Building deformation analysis method and system based on 3D laser scanning and deep learning

Abstract

The present invention discloses a method and system for analyzing building deformation based on three-dimensional laser scanning and deep learning. The method comprises obtaining original point cloud data, hierarchically processing through a feature decomposition method to obtain hierarchical feature data, combining environmental parameter data for fusion analysis to eliminate environmental influences, obtaining high-fidelity point cloud data through quality assessment optimization and spatiotemporal feature fusion; performing component segmentation and deformation feature extraction on the high-fidelity point cloud data, and obtaining fused deformation feature data through multi-scale analysis; then inputting the deformation feature data into a hybrid neural network for feature enhancement and deformation evolution analysis to generate deformation prediction data; performing physical law verification, multi-source data cross-validation, and spatiotemporal continuity optimization on the prediction results, and outputting the final optimized data and reliability report. The present invention improves the accuracy and reliability of building deformation monitoring through a multi-level data processing and verification mechanism.

Description

Technical Field

[0001] The present invention belongs to the field of building deformation monitoring and analysis, and in particular to a building deformation analysis method and system based on three-dimensional laser scanning and deep learning. Background Art

[0002] Deformation monitoring of building structures is of great significance to ensuring building safety. With the acceleration of urbanization, large and complex buildings continue to emerge. Traditional manual measurement and simple sensor monitoring can no longer meet the needs of high-precision, large-scale, and real-time monitoring. With its fast, non-contact, and high-precision characteristics, 3D laser scanning technology can obtain high-density point cloud data on the surface of buildings. Combined with deep learning methods for intelligent analysis, it can realize automatic monitoring and early warning of building deformation, which is of great practical significance for preventing building safety accidents, extending the service life of buildings, and reducing maintenance costs.

[0003] At present, building deformation monitoring mainly uses measuring robots, displacement sensors, inclination sensors and other equipment for fixed-point monitoring, or uses ground-based 3D laser scanners to obtain point cloud data and then perform simple geometric feature extraction and deformation analysis. In terms of data processing, it mainly relies on the ICP registration algorithm for point cloud matching, the RANSAC method for plane fitting and feature extraction, and the principal component analysis method for deformation feature identification. Although these methods have achieved the monitoring of building deformation to a certain extent, they still have many limitations when processing large-scale point cloud data in complex environments.

[0004] The existing technology has the following specific problems: First, point cloud data is easily affected by environmental factors such as temperature, humidity, and light, resulting in fluctuations in measurement accuracy, and the existing methods lack an effective environmental compensation mechanism; second, in terms of component segmentation, the existing methods are difficult to accurately identify the subtle components in complex buildings and the topological relationship between them, especially when dealing with irregular shapes and complex connections, segmentation errors are prone to occur; third, in terms of deformation feature extraction, traditional methods often only focus on geometric features of a single scale, cannot effectively capture deformation patterns under multiple scales, and lack an evaluation mechanism for the reliability of deformation features; fourth, in terms of early warning decision-making, the existing system usually uses a fixed threshold for simple judgment, lacks analysis of the overall stress state of the building and physical verification of the early warning results, and is prone to false alarms or missed alarms; finally, there is a lack of effective cross-validation and consistency optimization mechanisms between the various links of data processing, making it difficult to ensure the reliability and stability of the entire monitoring system. These technical problems have seriously restricted the application effect of 3D laser scanning technology in building deformation monitoring. Summary of the invention

[0005] The purpose of the invention is to propose a building deformation analysis method and system based on three-dimensional laser scanning and deep learning to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution is a building deformation analysis method based on 3D laser scanning and deep learning, which includes the following steps:

[0007] S1. Obtain the original point cloud data, and use the feature decomposition method to perform hierarchical processing based on the coordinate information and intensity value of the original point cloud data to obtain hierarchical feature data; fuse the hierarchical feature data with the pre-stored environmental parameter data to obtain environmental compensation feature data; perform quality assessment and optimization processing on the environmental compensation feature data to obtain optimized point cloud data; fuse the optimized point cloud data with the pre-stored historical monitoring data for spatiotemporal features to obtain high-fidelity point cloud data and spatiotemporal feature description data;

[0008] S2. Input the high-fidelity point cloud data and spatiotemporal feature description data into a preconfigured component segmentation model to generate component feature data and component relationship data; perform deformation feature extraction on the component feature data and component relationship data to obtain deformation feature data and feature weight data; perform multi-scale analysis based on the deformation feature data and feature weight data to obtain fused deformation feature data; perform consistency optimization on the fused deformation feature data and the spatiotemporal feature description data to obtain optimized deformation feature data and feature credibility data;

[0009] S3, inputting the optimized deformation feature data and feature credibility data into the hybrid neural network for processing to obtain hybrid feature data and network status data; performing multi-scale feature enhancement on the hybrid feature data and network status data to generate enhanced feature data and reliability scoring data; performing deformation evolution analysis based on the enhanced feature data and reliability scoring data, and outputting deformation prediction data and uncertainty data; using an early warning decision system based on the deformation prediction data and uncertainty data to generate early warning level data and decision recommendation data;

[0010] S4. Verify the warning level data, decision recommendation data and optimized deformation feature data according to physical laws to generate physical verification data and first verification correction data; perform multi-source data cross-validation on the physical verification data and spatiotemporal feature description data to obtain cross-validation data and second verification correction data; optimize spatiotemporal continuity based on the cross-validation data and deformation prediction data, and output continuity optimization data and third verification correction data; comprehensively optimize the continuity optimization data, first verification correction data, second verification correction data and third verification correction data to generate final optimization data and reliability report data.

[0011] Building deformation analysis system based on 3D laser scanning and deep learning, including:

[0012] at least one processor; and,

[0013] a memory communicatively connected to at least one of the processors; wherein,

[0014] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the building deformation analysis method based on three-dimensional laser scanning and deep learning.

[0015] Beneficial effects: The present invention improves the quality of data and realizes precise deformation monitoring; improves the accuracy of prediction and ensures the reliability of results; can provide more accurate and reliable building deformation monitoring and early warning services in complex engineering environments, forming a complete, adaptive and reliable building deformation monitoring and early warning system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flow chart of the method of the present invention.

[0017] Figure 2 This is a flow chart of step S1 of the present invention.

[0018] Figure 3 This is a flow chart of step S2 of the present invention.

[0019] Figure 4 This is a flow chart of step S3 of the present invention.

[0020] Figure 5 This is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 As shown, this application proposes a building deformation analysis method based on 3D laser scanning and deep learning, comprising the following steps:

[0022] S1. Obtain the original point cloud data, and use the feature decomposition method to perform hierarchical processing based on the coordinate information and intensity value of the original point cloud data to obtain hierarchical feature data; fuse the hierarchical feature data with the pre-stored environmental parameter data to obtain environmental compensation feature data; perform quality assessment and optimization processing on the environmental compensation feature data to obtain optimized point cloud data; fuse the optimized point cloud data with the pre-stored historical monitoring data for spatiotemporal features to obtain high-fidelity point cloud data and spatiotemporal feature description data;

[0023] S2. Input the high-fidelity point cloud data and spatiotemporal feature description data into a preconfigured component segmentation model to generate component feature data and component relationship data; perform deformation feature extraction on the component feature data and component relationship data to obtain deformation feature data and feature weight data; perform multi-scale analysis based on the deformation feature data and feature weight data to obtain fused deformation feature data; perform consistency optimization on the fused deformation feature data and the spatiotemporal feature description data to obtain optimized deformation feature data and feature credibility data;

[0024] S3, inputting the optimized deformation feature data and feature credibility data into the hybrid neural network for processing to obtain hybrid feature data and network status data; performing multi-scale feature enhancement on the hybrid feature data and network status data to generate enhanced feature data and reliability scoring data; performing deformation evolution analysis based on the enhanced feature data and reliability scoring data, and outputting deformation prediction data and uncertainty data; using an early warning decision system based on the deformation prediction data and uncertainty data to generate early warning level data and decision recommendation data;

[0025] S4. Verify the warning level data, decision recommendation data and optimized deformation feature data according to physical laws to generate physical verification data and first verification correction data; perform multi-source data cross-validation on the physical verification data and spatiotemporal feature description data to obtain cross-validation data and second verification correction data; optimize spatiotemporal continuity based on the cross-validation data and deformation prediction data, and output continuity optimization data and third verification correction data; comprehensively optimize the continuity optimization data, first verification correction data, second verification correction data and third verification correction data to generate final optimization data and reliability report data.

[0026] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0027] S11, receiving original point cloud data including three-dimensional coordinates and intensity values, and constructing a feature tensor; recursively decomposing the feature tensor into multiple levels, and setting corresponding feature weight coefficients for each level, to obtain a multi-level feature tensor with weight coefficients; based on the multi-level feature tensor, respectively extracting structural features, local features and noise features, and generating hierarchical feature data;

[0028] S12, obtaining environmental parameter data, constructing an environmental impact response matrix based on the hierarchical feature data and the environmental parameter data, calculating the degree of influence of each environmental factor on different features; based on the degree of influence, using a dynamic threshold method to distinguish environmental interference and actual deformation; based on the environmental interference and actual deformation, performing environmental impact compensation calculation, and outputting environmental compensation feature data;

[0029] S13. Based on the environmental compensation feature data, a spatial topological structure tensor is constructed to generate local geometric constraints; based on the local geometric constraints, a loss function is constructed; based on the loss function, an iterative optimization method is used to complete the data and improve the data quality to generate optimized point cloud data;

[0030] S14. Obtain historical monitoring data, construct a time series feature matrix based on the historical monitoring data and optimized point cloud data, and calculate the weight coefficient of current data and historical data; based on the weight coefficient, use the feature fusion method to enhance spatiotemporal consistency and output high-fidelity point cloud data and spatiotemporal feature description data.

