to 3.2% in this embodiment, and the true positive rate was increased from 85.3% to 94.8%. Through the regional adaptive partition PCA, the structural areas with different materials, orientations and stress conditions were correctly processed, so that the compensation model can achieve the best effect in each area. The displacement prediction errors of areas 1A and 1B were reduced from 0.085mm and 0.092mm to 0.021mm and 0.026mm respectively. From the limited data of only 5 environmental sensors, the environmental field distribution of the entire building was accurately reconstructed, and the RMSE of temperature field reconstruction was 0.38°C, which was better than the 1.25°C of the interpolation method. The cognitive compensation model of this embodiment combines physical and data-driven methods, and the generalization performance of the simple data-driven method under extreme environmental conditions (temperature change>15°C) is improved by 45%. Compared with the fixed threshold method, the adaptive threshold of this embodiment improves the F1 score of change detection under variable environmental conditions by 18.3%, especially during the seasonal transition period.

[0206] The present invention can effectively deal with the nonlinear, multi-factor coupling, spatial heterogeneity and sensor sparsity problems encountered in traditional environmental compensation. Its technical effects are mainly reflected in the following aspects: the introduction of nonlinear kernel function mapping and regional PCA feature extraction enables the complex coupling relationship of environmental factors to be modeled more accurately, so that the environmental parameter field can still be accurately reconstructed under extreme environmental conditions, and targeted structural response compensation can be provided. Through initial partitioning, response correlation matrix construction, spectral clustering analysis and optimized partitioning algorithm, sub-regions with high response similarity are divided; then, time-series weighted PCA and robust M-estimation are used in each sub-region to extract environmental features, achieving fine reconstruction of the local environmental field and significantly reducing the regional error caused by the overall unified model. By combining discrete environmental sensor data with registration point cloud data and material parameter data, a three-dimensional computational domain geometric model and a multi-resolution grid are constructed. Through data assimilation and physical information enhancement, the limited sensor information is expanded to the environmental parameter field of the entire structural surface; this not only solves the problem of insufficient sensor coverage, but also provides an accurate basis for environmental compensation for each point cloud point. The deterministic network parameters are converted into random variables. By setting prior distributions for weights and biases, and calculating ELBO under the framework of mean field variational inference, the posterior distribution of parameters is optimized. On the one hand, Bayesian inference is used to describe the model parameters probabilistically, quantifying the uncertainty of parameters and predictions. On the other hand, by extracting parameter samples for forward prediction, the distribution of structural response (displacement) is constructed, thus providing a reliable statistical basis for the subsequent dynamic threshold adaptive adjustment. The combination of three task losses, namely environmental reconstruction, response prediction and physical consistency, is adopted. By reasonably configuring weights, it not only ensures the full utilization of historical data and sensor data, but also enables the model to maintain internal consistency under physical constraints. By using the dynamic threshold adaptive adjustment mechanism, the model can automatically calibrate the detection threshold according to the difference between real-time prediction and monitoring data, thereby reducing the false positive rate and improving the reliability and stability of the monitoring system. It not only completes environmental field reconstruction and displacement prediction at a single scale, but also realizes the bridging of cross-scale data by integrating multi-scale information of microscopic materials, local structures and the whole structure; it uses a time series weighting mechanism to capture the changing trends and timeliness in historical data, and further improves the effect of time series difference compensation, so that monitoring not only accurately reflects the instantaneous state, but also takes into account the continuity of dynamic changes. The present invention improves the adaptability and accuracy of the monitoring system under complex structures and changing environmental conditions. By organically combining physical models, data-driven methods and Bayesian statistics, an accurate description of environmental compensation and structural response matching is achieved, reducing the risk of false alarms and missed detections, and providing a highly robust and reliable technical solution for structural health monitoring.

[0207] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for compensating for time series differences in three-dimensional monitoring data and adaptively adjusting dynamic thresholds, characterized in that: include: Collect and pre-process point cloud data to obtain registration point cloud data; Based on the registered point cloud data and pre-stored material parameter data, PCA environmental features are extracted by region to obtain regional environmental features; Using regional environmental features and pre-stored environmental sensor data, perform physical information enhanced environmental field reconstruction to obtain the final environmental field and build a cognitive physical-data dual-driven compensation model based on it to obtain a training model; When in use, the dynamic threshold adaptive adjustment and timing difference compensation are performed through the training model, and the environmental displacement field, the compensated displacement field and the adaptive threshold are calculated to obtain the change detection result.

2. The method according to claim 1, characterized in that: The steps of extracting PCA environmental features by region and obtaining regional environmental features include: Based on the registered point cloud data and material parameter data, the initial division of the structural response area is performed to obtain the initial partition; the response similarity is evaluated, and the optimized partition is obtained by constructing the response correlation matrix and spectral clustering analysis; Apply time-series weighted PCA feature extraction to the optimized partitions, construct a time-decay weight function and a weighted covariance matrix, obtain a PCA model and perform robust environmental feature extraction to obtain an environmental feature vector; Based on the environmental feature vector and PCA model, a nonlinear kernel function mapping is constructed to obtain the regional environmental characteristics.