[0031] In one embodiment of the present application, an original point cloud data set P = {p1, p2, ..., pn} is obtained, where pi = (xi, yi, zi, Ii), representing the three-dimensional coordinates and intensity value of the i-th point; a hierarchical feature decomposition algorithm (Hierarchical Feature Decomposition, HFD) is used to construct a feature tensor F to map the point cloud data to a multidimensional feature space; an adaptive feature weight matrix W is constructed, W = [wij]m×n, where wij represents the importance of the i-th feature at the j-th level; the data is decomposed into structural features Fs, local features Fl and noise features Fn through a recursive feature decomposition function φ(F, W); and a hierarchical feature set {Fs, Fl, Fn} is output.

[0032] Based on the hierarchical feature set {Fs, Fl, Fn} and the environmental parameter set E = {e1, e2, ..., ek} (temperature, humidity, light, etc.), a multi-modal environmental impact separator (MES) is used to construct the environmental impact response matrix R = [rij]k×m, where rij represents the influence of the i-th environmental factor on the j-th feature; a dynamic threshold function τ(R, E) is constructed to adaptively distinguish environmental interference from actual deformation; the environmental impact is eliminated through the compensation function C(F, R, E), where F is the feature matrix; and the feature set F' and the environmental impact mapping matrix M after environmental compensation are output.

[0033] Based on the feature set F' after environmental compensation, a topology-preserving reconstruction network (TPRN) is used to construct a topological structure tensor T to describe the spatial relationship between points. A loss function L(F', T) based on local geometric constraints is constructed, and data completion and quality improvement are achieved through an iterative optimization algorithm Ω(F', T, L), outputting a high-quality point cloud dataset P' and a credibility score matrix S.

[0034] Based on the high-quality point cloud dataset P' and the historical dataset H={h1, h2, ..., hm}, the spatial-temporal consistency enhancer (STCE) is used to construct the time series feature matrix Γ to describe the temporal pattern of building deformation. An adaptive weight function ω(P', H) is constructed to balance the importance of current data and historical data. The consistency of spatial-temporal features is enhanced through the fusion function ψ(P', H, Γ, ω), and the final processed high-fidelity point cloud dataset P* and spatial-temporal feature descriptor D are output.

[0035] This embodiment achieves the acquisition and processing of high-quality point cloud data by adopting the method of hierarchical feature extraction and environmental compensation. Specifically, hierarchical processing is performed through the feature decomposition method, so that the point cloud data can be effectively analyzed at different scales, so that the macro-structural characteristics and micro-deformation characteristics of the building can be captured at the same time. In terms of environmental compensation, an environmental impact response matrix is ​​established. By correlating and analyzing environmental parameters such as temperature, humidity, and light with the features of point cloud data, the interference of environmental factors on the measurement results is effectively identified and eliminated. The dynamic threshold method is used to distinguish between environmental interference and actual deformation to ensure the accuracy of the data. At the same time, by fusing the spatiotemporal features with historical monitoring data, not only the reliability of the current measurement data is improved, but also a time evolution model of building deformation is established. This embodiment makes the point cloud data finally obtained have higher accuracy and reliability, especially under complex environmental conditions, it can still maintain a stable monitoring effect, and improve the accuracy and reliability of building deformation monitoring.

[0036] According to one aspect of the present application, step S11 is further:

[0037] S111, receiving original point cloud data, constructing a local neighborhood covariance matrix, calculating the main eigenvector and eigenvalue, generating local feature data; performing nonlinear transformation on the local feature data, constructing a high-dimensional feature map, and outputting feature tensor data;

[0038] S112, based on the feature tensor data, calculate the hierarchical weights through an adaptive feature importance evaluation function to generate initial weight data; perform multi-scale consistency optimization on the initial weight data, and output feature weight data;

[0039] S113, based on the feature tensor data and the feature weight data, construct a recursive neural network structure, perform feature decomposition operation, and generate hierarchical decomposition data; perform feature correlation analysis on the hierarchical decomposition data, and output multi-level feature data;

[0040] S114. Based on the multi-level feature data, a feature clustering algorithm is used to classify the features of each level to obtain feature category data; the feature category data is subjected to inter-level correlation analysis and integration to output hierarchical feature data.

[0041] In one embodiment of the present application, the characteristic decomposition operation is performed specifically as follows: construct a feature tensor T(x, y, z) = ∑[α_i·F_i(x, y, z) + β_i·G_i(x, y, z) + γ_i·H_i(x, y, z)]; calculate the feature importance R(i) = λ1·E(i) + λ2·V(i) + λ3·C(i); perform recursive decomposition D(k+1) = W·D(k)·S(k) +ε(k); where: F_i(x, y, z) is the spatial structure feature function, which represents the geometric features of the point cloud at the spatial position (x, y, z); G_i(x, y, z) is the local deformation feature function, which represents the deformation features of the local area; H_i(x, y, z) is the noise feature function, which represents the non-structural change features; α_i, β_i, γ_i are feature weight coefficients used to balance the contribution of different features; E(i) is the feature entropy, which represents the amount of information of the feature; V(i) is the variance contribution rate, which represents the discrimination of the feature; C(i) is the spatial correlation, which represents the spatial distribution characteristics of the feature; λ1, λ2, λ3 are importance weight coefficients, which are used to adjust the importance of different evaluation indicators; W is the recursive weight matrix, which controls the feature transfer strength; S(k) is the scale factor, which changes with the recursive level k; ε(k) is the residual term, which represents the feature components that have not been decomposed; D(k) is the decomposition result of the kth layer.

[0042] This embodiment realizes efficient processing and feature expression of point cloud data through feature extraction and hierarchical decomposition. First, by constructing a local neighborhood covariance matrix and a nonlinear feature map, the system can capture the local geometric features and global structural features of point cloud data, which is particularly suitable for processing complex geometric shapes of buildings and can accurately describe the features of each component of the building. In terms of feature weight calculation, an adaptive feature importance evaluation function is adopted, and multi-scale consistency optimization is performed to ensure that the weight distribution of features at different levels is more reasonable. In particular, in the feature decomposition process of the recursive neural network, the system can automatically learn and adapt to the feature expression of different levels, and extract the most representative multi-level features through feature correlation analysis. Finally, through feature clustering and inter-level correlation analysis, the system realizes automatic classification and integration of features, which not only improves the efficiency of feature extraction, but also ensures the integrity and accuracy of feature expression, providing a reliable data basis for subsequent deformation analysis.

[0043] According to one aspect of the present application, step S12 is further:

[0044] S121, obtaining environmental parameter data, constructing a spatiotemporal continuous field based on the environmental parameter data using an interpolation algorithm, and generating environmental parameter field data; calculating a correlation coefficient matrix based on each characteristic component in the hierarchical characteristic data and the environmental parameter field data, and generating correlation coefficient matrix data; performing singular value decomposition based on the correlation coefficient matrix data, determining the main impact mode, and outputting an environmental impact response matrix;

[0045] S122. Based on the environmental impact response matrix and the hierarchical characteristic data, matrix operations are performed to calculate the impact intensity coefficient of each environmental factor and generate environmental impact intensity data; a probability distribution model based on kernel density estimation is constructed to perform statistical analysis on the environmental impact intensity data and output environmental impact distribution data; based on the environmental impact distribution data, an adaptive threshold function is constructed to generate dynamic threshold data;

[0046] S123, comparing and analyzing the variation of the hierarchical feature data with the dynamic threshold data, identifying deformation features beyond the environmental impact range, and generating actual deformation data; constructing a compensation model based on the environmental impact response matrix and the environmental impact intensity data, calculating the environmental impact compensation amount, and outputting environmental compensation coefficient data; performing feature reconstruction operation based on the actual deformation data and the environmental compensation coefficient data, and generating compensated reconstruction data;

[0047] S124. Based on the compensation reconstruction data, a physical constraint-based optimization algorithm is used to generate compensation optimization data; the compensation optimization data is checked for consistency with pre-stored historical compensation results, and compensation consistency data is output; based on the compensation consistency data, the compensation optimization data is fine-tuned, and finally environmental compensation characteristic data is output.

[0048] In one embodiment of the present application, the construction of the environmental impact response matrix is ​​specifically as follows: constructing the environmental impact function M(p, e) = ∑[μ_i·P_i(p, t)·E_i(e, t)]; calculating the response intensity I(p, e) =θ1·L(p, e) +θ2·N(p, e) +θ3·Q(p, e); generating the response matrix R(t) = U(t)·I(p, e)·V(t) T+η(t); where: P_i(p, t) is the point cloud feature function, which represents the change of the point cloud feature at position p with time t; E_i(e, t) is the environmental parameter function, which represents the change of environmental factor e with time t; μ_i is the environmental impact weight, which represents the degree of influence of different environmental factors; L(p, e) is the linear response term, which represents the direct influence of environmental factors; N(p, e) is the nonlinear response term, which represents the coupling effect of environmental factors; Q(p, e) is the lag response term, which represents the delayed effect of environmental influence; θ1, θ2, θ3 are response weight coefficients, which are used to balance different response characteristics; U(t) and V(t) are time-varying feature vectors, which describe the dynamic changes of response characteristics; η(t) is the noise term, which represents the unmodeled environmental influence.

[0049] This embodiment realizes the accurate quantification of the impact of environmental factors on point cloud data by constructing a spatiotemporal continuous environmental parameter field and environmental impact response matrix. The system first uses an interpolation algorithm to establish a continuous distribution field of environmental parameters to ensure the spatial integrity of environmental data; then the main impact modes are extracted through the correlation coefficient matrix and singular value decomposition, and an environmental impact response matrix is ​​established to achieve an accurate association between environmental factors and deformation features. Using kernel density estimation and adaptive threshold methods, the system can accurately distinguish between environmental interference and actual deformation, avoiding false deformation identification caused by environmental factors. By establishing a compensation model and performing feature reconstruction, combined with physical constraint optimization and consistency verification, accurate compensation for environmental impacts is ultimately achieved, improving the accuracy of deformation monitoring.

[0050] According to one aspect of the present application, step S13 is further:

[0051] S131. Acquire environmental compensation feature data, establish local topological connection relationships through neighborhood point set analysis, and generate topological structure data; calculate local geometric features of the topological structure data, including curvature and normal vectors, and output geometric feature data; construct a spatial topological structure tensor based on the geometric feature data to generate structure tensor data.

[0052] S132, receiving structure tensor data, constructing a loss function according to local geometric consistency constraints, and generating loss function data; combining the loss function data with preset physical constraints, constructing an optimization objective function, and outputting optimization function data; performing parameter sensitivity analysis on the optimization function data, and generating parameter sensitivity data.

[0053] S133. Based on the optimization function data and parameter sensitivity data, perform iterative optimization calculations to generate optimization iteration data for each iteration; perform convergence analysis on the optimization iteration data, determine the optimal iteration step, and output convergence parameter data; dynamically adjust the optimization iteration data in combination with the convergence parameter data to generate optimization result data.

[0054] S134, obtaining optimization result data, evaluating the optimization effect through residual analysis, and generating residual evaluation data; performing statistical analysis on the residual evaluation data, calculating the optimization credibility, and outputting credibility data; performing quality evaluation and screening on the optimization result data based on the credibility data, and finally outputting the optimized point cloud data.

[0055] This embodiment achieves high-quality optimization of point cloud data by establishing local topological structure and geometric feature analysis. The system first establishes topological connection relationships through neighborhood analysis, calculates local geometric features, and constructs a complete spatial topological structure tensor, providing a reliable mathematical basis for subsequent optimization; then, a loss function is constructed based on local geometric consistency constraints, and an optimization objective function is established by combining physical constraints and parameter sensitivity analysis. Through iterative optimization calculations and convergence analysis, the system can adaptively adjust optimization parameters to ensure the stability of the optimization process. Finally, the optimization effect is evaluated through residual analysis and statistical methods, and the quality of point cloud data is improved.

[0056] According to one aspect of the present application, step S14 is further:

[0057] S141. Receive optimized point cloud data and historical monitoring data, construct a time window sequence, calculate a time series feature matrix, and obtain window feature data; perform time series pattern decomposition on the window feature data, extract key time series features through principal component analysis, and output time series feature data.

[0058] S142, obtaining time series feature data, calculating the feature sequence similarity through a dynamic time warping algorithm, and generating similarity data; performing adaptive weight allocation on the similarity data, combining time series correlation analysis, and outputting time series weight data.

[0059] S143, receiving time series feature data and time series weight data, constructing a spatiotemporal feature fusion network, performing weighted feature fusion operations, generating initial fused data; performing residual feature compensation on the initial fused data, and outputting fused feature data.

[0060] S144, obtaining fused feature data, performing feature verification through a spatiotemporal consistency constraint function, and obtaining consistency score data; optimizing and adjusting the consistency score data and fused feature data, and finally outputting high-fidelity point cloud data and spatiotemporal feature description data.

[0061] This embodiment realizes the spatiotemporal fusion of point cloud data and historical monitoring data. By constructing a time window sequence and a time series feature matrix, the system can effectively capture the time evolution characteristics of building deformation. In particular, in terms of time series pattern decomposition, the key time series features are extracted through principal component analysis, and the main patterns and trends of building deformation are successfully identified. In the calculation of feature similarity, the dynamic time warping algorithm is adopted, which can effectively handle the nonlinear deformation between different time series, making the alignment of time series features more accurate. Through adaptive weight allocation and time series correlation analysis, the system can dynamically adjust the importance of data at different time points, effectively reducing the influence of noise data. In the spatiotemporal feature fusion link, through the residual feature compensation mechanism, the system not only improves the accuracy of the fusion result, but also maintains the continuity of the spatiotemporal features. Finally, through the verification of the spatiotemporal consistency constraint function, the physical rationality of the fusion result is ensured, and the high-fidelity point cloud data and spatiotemporal feature description data finally output provide reliable data support for the accurate analysis of building deformation.

[0062] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0063] S21. Based on the high-fidelity point cloud data and the spatiotemporal feature description data, a spatial feature pyramid is constructed; regional growth calculation is performed on the data in the spatial feature pyramid to obtain component region identification; based on the component region identification, a topological relationship diagram between components is established, and the organizational relationship between components is constructed through hierarchical indexing, and component feature data and component relationship data are output;

[0064] S22, constructing a deformation feature tensor based on the component feature data and the component relationship data; based on the deformation feature tensor, extracting the deformation feature of the component through geometric invariant calculation; based on the deformation feature, extracting and fusing local deformation information and global deformation information, calculating feature significance scores, and generating deformation feature data and feature weight data;

[0065] S23, constructing a multi-scale observation window based on the deformation feature data and feature weight data; based on the multi-scale observation window, performing aggregation calculation on deformation features at different scales to construct a deformation pattern correlation matrix; based on the deformation pattern correlation matrix, highlighting key deformation patterns through feature enhancement operations, and outputting fused deformation feature data;

[0066] S24. Based on the fused deformation feature data and the spatiotemporal feature description data, a spatiotemporal feature tensor is constructed; based on the spatiotemporal feature tensor, a consistency function based on physical constraints is constructed, and feature optimization calculations are performed to generate optimized deformation feature data and feature credibility data.

[0067] In one embodiment of the present application, based on a high-fidelity point cloud dataset P* and a spatiotemporal feature descriptor D, a hierarchical component segmentation network (HCSN) is used to construct a spatial feature pyramid Ψ={ψ1, ψ2, ..., ψk}, where ψi represents the feature expression of the i-th layer; an adaptive region growing algorithm G(Ψ) is constructed to perform component segmentation based on local similarity; the topological relationship between components is described by a component relationship graph R(V, E), where V is a component set and E is a relationship edge set; a hierarchical component index tree T(V, H) is constructed, where H represents the hierarchical relationship between components, and the component feature set C={c1, c2, ..., cm} and the component relationship graph R are output.

[0068] Based on the component feature set C and the component relationship graph R, a deformation-sensitive feature extractor (DSFE) is used to construct a deformation feature tensor D=[dij]m×n, where dij represents the j-th deformation feature of the ith component. A feature function φ(C) based on geometric invariants is constructed to extract rotational and translational invariant deformation features. The local and global deformation information is fused through the feature enhancement function E(D, R). A feature significance evaluation function S(D) is constructed to highlight key deformation features, and the deformation feature set V={v1, v2, ..., vk} and feature importance weight W are output.

[0069] Based on the deformation feature set V and feature importance weight W, a multi-scale deformation pattern analyzer (MDPA) is used to construct a multi-scale observation window set Ω={ω1, ω2, ..., ωs} to capture deformation patterns of different spatial scales. A scale-adaptive feature aggregation function A(V, Ω) is constructed to achieve cross-scale feature fusion. The pattern correlation matrix M=[mij]k×k is used to describe the correlation between different deformation features, and a feature enhancement function F(V, M, W) is constructed to enhance the key deformation patterns. The fused deformation feature set F and the pattern correlation matrix M are output.

[0070] Based on the fusion deformation feature set F and spatiotemporal feature descriptor D, the spatial-temporal consistency optimizer (Spatial-Temporal Consistency Optimizer, STCO) is used to construct the spatiotemporal feature tensor T to comprehensively express the deformation features in spatial and temporal dimensions; a consistency function C(F, T) based on physical constraints is constructed to ensure that the deformation features conform to physical laws, and the spatiotemporal consistency of the features is enhanced by optimizing the function O(T, C); a credibility evaluation function R(T) is constructed to quantify the feature reliability and output the final deformation feature set F* and credibility matrix Q.

[0071] This embodiment realizes accurate deformation analysis at the building component level by combining component segmentation and deformation feature extraction technology. First, by constructing a spatial feature pyramid and a regional growth algorithm, intelligent segmentation of point cloud data is realized, and different building components are accurately identified. Not only the accuracy of segmentation is improved, but also the topological relationship between components is maintained. In terms of deformation feature extraction, through geometric invariant calculation and multi-scale analysis, local deformation information and global deformation features can be obtained simultaneously, overcoming the limitation of traditional methods that only focus on single-scale deformation. In particular, in the calculation of feature weights, an adaptive weight allocation mechanism is adopted to dynamically adjust feature weights according to the importance of different components and the significance of deformation features, which enables the system to more accurately identify key deformation areas. Through multi-scale feature fusion and consistency optimization based on physical constraints, the deformation features finally obtained have higher reliability and physical rationality, providing a more reliable data basis for subsequent deformation warnings.

[0072] According to one aspect of the present application, step S21 is further:

[0073] S211. Based on the high-fidelity point cloud data, a feature pyramid network is constructed through multi-scale analysis to generate pyramid feature data of different scales; feature fusion is performed on the pyramid feature data and the spatiotemporal feature description data to output multi-scale fusion data; based on the multi-scale fusion data, a local feature descriptor is calculated to generate feature description data;

[0074] S212, based on the feature description data, using a region growing algorithm, calculating the point cloud segmentation seed points, and generating seed point data; based on the seed point data, performing an adaptive region growing operation, and outputting the region growing data; performing boundary optimization processing on the region growing data, and generating segmentation region data;

[0075] S213, based on the segmented area data, using a graph theory method, establishing a connection relationship diagram between components, and generating connection relationship data; performing a topological relationship analysis on the connection relationship data, determining the spatial position relationship between components, and outputting the topological relationship data; based on the topological relationship data, calculating the interaction strength between components, and generating interaction strength data;

[0076] S214. Based on the segmented area data and topological relationship data, a hierarchical index structure is constructed to generate index structure data; the index structure data is combined with the action intensity data to establish a component organizational relationship and output the organizational relationship data; the organizational relationship data is optimized for consistency and finally the component feature data and component relationship data are output.

[0077] This embodiment achieves accurate component segmentation of point cloud data through a multi-scale feature pyramid and region growing algorithm. The system first constructs a feature pyramid network, integrates spatiotemporal feature descriptions, generates multi-scale feature expressions, and provides rich feature information for component segmentation; then, it performs point cloud segmentation through an adaptive region growing algorithm, accurately identifies component boundaries, and establishes a complete component division system. The topological relationship and interaction strength between components are analyzed using graph theory methods, a hierarchical index structure is constructed, and an accurate description of the relationship between components is achieved. This segmentation method based on multi-scale analysis and hierarchical organization improves the accuracy and reliability of component identification.