3. The method according to claim 2, characterized in that The steps for evaluating response similarity and obtaining optimized partitions include: Extract environmental change data and registration point cloud data at historical time points to establish an environment-response mapping table; Based on the environment-response mapping table, multivariate regression analysis is performed on the point cloud within the initial partition to obtain a sensitivity coefficient matrix; based on this, the point-to-point response pattern similarity is calculated and a response correlation matrix is ​​constructed; Laplace matrix transformation and eigenvalue decomposition are applied to the response correlation matrix, and spectral clustering analysis is performed to obtain the sub-partition consistency matrix; based on this, the partition consistency index and the similarity between partitions are calculated, and partition merging or splitting operations are performed to obtain the optimized partition.

4. The method according to claim 3, characterized in that: The steps of applying time-series weighted PCA feature extraction to obtain the PCA model include: Based on the optimized partition and environment-response mapping table, the environment-displacement joint matrix is ​​constructed; Based on the preset time decay coefficient, the exponential decay function is applied to calculate the weight of the historical time point to obtain the normalized time weight; based on this, the weighted mean vector and weighted outer product of the environment-displacement joint matrix are calculated to construct the regularized covariance matrix; Perform eigenvalue decomposition on the regularized covariance matrix, sort it by contribution rate and select the main eigenvectors to obtain the principal component matrix, and then construct the projection transformation matrix and reconstruction transformation matrix based on it. Combined with the weighted mean vector, the PCA model is generated.

5. The method according to claim 4, characterized in that The steps of extracting robust environmental features and obtaining environmental feature vectors include: Based on the environment-displacement joint matrix, the Huber loss function is used as a robust M-estimator to obtain the weight function; The weight function is used to calculate the weight of data points, identify outliers and make corrections to obtain a robust data matrix; Project the robust data matrix through the projection matrix of the PCA model to generate a robust coefficient matrix; The corresponding relationship between environmental parameters and principal components is analyzed based on the robust coefficient matrix, key environmental features are extracted, and environmental feature vectors are obtained.

6. The method according to claim 1, characterized in that The steps of performing physical information enhanced environment field reconstruction to obtain the final environment field include: Based on regional environmental characteristics, registered point cloud data and material parameter data, a three-dimensional computational domain geometric model is constructed, the spatial gradient of environmental parameters is calculated, and a multi-resolution grid is generated; For the temperature field, humidity field and light field, a heterogeneous physical model is constructed in combination with material parameter data, including the heat conduction equation, convection-diffusion equation and radiation transfer equation, to obtain the environmental field physical model set; it is then fused with the environmental sensor data, data assimilation is performed, and the posterior environmental field is obtained; The theoretical structural response is calculated based on the a posteriori environmental field and compared with the measured displacement. The environmental field parameters are adjusted through feedback optimization to obtain the final environmental field.

7. The method according to claim 6, characterized in that The steps to obtain the environmental field physical model set include: Based on the material parameter data, the inhomogeneous heat conduction equation is established through the position-dependent heat conductivity field and heat source function to obtain the temperature field model; Based on the material parameter data and pre-stored spatial geometric characteristics, the humidity convection-diffusion equation is established through the humidity diffusion coefficient field and the air flow field to obtain the humidity field model; Based on the optical characteristic parameters in the material parameter data, the radiation transfer equation is constructed through the bidirectional reflectance distribution function to obtain the illumination field model; Based on multi-resolution grids, a numerical solution framework is constructed, the coupling relationship between physical fields is analyzed, and the temperature field, humidity field and light field models are integrated to obtain a set of environmental field physical models.

8. The method according to claim 6, characterized in that The steps to obtain the posterior environment field include: Determine the location of environmental sensors in a multi-resolution grid and construct a unified observation operator through the observation error model; combine it with the environmental sensor data to construct the initial state vector, perform forward time integration through the environmental field physical model set, and obtain the model prediction state and observation residual; Based on the model prediction state, a predetermined set of members is generated and the prior mean and covariance matrix are calculated to construct the state-observation cross-covariance matrix; the Kalman gain is calculated in combination with the observation residual, the set members are updated, and the posterior set is obtained; The global state vector of the posterior set is decomposed into local region state vectors, and the posterior environmental field is obtained through ensemble Kalman filtering and overlapping region information fusion.

9. The method according to claim 6, characterized in that The steps to obtain the training model include: Based on the set of environmental field physical models, the physically constrained neural ordinary differential equations are constructed to obtain a hybrid model; Construct environmental field reconstruction, structural response prediction and physical consistency loss functions to form a multi-task loss function; The hybrid model is converted into a Bayesian model, the posterior distribution of parameters is calculated, and the displacement distribution is constructed by combining the final environmental field. Based on this, a multi-scale model from the microscopic material to the macroscopic structure is constructed, and a cross-scale fusion model is obtained through scale bridging. The hybrid model, Bayesian model and cross-scale fusion model are integrated, and the parameters are optimized through the multi-task loss function to obtain the training model.

10. The method according to claim 9, characterized in that The steps to construct the displacement distribution include: The deterministic network parameters of the hybrid model are converted into random variables, and appropriate prior distributions are constructed for different types of parameters to obtain a Bayesian model. Construct the evidence lower bound objective function, and obtain the ELBO function and SGVI optimizer through stochastic gradient variational inference and reparameterization; Use the SGVI optimizer to maximize the ELBO function, obtain the optimized variational parameters, construct the parameter distribution and extract the parameter sample set; For each set of parameter samples, the hybrid model is used for forward prediction to generate an environmental field sample set; based on this, the corresponding structural response is calculated, the displacement sample set and its probability distribution are constructed, and the displacement distribution is obtained.

Images