[0078] According to one aspect of the present application, step S22 is further:

[0079] S221, receiving component feature data and component relationship data, constructing a multi-dimensional deformation description matrix, calculating relative deformation parameters between components, and obtaining relative deformation data; performing tensor mapping transformation on the relative deformation data, and outputting initial deformation feature data through nonlinear feature extraction.

[0080] S222, obtaining initial deformation feature data, calculating geometric features through a rotation invariance operator, and generating rotation feature data; performing a scale invariance transformation on the rotation feature data, combining with a displacement invariance analysis, and outputting geometric feature data.

[0081] S223, receiving geometric feature data, constructing a local feature aggregation network, performing neighborhood feature fusion, and generating local deformation data; integrating the local deformation data with global structural features at multiple scales, and outputting fused deformation data.

[0082] S224, obtaining fused deformation data, calculating feature importance through deformation significance evaluation function, and obtaining significance score data; performing feature screening and optimization on the significance score data and fused deformation data, and finally outputting deformation feature data and feature weight data.

[0083] This embodiment realizes accurate deformation analysis at the building component level. First, by constructing a multi-dimensional deformation description matrix, the system can fully capture the relative deformation relationship between components, which is particularly suitable for dealing with the mutual influence between components in complex building structures. In terms of geometric feature extraction, rotation invariance operators, scale invariance transformations and displacement invariance analysis are introduced. This multiple invariance feature extraction method ensures that consistent deformation features can be obtained at different observation angles and scales. Through the design of a local feature aggregation network, the system can adaptively fuse neighborhood feature information and effectively capture local deformation patterns. At the same time, through multi-scale integration with global structural features, the integrity of deformation analysis is ensured. In particular, in the feature importance evaluation link, through the deformation significance evaluation function, the system can automatically identify key deformation areas, and through the feature screening optimization mechanism, retain the most representative deformation features. This embodiment not only improves the accuracy of deformation detection, but also provides reliable feature data support for subsequent early warning analysis.

[0084] According to one aspect of the present application, step S23 is further:

[0085] S231, obtain deformation feature data, construct observation windows of multiple scales through the sliding window method, and generate observation window data; calculate local statistical features for each observation window data, and output window feature data; perform weighted processing on the window feature data in combination with feature weight data to generate weighted window data.

[0086] S232, receiving weighted window data, and using a hierarchical clustering method to aggregate features of different scales to generate feature clustering data; performing principal component analysis on the feature clustering data to extract the main deformation mode, and outputting deformation mode data; and calculating a mode weight coefficient based on the deformation mode data to generate mode weight data.

[0087] S233. Perform correlation analysis on the deformation pattern data and pattern weight data, calculate the correlation strength between patterns, and generate correlation strength data; use the correlation strength data to construct a deformation pattern correlation matrix and output the correlation matrix data; identify key deformation patterns based on the correlation matrix data and generate key pattern data.

[0088] S234, obtaining key pattern data, performing pattern enhancement processing through a feature enhancement network, and generating enhanced pattern data; comparing and verifying the enhanced pattern data with the original deformation feature data, and outputting verification result data; optimizing and adjusting the enhanced pattern data based on the verification result data, and finally outputting fused deformation feature data.

[0089] In one embodiment of the present application, constructing the deformation mode correlation matrix is ​​specifically as follows: constructing the correlation function C(m,n) = ∑[ψ_i·X_i(m)·Y_i(n)]; calculating the mode strength S(m,n) = κ1·P(m,n) +κ2·R(m,n) +κ3·H(m,n); generating the correlation matrix M(t) = W(t)·S(m,n)·Z(t) T +ε(t); where: X_i(m), Y_i(n) are deformation mode characteristic functions; ψ_i is the mode weight coefficient; P(m, n) is the position correlation; R(m, n) is the morphological correlation; H(m, n) is the historical correlation; κ1, κ2, κ3 are intensity weight coefficients; W(t), Z(t) are time-varying eigenvectors; ε(t) is the residual term.

[0090] This embodiment realizes multi-scale analysis and enhancement of deformation features through sliding window and hierarchical clustering methods. The system first uses the sliding window method to construct a multi-scale observation window, and combines the weight to calculate the local statistical features to obtain a complete multi-scale feature expression; then, the main deformation mode is extracted through hierarchical clustering and principal component analysis, the mode weight coefficient is determined, and a hierarchical expression of deformation features is established. Through correlation analysis and correlation matrix construction, the system can accurately identify key deformation modes, and enhance the mode through the feature enhancement network, and finally realize the effective fusion of deformation features, improving the accuracy and robustness of deformation analysis.

[0091] According to one aspect of the present application, step S24 is further:

[0092] S241. Receive fused deformation feature data and spatiotemporal feature description data, construct a spatiotemporal feature tensor through a tensor decomposition method, and generate spatiotemporal tensor data; perform a rank decomposition operation on the spatiotemporal tensor data, extract the main characteristic modes, and output characteristic mode data; calculate the spatiotemporal correlation index based on the characteristic mode data, and generate correlation index data.

[0093] S242. Acquire characteristic modal data, construct a physical constraint function according to the principles of building structure mechanics, and generate physical constraint data; combine the physical constraint data with the correlation index data, establish a consistent optimization objective function, and output the optimization target data; perform gradient calculation on the optimization target data to generate gradient characteristic data.

[0094] S243. Based on the optimization target data and gradient feature data, perform iterative optimization calculations to generate optimization iteration data; perform convergence analysis on the optimization iteration data, determine the optimal iteration step, and output convergence control data; dynamically adjust the optimization process according to the convergence control data to generate optimization result data.

[0095] S244, receiving the optimization result data, evaluating the feature reliability through the information entropy method, and generating reliability score data; performing threshold segmentation on the reliability score data, identifying high-credibility features, and outputting feature screening data; making final adjustments to the optimization result data in combination with the feature screening data, and outputting the optimized deformation feature data and feature credibility data.

[0096] In one embodiment of the present application, constructing a physical constraint function is specifically as follows: constructing a loss function L(x) = ∑[φ1·G(x) + φ2·B(x) + φ3·C(x)]; calculating the optimization gradient ▽L(x) = ω1·▽G(x) +ω2·▽B(x) +ω3·▽C(x); performing iterative optimization x(t+1) = x(t) - η·▽L(x) + δ(t); wherein G(x) is a geometric consistency loss term, indicating the continuity of local geometric features; B(x) is a boundary constraint loss term, indicating the degree of satisfaction of physical boundary conditions; C(x) is a continuity loss term, indicating the smoothness of feature distribution; φ1, φ2, φ3 are loss weight coefficients; ω1, ω2, ω3 are gradient weight coefficients; η is a learning rate; δ(t) is a dynamic correction term; x(t) is the result of the tth iteration.

[0097] This embodiment achieves consistency optimization of deformation features through space-time feature tensor decomposition and physical constraint optimization. The system first extracts the main feature modes through tensor decomposition, calculates the space-time correlation index, and establishes a complete feature expression; then constructs a physical constraint function based on the principles of building structural mechanics, combines the space-time correlation to design the optimization target, and achieves the physical rationality of feature optimization. Through iterative optimization and convergence analysis, the system can dynamically adjust the optimization process to ensure the stability of the results. Finally, the information entropy method is used to evaluate the reliability of the features, and the screening of high-confidence features is achieved, ensuring the reliability of the optimization results.

[0098] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0099] S31, constructing a feature conversion matrix based on the optimized deformation feature data and feature credibility data; based on the feature conversion matrix, inputting the optimized deformation feature data and feature credibility data into the spatial feature extraction network, the temporal feature extraction network and the physical constraint network respectively, and calculating the sub-network weight coefficients; based on the sub-network weight coefficients, performing feature interactive fusion through the hybrid layer, and outputting the hybrid feature data and network status data;

[0100] S32. Based on the hybrid feature data and the network status data, a multi-scale feature pyramid is constructed, and the cross-scale attention weight is calculated; based on the cross-scale attention weight, a feature enhancement operation is performed to generate enhanced feature data; based on the enhanced feature data, feature reliability verification is performed to obtain reliability score data;

[0101] S33, based on the enhanced feature data and the reliability score data, construct a time series feature sequence and generate a nonlinear dynamic model; based on the nonlinear dynamic model, perform multi-path prediction calculation to obtain deformation prediction data; based on the deformation prediction data, perform uncertainty assessment to generate uncertainty data;

[0102] S34. Based on the deformation prediction data and uncertainty data, a risk assessment matrix is ​​constructed; based on the risk assessment matrix and pre-stored expert knowledge base data, decision analysis is performed, warning thresholds are calculated, and a graded warning strategy is generated; based on the graded warning strategy, warning level data and decision recommendation data are generated.

[0103] In one embodiment of the present application, based on the final deformation feature set F* and the credibility matrix Q, an adaptive hybrid-connected neural network (AHCNN) is used to construct a feature transformation matrix Θ=[θij]n×m to map the input features to the latent space; a parallel sub-network set N={n1, n2, ..., nk} is constructed: n1 is a spatial feature extraction network that captures spatial dependencies; n2 is a temporal feature extraction network that learns the time evolution pattern; n3 is a physical constraint network that ensures that the prediction conforms to the laws of mechanics; the importance of each sub-network is adjusted through a dynamic weight function W(N, Q), and a hybrid layer L(N, W) is constructed to realize feature interaction and fusion between sub-networks, and the hybrid feature expression H and the network state matrix S are output.

[0104] Based on the hybrid feature expression H and the network state matrix S, a hierarchical feature enhancer (HFE) is used to construct a multi-scale feature pyramid P={p1, p2, ..., pm}, where pi represents the feature of the i-th scale; a cross-scale attention mechanism A(P) is constructed to achieve adaptive fusion of features of different scales, and the key deformation patterns are strengthened through the feature enhancement function E(H, P, A); a feature verification function V(E) is constructed to evaluate the reliability of the features, and the enhanced feature set E* and the reliability score R are output.

[0105] Based on the enhanced feature set E* and reliability score R, the dynamic deformation evolution predictor (DDEP) is used to construct a time series feature sequence T={t1, t2, ..., tk} to describe the time evolution of deformation; a nonlinear dynamic model D(T) is constructed to capture the dynamic characteristics of deformation; multi-path prediction results are generated through the probability prediction function P(E*, D); an uncertainty quantification function U(P) is constructed to evaluate the credibility of the prediction, and the deformation prediction set Y and uncertainty estimate U are output.

[0106] Based on the deformation prediction set Y and uncertainty estimate U, an adaptive warning decision system (AWDS) is used to construct a risk assessment matrix R = [rij] m × n to quantify the risk levels of different types of deformation. A decision function D (Y, U, R) based on evidence theory is constructed to comprehensively evaluate the warning needs. The decision threshold is optimized through the expert knowledge base K. A warning strategy generation function W (D, K) is constructed to output graded warning recommendations. Based on the graded warning recommendations, a warning level evaluation L and a decision recommendation B are output.

[0107] This embodiment adopts a hybrid neural network architecture and multi-scale feature enhancement technology to achieve high-precision deformation prediction and reliable early warning decisions. By hybridizing the spatial feature extraction network, the temporal feature extraction network and the physical constraint network, the system can simultaneously consider information in three dimensions: space, time and physical constraints. This enables the model to better understand the physical nature of building deformation and improve the accuracy of prediction. In terms of feature enhancement, by constructing a multi-scale feature pyramid and calculating cross-scale attention weights, the system can adaptively focus on key deformation features at different scales. In particular, in the deformation evolution analysis, a nonlinear dynamic model is introduced. Through multi-path prediction and uncertainty assessment, not only can the deformation trend be predicted, but also the reliability of the prediction results can be quantified. Finally, through the combination with the expert knowledge base, intelligent early warning decisions are realized. The system can automatically generate hierarchical early warning strategies and decision recommendations based on the prediction results and uncertainty assessment, thereby improving the practicality and reliability of the early warning system.

[0108] According to one aspect of the present application, step S31 is further:

[0109] S311, constructing an orthogonal transformation matrix based on the optimized deformation feature data and the feature credibility data, performing feature projection operation, and obtaining projection feature data; performing nonlinear spatial mapping on the projection feature data, and outputting conversion feature data through dimensional adaptive transformation;

[0110] S312, input the converted feature data into the spatial feature extraction branch, the temporal feature extraction branch and the physical constraint branch respectively, generate spatial feature data, temporal feature data and physical constraint data; preliminarily align the spatial feature data, the temporal feature data and the physical constraint data, and output multi-dimensional feature data;

[0111] S313, based on the multi-dimensional feature data, calculating the weight of each branch through a feature importance evaluation function to obtain initial weight data; performing a dynamic adjustment operation on the initial weight data, and outputting weight adjustment data;

[0112] S314. Based on the multi-dimensional feature data and weight adjustment data, a feature interaction attention network is constructed, cross-branch feature fusion is performed, and fusion state data is generated; the fusion state data is updated for network state and feature integration, and finally mixed feature data and network state data are output.

[0113] This embodiment constructs a hybrid neural network architecture and realizes the effective fusion of multi-dimensional features. By constructing an orthogonal transformation matrix and nonlinear space mapping, the system first realizes the dimensionality reduction and transformation of the feature space. This feature preprocessing method significantly improves the efficiency of subsequent network processing. A three-branch parallel processing architecture is adopted to perform special feature extraction for spatial features, temporal features and physical constraints, respectively, and fully consider the multi-dimensional information in building deformation analysis. Through the feature importance evaluation function and dynamic weight adjustment mechanism, the system can adaptively balance the contribution of different branches to ensure the rationality of feature fusion. In the design of the feature interactive attention network, a cross-branch feature fusion mechanism is introduced, so that different types of features can effectively exchange and complement information. This embodiment not only improves the comprehensiveness of feature extraction, but also realizes the adaptive optimization of feature expression through the dynamic update of the network state, providing high-quality feature input for subsequent deformation prediction.

[0114] According to one aspect of the present application, step S32 is further:

[0115] S321. Receive mixed feature data, use the Laplace pyramid algorithm to build a multi-scale feature pyramid, and generate pyramid level data; perform feature fusion on the pyramid level data and network status data, and output fused level data; calculate the inter-layer feature mapping relationship based on the fused level data, and generate level mapping data.

[0116] S322. Obtain fusion hierarchical data, calculate cross-scale feature weights through the self-attention mechanism, and generate attention weight data; combine the attention weight data with the hierarchical mapping data, construct a feature enhancement matrix, and output the enhanced matrix data; perform feature selection operations on the enhanced matrix data to generate feature selection data.

[0117] S323. Based on the enhancement matrix data and feature selection data, perform feature enhancement operation to generate enhancement result data; perform nonlinear activation processing on the enhancement result data, extract key feature patterns, and output feature pattern data; calculate enhancement effect index based on the feature pattern data to generate enhancement index data.

[0118] S324, receiving enhanced result data, performing feature reliability verification through information entropy and mutual information methods, and generating verification score data; performing threshold screening on the verification score data, determining high reliability features, and outputting reliable feature data; performing feature optimization based on the reliable feature data and enhanced indicator data, and finally outputting enhanced feature data and reliability score data.

[0119] In one embodiment of the present application, constructing a multi-scale feature pyramid is specifically as follows: constructing a pyramid expression P(l) = ∑[ρ_i·K_i(x, l)·F(x)]; calculating the inter-layer mapping M(l, l+1) =σ1·T(l, l+1) +σ2·S(l, l+1) +σ3·D(l, l+1); performing feature fusion F(l) = A(l)·F(l+1)↑+B(l)·F(l-1)↓; wherein K_i(x, l) is a scale kernel function, which represents the filtering effect on the feature at level l; ρ_i is a scale weight coefficient; T(l, l+1) is a feature transfer operator; S(l, l+1) is a feature selection operator; D(l, l+1) is a feature dimensionality reduction operator; σ1, σ2, σ3 are mapping weight coefficients; A(l) and B(l) are inter-layer fusion coefficients; ↑ and ↓ respectively represent upsampling and downsampling operations.

[0120] The specific steps of feature fusion are as follows: construct fusion function F(x) =∑[α_i·S_i(x) +β_i·T_i(x) +γ_i·P_i(x)]; calculate feature weight W(i) =λ1·E(i) +λ2·V(i) +λ3·C(i); perform feature hybridization H(t) =U(t)·F(x)·V(t) T +η(t); where S_i(x) is the spatial feature function; T_i(x) is the temporal feature function; P_i(x) is the physical constraint function; α_i, β_i, γ_i are feature weight coefficients; E(i) is the feature entropy; V(i) is the variance contribution; C(i) is the feature correlation; λ1, λ2, λ3 are weight coefficients; U(t) and V(t) are hybrid matrices; η(t) is the noise term.

[0121] This embodiment achieves effective enhancement of hybrid features through feature pyramid and self-attention mechanism. The system first uses the Laplace pyramid algorithm to construct multi-scale feature expression, integrates network state information, and establishes a complete feature hierarchy structure; then calculates cross-scale feature weights through the self-attention mechanism, constructs a feature enhancement matrix, and achieves adaptive enhancement of features. Through nonlinear activation and key feature extraction, the system can highlight important feature patterns, and verify feature reliability through information entropy and mutual information methods, ultimately achieving high-quality feature enhancement and improving the accuracy of subsequent analysis.

[0122] According to one aspect of the present application, step S33 is further:

[0123] S331. Receive enhanced feature data and reliability score data, construct a multi-scale time window, perform time series feature decomposition, and obtain decomposed sequence data; perform time series pattern mining on the decomposed sequence data, extract trend features, and output time series representation data.

[0124] S332, obtain time series characterization data, construct nonlinear dynamic state equations, calculate system evolution parameters, and generate state parameter data; perform dynamic feature extraction on the state parameter data, combine with stability analysis, and output dynamic feature data.

[0125] S333, receiving dynamic characteristic data, generating multiple evolution paths through Monte Carlo sampling, executing parallel prediction calculations, and obtaining path set data; performing consistency evaluation and optimization screening on the path set data, and outputting predicted path data.

[0126] S334, obtain the predicted path data, calculate the predicted confidence interval through the Bayesian probability model, and obtain the confidence interval data; comprehensively evaluate the confidence interval data and the predicted deviation analysis results, and finally output the deformation prediction data and uncertainty data.

[0127] In one embodiment of the present application, constructing the nonlinear dynamic state equation is specifically as follows: constructing the state equation dx / dt = f(x, t) + g(x, t)u(t) + w(t); calculating the prediction error e(t) = y(t) – y*(t); updating the prediction model x(t+1) = Φ(t)x(t) + Γ(t)u(t) + K(t)e(t); wherein f(x, t) is the system dynamics function; g(x, t) is the control input function; u(t) is the control variable; w(t) is the system noise; y(t) is the actual observation value; y*(t) is the predicted value; Φ(t) is the state transfer matrix; Γ(t) is the input matrix; K(t) is the Kalman gain matrix.

[0128] This embodiment combines time series analysis and dynamic modeling to achieve high-precision deformation prediction. By constructing multi-scale time windows and time series feature decomposition, the system can effectively capture the multi-scale time series characteristics of building deformation. In terms of dynamic modeling, the nonlinear dynamic state equation is introduced, which can more accurately describe the complex evolution process of building deformation. Especially in the prediction path generation link, through the Monte Carlo sampling method, the system can generate multiple possible evolution paths. This multi-path prediction strategy improves the reliability of the prediction results. Through parallel prediction calculation and consistency evaluation, the system can quickly screen out the most likely deformation development trend. In terms of uncertainty assessment, the Bayesian probability model is adopted. By calculating the prediction confidence interval, the system can not only give specific prediction results, but also quantify the uncertainty of the prediction, which has important guiding significance for risk assessment and decision-making in engineering practice.

[0129] According to one aspect of the present application, step S34 is further:

[0130] S341. Receive deformation prediction data and uncertainty data, construct a risk assessment matrix through fuzzy comprehensive evaluation method, and generate risk matrix data; perform hierarchical analysis operation on the risk matrix data, calculate risk level weights, and output risk weight data; construct a risk scoring model based on risk weight data and generate risk scoring data.

[0131] S342. Obtain pre-stored expert knowledge base data, perform knowledge rule matching in combination with risk score data, and generate rule matching data; perform similarity analysis on the rule matching data and historical warning cases, and output similar case data; extract decision rules based on similar case data, and generate decision rule data.

[0132] S343. Calculate warning threshold parameters using risk score data and decision rule data to generate threshold parameter data; perform dynamic adjustment operations on threshold parameter data, consider real-time monitoring status, and output dynamic threshold data; formulate a graded warning strategy based on dynamic threshold data to generate warning strategy data.

[0133] S344. Receive early warning strategy data, generate specific early warning plans through a decision tree model, and generate early warning plan data; conduct a feasibility assessment on the early warning plan data, calculate the priority of plan execution, and output plan priority data; optimize and combine early warning plans based on plan priority data, and finally output warning level data and decision recommendation data.

[0134] This embodiment achieves accurate generation of early warning decisions through fuzzy comprehensive evaluation and hierarchical analysis methods. The system first constructs a risk assessment matrix, calculates risk level weights, and establishes a complete risk scoring system; then combines expert knowledge bases to perform rule matching and similar case analysis, extracts decision rules, and achieves experience-based decision support. Through dynamic threshold adjustment and the formulation of graded early warning strategies, the system can flexibly adjust early warning parameters according to real-time monitoring status, and generate specific early warning plans through decision tree models, ultimately achieving accurate determination of early warning levels and reasonable generation of decision recommendations, improving the reliability and practicality of early warning decisions.

[0135] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0136] S41. Based on the warning level data, the decision suggestion data and the optimized deformation characteristic data, a building mechanics constraint matrix is ​​constructed; based on the building mechanics constraint matrix, a stress-strain relationship verification calculation is performed to obtain physical verification data; based on the physical verification data, an energy conservation test is performed, the physical feasibility score is evaluated, and the first verification correction data is output;

[0137] S42, based on the physical verification data and the spatiotemporal feature description data, construct a multi-source data feature set and calculate the data consistency score; based on the data consistency score, perform anomaly detection operations and generate cross-validation data; based on the cross-validation data, evaluate the trust weight of each data source and generate second verification correction data;

[0138] S43, constructing a spatiotemporal continuity constraint tensor based on the cross-validation data and the deformation prediction data; performing variational optimization calculation based on the spatiotemporal continuity constraint tensor to obtain optimized data; identifying discontinuities in the optimized data, performing adaptive smoothing processing, and obtaining continuous optimized data and third validation correction data;

[0139] S44. Construct a comprehensive decision tree based on the continuous optimization data, the first verification correction data, the second verification correction data and the third verification correction data; based on the comprehensive decision tree, perform conflict resolution calculations to generate optimization results; perform reliability evaluation on the optimization results to obtain evaluation results; based on the evaluation results, perform iterative optimization processing to generate final optimization data and reliability report data.

[0140] In one embodiment of the present application, based on the warning level assessment L, decision recommendation B and deformation feature set F*, a physical law constraint validator (PLCV) is used to construct a building mechanics constraint matrix C=[cij]m×n to describe the physical boundary conditions of deformation development; a stress-strain relationship verification function V(F*, C) is constructed to verify the rationality of deformation and load; the energy balance of deformation development is verified through the energy conservation verification function E(F*, C); a physical feasibility scoring function P(V, E) is constructed to quantify the physical rationality of the results, and the physical verification result matrix P and the correction recommendation set R1 are output.

[0141] Based on the physical verification result matrix P and the spatiotemporal feature descriptor D, a multi-source information cross validator (MCV) is used to construct a multi-source data feature set Ω={ω1, ω2, ..., ωk}, which contains data from different monitoring methods; a data consistency evaluation function C(Ω) is constructed to test the coordination of data from different sources; potential data conflicts are identified through the anomaly detection function A(Ω, C); an evidence weight function W(Ω) is constructed to evaluate the credibility of different data sources, and the cross-validation result set V and the correction suggestion set R2 are output.

[0142] Based on the cross-validation result set V and the deformation prediction set Y, the Spatial-Temporal Continuity Optimizer (STCO) is used to construct the spatiotemporal continuity constraint tensor T to describe the continuity requirements of deformation development; the optimization function O(V, Y, T) based on the variational principle is constructed to ensure the spatiotemporal smoothness of the results; the potential discontinuities are identified through the fracture detection function B(O); and an adaptive smoothing function S(O, B) is constructed to achieve smooth transition while maintaining details, and the optimization result set O and the correction suggestion set R3 are output.

[0143] Based on the optimization result set O and the correction suggestion set {R1, R2, R3}, a comprehensive optimization decision system (CODS) is used to construct a decision tree D to integrate various correction suggestions; a conflict resolution function R(D) is constructed to deal with the contradictions between different correction suggestions; the reliability evaluation function E(O, R) is used to quantify the credibility of the final result; a feedback optimization loop F(E) is constructed to achieve iterative optimization of the results, and the final optimization result F and reliability evaluation report Q are output.

[0144] This embodiment constructs a multi-level verification and optimization mechanism to ensure the reliability and accuracy of the output results of the entire system. First, the physical laws are verified through the building mechanics constraint matrix to ensure that the warning results conform to the principles of building structural mechanics, avoiding the physical irrationality that may be caused by relying solely on data-driven. Secondly, through the multi-source data cross-validation mechanism, the system can comprehensively evaluate the credibility of different data sources, and effectively identify and process abnormal data through data consistency scoring and anomaly detection. In terms of spatiotemporal continuity optimization, a variational optimization method is introduced. Through adaptive smoothing processing, the continuity of the data is maintained and the information loss caused by excessive smoothing is avoided. Finally, through the comprehensive decision tree and iterative optimization processing, the system can automatically resolve possible conflicts between different verification levels and generate the final optimization results and reliability reports. This embodiment improves the credibility of the system output results and provides more reliable decision support for engineering practice.

[0145] According to one aspect of the present application, step S41 is further:

[0146] S411. Obtain optimized deformation feature data, construct a building mechanics constraint matrix based on elastic mechanics theory, and generate mechanical constraint data; perform feature decomposition on the mechanical constraint data, extract the main deformation mode, and output deformation mode data; calculate the mechanical compatibility index by combining the warning level data and the deformation mode data, and generate compatibility data.

[0147] S412, receiving deformation mode data, calculating stress-strain distribution by finite element method, and generating stress-strain data; comparing and analyzing the stress-strain data with theoretical prediction values, and outputting stress deviation data; constructing a verification scoring model based on the stress deviation data and compatibility data, and generating verification scoring data.

[0148] S413. Obtain verification score data and decision recommendation data, perform energy conservation test calculations, and generate energy balance data; perform error analysis on the energy balance data, evaluate energy conservation deviation, and output energy deviation data; calculate the physical feasibility score based on the energy deviation data and generate feasibility score data.

[0149] S414. Based on the verification score data and the feasibility score data, a comprehensive evaluation model is constructed to generate comprehensive evaluation data; threshold analysis is performed on the comprehensive evaluation data to determine the correction requirements and output the correction requirement data; the original data is corrected according to the correction requirement data and finally the physical verification data and the first verification correction data are output.

[0150] In one embodiment of the present application, constructing a building mechanics constraint matrix is ​​specifically as follows: constructing a constraint function C(x) =∑[μ_i·F_i(x) +ν_i·M_i(x) +ξ_i·E_i(x)]; calculating the constraint strength I(x) =θ1·S(x) +θ2·D(x) +θ3·V(x); generating a constraint matrix A(t) = P(t)·I(x)·Q(t) T +ε(t); where F_i(x) is the force balance constraint; M_i(x) is the moment balance constraint; E_i(x) is the energy conservation constraint; μ_i, ν_i, ξ_i are constraint weight coefficients; S(x) is the stress intensity; D(x) is the deformation intensity; V(x) is the vibration intensity; θ1, θ2, θ3 are intensity weight coefficients; P(t) and Q(t) are time-varying eigenvectors; ε(t) is the residual term.

[0151] This embodiment achieves physical verification of the early warning results through architectural mechanics constraints and energy conservation tests. The system first constructs a mechanical constraint matrix based on elastic mechanics theory, extracts deformation patterns, calculates mechanical compatibility indicators, and establishes a complete mechanical verification system; then, the stress-strain distribution is calculated by the finite element method, and the accuracy of stress verification is achieved by comparing the theoretical prediction values. Through energy conservation tests and error analysis, the system can evaluate physical feasibility and make corrections based on a comprehensive evaluation model, ultimately achieving physical rationality verification of the early warning results and improving the reliability of the early warning system.

[0152] According to one aspect of the present application, step S42 is further:

[0153] S421. Receive physical verification data and spatiotemporal feature description data, construct a multi-source data feature space through a feature mapping method, and generate feature space data; perform dimensionality reduction processing on the feature space data, extract common features, and output common feature data; calculate feature distribution parameters based on the common feature data and generate distribution parameter data.

[0154] S422. Acquire common feature data, calculate data consistency score using information entropy method, and generate consistency score data; combine the consistency score data with distribution parameter data, construct a score matrix, and output score matrix data; perform cluster analysis on the score matrix data to generate cluster result data.

[0155] S423. Based on the clustering result data, anomaly detection operation is performed through the isolation forest algorithm to generate anomaly detection data; statistical analysis is performed on the anomaly detection data, anomaly degree index is calculated, and anomaly index data is output; and a data credible interval is determined in combination with the anomaly index data to generate credible interval data.

[0156] S424, receiving anomaly detection data and credible interval data, calculating the credible weight of each data source through Bayesian reasoning, and generating credible weight data; performing normalization processing on the credible weight data, optimizing weight distribution, and outputting weight optimized data; performing weighted fusion on multi-source data based on the weight optimized data, and finally outputting cross-validation data and second validation correction data.

[0157] In one embodiment of the present application, the data consistency score is calculated as follows: construct a consistency function H(d) =∑[π_i·X_i(d) +ρ_i·Y_i(d) +τ_i·Z_i(d)]; calculate the scoring index S(d) =ω1·C(d) +ω2·R(d) +ω3·T(d); generate a scoring matrix E(t) = U(t)·S(d)·V(t) T +η(t); where X_i(d) is the data similarity function; Y_i(d) is the data reliability function; Z_i(d) is the spatiotemporal consistency function; π_i, ρ_i, τ_i are feature weight coefficients; C(d) is the correlation score; R(d) is the reliability score; T(d) is the timeliness score; ω1, ω2, ω3 are scoring weight coefficients; U(t), V(t) are time-varying feature vectors; η(t) is the noise term.

[0158] This embodiment realizes cross-checking of verification results through multi-source data fusion and anomaly detection methods. The system first constructs a multi-source data feature space, extracts common features, calculates feature distribution parameters, and establishes a complete feature expression; then the data consistency score is calculated through the information entropy method, and the data features are identified by cluster analysis to realize the evaluation of data reliability. Through the isolation forest algorithm for anomaly detection and the calculation of trust weights by Bayesian reasoning, the system can accurately identify and process abnormal data, and finally realize the reliable fusion of multi-source data, improving the accuracy of verification results.

[0159] According to one aspect of the present application, step S43 is further:

[0160] S431, receiving cross-validation data and deformation prediction data, constructing a continuity constraint matrix, calculating space-time dimension constraint parameters, and obtaining constraint parameter data; performing tensor mapping conversion on the constraint parameter data, and outputting constraint representation data through multi-dimensional constraint fusion.

[0161] S432, obtaining constraint characterization data, constructing a variational energy function, executing gradient optimization calculations, and generating gradient feature data; performing iterative optimization and solving of the gradient feature data, combining convergence analysis, and outputting optimized feature data.

[0162] S433, receiving optimized feature data, calculating feature discontinuity through singular value decomposition, performing threshold segmentation operation, and obtaining discontinuous point data; performing topological connection analysis on the discontinuous point data, and outputting fracture feature data through regional growing algorithm.

[0163] S434, obtaining fracture characteristic data, constructing an adaptive kernel function, calculating local smoothing parameters, and generating smoothing parameter data; adaptively smoothing and optimizing the smoothing parameter data and fracture area characteristics, and finally outputting continuity optimization data and third verification correction data.

[0164] In one embodiment of the present application, constructing the continuity constraint matrix is ​​specifically as follows: constructing a continuity function L(x, t) = ∑[α_i·S_i(x, t) + β_i·T_i(x, t) + γ_i·C_i(x, t)]; calculating the optimization target O(x, t) = λ1·G(x, t) + λ2·V(x, t) + λ3·E(x, t); performing the optimization iteration x(t+1) = x(t) + η·▽O(x, t) +δ(t); where S_i(x, t) is the spatial continuity function; T_i(x, t) is the temporal continuity function; C_i(x, t) is the constraint continuity function; α_i, β_i, γ_i are continuity weight coefficients; G(x, t) is the gradient term; V(x, t) is the velocity term; E(x, t) is the energy term; λ1, λ2, λ3 are the target weight coefficients; η is the learning rate; δ(t) is the dynamic correction term.

[0165] This embodiment achieves high-quality optimization of deformation data. By constructing a continuity constraint matrix and multi-dimensional constraint fusion, the system first establishes a complete spatiotemporal constraint system. In terms of optimization solution, the variational energy function and gradient optimization calculation are used to effectively handle local discontinuities while maintaining data continuity. Especially in the fracture feature analysis link, through the methods of singular value decomposition and threshold segmentation, the system can accurately identify discontinuous areas in deformation data, which is of great significance for monitoring potential dangerous parts of building structures. Through the design of adaptive kernel functions and local smoothing parameters, the system realizes adaptive smoothing processing for different regions, which not only maintains important deformation features, but also eliminates the influence of noise. The final continuity optimization result not only meets the physical constraint requirements, but also maintains the spatiotemporal continuity of the data, providing reliable data support for engineering decision-making.

[0166] According to one aspect of the present application, step S44 is further:

[0167] S441. Receive continuous optimization data and all verification and correction data, build a multi-level decision model through a decision tree algorithm, and generate decision tree data; perform rule extraction operations on the decision tree data, identify key decision nodes, and output decision rule data; build an optimization strategy matrix based on the decision rule data and generate optimization strategy data.

[0168] S442. Obtain optimization strategy data, use a conflict detection algorithm to identify conflicting items between verification results, and generate conflict item data; sort the conflict item data by level according to the degree of impact, and output conflict level data; formulate a solution based on the conflict level data, and generate solution data.

[0169] S443. Utilize the solution data to perform reliability evaluation calculations of the optimization results and generate reliability evaluation data; perform statistical analysis on the reliability evaluation data, calculate the credibility index, and output the credibility index data; construct an evaluation report template in combination with the credibility index data and generate evaluation report data.

[0170] S444. Based on the evaluation report data and solution data, perform iterative optimization operations to generate iterative optimization data; perform convergence analysis on the iterative optimization data to determine the optimal solution and output the optimal solution data; perform final optimization adjustments based on the optimal solution data and credibility index data, and finally output the final optimization data and reliability report data.

[0171] This embodiment achieves comprehensive optimization of verification results through decision trees and conflict detection methods. The system first builds a multi-level decision model, extracts decision rules, establishes an optimization strategy matrix, and forms a complete optimization strategy system; then, it uses a conflict detection algorithm to identify contradictions between verification results, formulates solutions, and achieves effective conflict handling. Through reliability evaluation and iterative optimization calculations, the system can continuously improve the optimization results, and through the determination and final adjustment of the optimal solution, it finally achieves high-quality optimization of verification results and improves the reliability of system output.

[0172] According to one aspect of the present application, a building deformation analysis system based on 3D laser scanning and deep learning includes:

[0173] at least one processor; and,

[0174] a memory communicatively connected to at least one of the processors; wherein,

[0175] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the building deformation analysis method based on three-dimensional laser scanning and deep learning described in any of the above embodiments.

[0176] The present invention integrates three-dimensional laser scanning and deep learning technology to build a complete building deformation analysis and early warning system. It adopts a hierarchical architecture of "data acquisition-feature extraction-prediction analysis-verification optimization". At the data acquisition level, the quality of the original data is improved through environmental compensation and spatiotemporal feature fusion; at the feature extraction level, more sophisticated deformation monitoring is achieved through multi-scale analysis and component-level deformation feature extraction; at the prediction analysis level, hybrid neural networks and nonlinear dynamic models are used to improve the accuracy of predictions; at the verification optimization level, the reliability of the results is ensured through a multi-level verification mechanism. The present invention can provide more accurate and reliable building deformation monitoring and early warning services in complex engineering environments. Especially when dealing with deformation monitoring of large buildings, the system can take into account both local details and overall deformation trends, providing more comprehensive technical support for building safety management, and forming a complete, adaptive and reliable building deformation monitoring and early warning system.

[0177] 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 building deformation analysis method based on 3D laser scanning and deep learning, characterized in that: The steps include: S1. Obtaining original point cloud data, and using a feature decomposition method to perform hierarchical processing based on the coordinate information and intensity value of the original point cloud data to obtain hierarchical feature data; The layered feature data is integrated and analyzed with the pre-stored environmental parameter data to obtain environmental compensation feature data; the environmental compensation feature data is quality evaluated and optimized to obtain optimized point cloud data; the optimized point cloud data is integrated with the pre-stored historical monitoring data for spatiotemporal features to obtain high-fidelity point cloud data and spatiotemporal feature description data; S2, inputting the high-fidelity point cloud data and the spatiotemporal feature description data into a preconfigured component segmentation model to generate component feature data and component relationship data; performing deformation feature extraction on the component feature data and the component relationship data to obtain deformation feature data and feature weight data; Based on the deformation feature data and feature weight data, multi-scale analysis is performed to obtain fused deformation feature data; Optimize the consistency of the fused deformation feature data and the spatiotemporal feature description data to obtain optimized deformation feature data and feature credibility data; S3, inputting the optimized deformation feature data and feature credibility data into the hybrid neural network for processing to obtain hybrid feature data and network status data; performing multi-scale feature enhancement on the hybrid feature data and network status data to generate enhanced feature data and reliability score data; Based on the enhanced feature data and reliability score data, deformation evolution analysis is performed to output deformation prediction data and uncertainty data; Based on deformation prediction data and uncertainty data, an early warning decision system is used to generate early warning level data and decision recommendation data; S4, verify the warning level data, decision suggestion data and optimized deformation feature data according to physical laws, generate physical verification data and first verification correction data; perform multi-source data cross-validation on the physical verification data and spatiotemporal feature description data to obtain cross-validation data and second verification correction data; perform spatiotemporal continuity optimization based on the cross-validation data and deformation prediction data, and output continuity optimization data and third verification correction data; Comprehensively optimize the continuous optimization data, the first verification correction data, the second verification correction data and the third verification correction data to generate final optimization data and reliability report data; Step S3 is further as follows: S31, constructing a feature conversion matrix based on the optimized deformation feature data and feature credibility data; Based on the feature conversion matrix, the optimized deformation feature data and feature credibility data are respectively input into the spatial feature extraction network, the temporal feature extraction network and the physical constraint network, and the sub-network weight coefficients are calculated; based on the sub-network weight coefficients, the features are interactively fused through the hybrid layer, and the hybrid feature data and network status data are output; S32. Based on the hybrid feature data and the network status data, a multi-scale feature pyramid is constructed, and the cross-scale attention weight is calculated; based on the cross-scale attention weight, a feature enhancement operation is performed to generate enhanced feature data; based on the enhanced feature data, feature reliability verification is performed to obtain reliability score data; S33, constructing a time series feature sequence based on the enhanced feature data and the reliability score data, and generating a nonlinear dynamic model; performing multi-path prediction calculation based on the nonlinear dynamic model to obtain deformation prediction data; Based on the deformation prediction data, uncertainty assessment is performed to generate uncertainty data; S34. Construct a risk assessment matrix based on deformation prediction data and uncertainty data; Based on the risk assessment matrix and pre-stored expert knowledge base data, conduct decision analysis, calculate warning thresholds, and generate graded warning strategies; Based on the graded warning strategy, warning level data and decision-making recommendation data are generated.

2. The building deformation analysis method based on three-dimensional laser scanning and deep learning according to claim 1 is characterized in that: Step S1 is further as follows: S11, receiving original point cloud data including three-dimensional coordinates and intensity values, and constructing a feature tensor; recursively decomposing the feature tensor into multiple levels, and setting corresponding feature weight coefficients for each level, to obtain a multi-level feature tensor with weight coefficients; based on the multi-level feature tensor, respectively extracting structural features, local features and noise features, and generating hierarchical feature data; S12, obtaining environmental parameter data, constructing an environmental impact response matrix based on the hierarchical feature data and the environmental parameter data, calculating the degree of influence of each environmental factor on different features; based on the degree of influence, using a dynamic threshold method to distinguish environmental interference and actual deformation; based on the environmental interference and actual deformation, performing environmental impact compensation calculation, and outputting environmental compensation feature data; S13. Based on the environmental compensation feature data, a spatial topological structure tensor is constructed to generate local geometric constraints; based on the local geometric constraints, a loss function is constructed; based on the loss function, an iterative optimization method is used to complete the data and improve the data quality to generate optimized point cloud data; S14. Obtain historical monitoring data, construct a time series feature matrix based on the historical monitoring data and optimized point cloud data, and calculate the weight coefficient of current data and historical data; based on the weight coefficient, use the feature fusion method to enhance spatiotemporal consistency and output high-fidelity point cloud data and spatiotemporal feature description data.

3. The building deformation analysis method based on three-dimensional laser scanning and deep learning according to claim 2 is characterized in that: Step S2 is further as follows: S21. Based on the high-fidelity point cloud data and the spatiotemporal feature description data, a spatial feature pyramid is constructed; regional growth calculation is performed on the data in the spatial feature pyramid to obtain component region identification; based on the component region identification, a topological relationship diagram between components is established, and the organizational relationship between components is constructed through hierarchical indexing, and component feature data and component relationship data are output; S22, constructing a deformation feature tensor based on the component feature data and the component relationship data; Based on the deformation feature tensor, the deformation features of the component are extracted through geometric invariant calculation; Based on the deformation features, local deformation information and global deformation information are extracted and fused, feature significance scores are calculated, and deformation feature data and feature weight data are generated; S23, constructing a multi-scale observation window based on the deformation feature data and feature weight data; Based on the multi-scale observation window, the deformation features at different scales are aggregated and calculated to construct the deformation mode correlation matrix; Based on the deformation mode correlation matrix, the key deformation modes are highlighted through feature enhancement operation, and the fused deformation feature data is output; S24, constructing a spatiotemporal feature tensor based on the fused deformation feature data and the spatiotemporal feature description data; Based on the spatiotemporal feature tensor, a consistency function based on physical constraints is constructed, and feature optimization calculations are performed to generate optimized deformation feature data and feature credibility data.

4. The building deformation analysis method based on three-dimensional laser scanning and deep learning according to claim 3 is characterized in that: Step S4 is further as follows: S41. Construct a building mechanics constraint matrix based on warning level data, decision suggestion data and optimized deformation feature data; Based on the building mechanics constraint matrix, perform stress-strain relationship verification calculations to obtain physical verification data; Based on the physical verification data, perform energy conservation test, evaluate the physical feasibility score, and output the first verification correction data; S42, based on the physical verification data and the spatiotemporal feature description data, construct a multi-source data feature set and calculate the data consistency score; based on the data consistency score, perform anomaly detection operations and generate cross-validation data; based on the cross-validation data, evaluate the trust weight of each data source and generate second verification correction data; S43, constructing a spatiotemporal continuity constraint tensor based on the cross-validation data and the deformation prediction data; Based on the space-time continuity constraint tensor, variational optimization calculation is performed to obtain optimized data; discontinuity points in the optimized data are identified, and adaptive smoothing is performed to obtain continuous optimized data and third verification correction data; S44, constructing a comprehensive decision tree based on the continuous optimization data, the first verification correction data, the second verification correction data and the third verification correction data; Based on the comprehensive decision tree, conflict resolution calculations are performed to generate optimization results; Perform reliability evaluation on the optimization results to obtain evaluation results; based on the evaluation results, perform iterative optimization processing to generate final optimization data and reliability report data.

5. The building deformation analysis method based on three-dimensional laser scanning and deep learning according to claim 4 is characterized in that: Step S11 is further as follows: S111, receiving original point cloud data, constructing a local neighborhood covariance matrix, calculating the main eigenvector and eigenvalue, generating local feature data; performing nonlinear transformation on the local feature data, constructing a high-dimensional feature map, and outputting feature tensor data; S112, based on the feature tensor data, calculating the hierarchical weights through an adaptive feature importance evaluation function to generate initial weight data; Perform multi-scale consistency optimization on the initial weight data and output feature weight data; S113, based on the feature tensor data and the feature weight data, construct a recursive neural network structure, perform feature decomposition operation, and generate hierarchical decomposition data; perform feature correlation analysis on the hierarchical decomposition data, and output multi-level feature data; S114, based on the multi-level feature data, using a feature clustering algorithm to classify the features of each level to obtain feature category data; Perform inter-level correlation analysis and integration on feature category data to output hierarchical feature data.

6. The building deformation analysis method based on 3D laser scanning and deep learning according to claim 4 is characterized in that: Step S12 is further as follows: S121, obtaining environmental parameter data, and based on the environmental parameter data, using an interpolation algorithm to construct a spatiotemporal continuous field to generate environmental parameter field data; based on each characteristic component in the hierarchical characteristic data and the environmental parameter field data, calculating a correlation coefficient matrix to generate correlation coefficient matrix data; Based on the correlation coefficient matrix data, singular value decomposition is performed to determine the main impact mode and output the environmental impact response matrix; S122. Based on the environmental impact response matrix and the hierarchical characteristic data, matrix operations are performed to calculate the impact intensity coefficient of each environmental factor and generate environmental impact intensity data; a probability distribution model based on kernel density estimation is constructed to perform statistical analysis on the environmental impact intensity data and output environmental impact distribution data; based on the environmental impact distribution data, an adaptive threshold function is constructed to generate dynamic threshold data; S123, comparing and analyzing the change amount of the layered feature data with the dynamic threshold data, identifying deformation features beyond the environmental impact range, and generating actual deformation data; Based on the environmental impact response matrix and environmental impact intensity data, a compensation model is constructed, the environmental impact compensation amount is calculated, and the environmental compensation coefficient data is output; Based on the actual deformation data and the environmental compensation coefficient data, a feature reconstruction operation is performed to generate compensated reconstruction data; S124, based on the compensation reconstruction data, using an optimization algorithm based on physical constraints to generate compensation optimization data; Perform consistency check on compensation optimization data and pre-stored historical compensation results, and output compensation consistency data; Based on the compensation consistency data, the compensation optimization data is fine-tuned and finally the environmental compensation characteristic data is output.

7. The building deformation analysis method based on 3D laser scanning and deep learning according to claim 4 is characterized in that: Step S21 is further as follows: S211. Based on the high-fidelity point cloud data, a feature pyramid network is constructed through multi-scale analysis to generate pyramid feature data of different scales; feature fusion is performed on the pyramid feature data and the spatiotemporal feature description data to output multi-scale fusion data; based on the multi-scale fusion data, a local feature descriptor is calculated to generate feature description data; S212, based on the feature description data, using a region growing algorithm, calculating the point cloud segmentation seed points, and generating seed point data; Based on the seed point data, perform adaptive region growing operation and output region growing data; perform boundary optimization processing on the region growing data to generate segmented region data; S213, based on the segmented area data, using a graph theory method, establishing a connection relationship diagram between components and generating connection relationship data; Perform topological relationship analysis on the connection relationship data, determine the spatial position relationship between components, and output topological relationship data; Based on the topological relationship data, the interaction strength between components is calculated to generate interaction strength data; S214, constructing a hierarchical index structure based on the segmented region data and the topological relationship data, and generating index structure data; Combine the index structure data with the action intensity data, establish the component organization relationship, and output the organization relationship data; The consistency of organizational relationship data is optimized, and component feature data and component relationship data are finally output.

8. The building deformation analysis method based on 3D laser scanning and deep learning according to claim 4 is characterized in that: Step S31 is further as follows: S311, constructing an orthogonal transformation matrix based on the optimized deformation feature data and the feature credibility data, performing feature projection operation, and obtaining projection feature data; performing nonlinear spatial mapping on the projection feature data, and outputting conversion feature data through dimensional adaptive transformation; S312, input the converted feature data into the spatial feature extraction branch, the temporal feature extraction branch and the physical constraint branch respectively, generate spatial feature data, temporal feature data and physical constraint data; preliminarily align the spatial feature data, the temporal feature data and the physical constraint data, and output multi-dimensional feature data; S313, based on the multi-dimensional feature data, calculating the weight of each branch through a feature importance evaluation function to obtain initial weight data; Perform dynamic adjustment operation on the initial weight data and output weight adjustment data; S314, based on the multi-dimensional feature data and weight adjustment data, construct a feature interaction attention network, perform cross-branch feature fusion, and generate fusion state data; The fused state data is used to update the network state and integrate the features, and finally the hybrid feature data and network state data are output.

9. Building deformation analysis system based on 3D laser scanning and deep learning, characterized by: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the building deformation analysis method based on three-dimensional laser scanning and deep learning as described in any one of claims 1 to 8.

Images