Early identification method and system for cracks on the crest of high core rockfill dam

Abstract

The present invention discloses an early identification method and system for cracks on the top of a high core rockfill dam. The method comprises the following steps: labeling InSAR data, drone oblique photography data and three-dimensional laser scanning data through a spatiotemporal index structure, using a dynamic multidimensional threshold system to clean abnormal values, and obtaining a registration data set through a segmented adaptive standardization processing unit and a spatiotemporal registration unit; extracting local features by using probability tensor decomposition, building a global feature matrix in combination with graph entropy calculation, and obtaining a fusion feature set through multidimensional probability combination; building a multi-level early warning index based on hierarchical sparse representation, identifying critical states by using a probability migration matrix, and determining the warning level by using an adaptive threshold network and probability combination discrimination; using a combined verification matrix for multidimensional verification, and realizing system optimization and updating through a multi-objective optimization network and an incremental learning matrix. The method improves the accuracy and reliability of early identification of cracks.

Description

Technical Field

[0001] The invention relates to image processing technologies such as laser and video, and in particular to an early identification method and system for cracks on the crest of a high core rockfill dam. Background Art

[0002] High core rockfill dams are widely used in dam construction due to their advantages such as high material utilization, simple construction technology, and good seismic performance. However, due to the heterogeneity of rockfill materials and the uncertainty in the filling process, cracks on the dam crest have become an important hidden danger affecting the safe operation of the dam. Such cracks often start in the form of microscopic damage and gradually develop under the influence of external loads and environmental factors. Once through cracks are formed, they will seriously threaten the structural safety of the dam. Therefore, early identification of cracks on the crest of high core rockfill dams has important engineering significance and social value for timely discovery of potential risks and prevention of major accidents.

[0003] At present, the monitoring methods of high core rockfill dams mainly include traditional measurement methods and modern monitoring technologies. Traditional methods mainly rely on manual inspections and single-point measurements, such as displacement measurement and strain monitoring. Although these methods have high accuracy, they have limited spatial coverage and it is difficult to detect local anomalies in a timely manner. With the development of monitoring technology, new methods such as InSAR, drone oblique photography and three-dimensional laser scanning have gradually been applied to dam monitoring. These technologies can obtain deformation data over a wide range. In terms of data processing, traditional methods such as statistical analysis, threshold discrimination and trend analysis are mainly used. Some studies have begun to try to introduce machine learning algorithms for anomaly detection, but they are mostly limited to the processing of a single data source and simple feature extraction.

[0004] However, there are still the following technical difficulties in the early identification of cracks on the crest of high core rockfill dams: first, there is a lack of unified standards for the quality control of multi-source heterogeneous data, especially in the process of data cleaning and standardization, it is difficult to effectively handle outliers of data of different scales and different modes; second, the existing feature extraction methods often split different types of deformation features and fail to establish a dynamic correlation mechanism between features, resulting in insufficient overall characterization of the crack development process; third, the early warning indicator system lacks hierarchy, making it difficult to describe the evolution process from microscopic damage to macroscopic instability, and the setting of the early warning threshold is too dependent on experience; fourth, the reliability verification mechanism of the early warning results is imperfect, especially in terms of physical constraints and spatiotemporal consistency, which affects the credibility of the early warning results; finally, the optimization process of system parameters lacks adaptive capabilities, making it difficult to dynamically adjust and continuously improve according to new data. These technical problems directly affect the accuracy and reliability of early identification of dam crest cracks, and more advanced identification methods are urgently needed. Summary of the invention

[0005] The purpose of the invention is to provide a method and system for early identification of cracks on the crest of a high core rockfill dam, in order to solve one of the above-mentioned problems existing in the prior art.

[0006] The technical solution is a method for early identification of cracks on the crest of a high core rockfill dam, comprising the following steps:

[0007] S1. Obtain geospatial data and input it into a unified spatiotemporal index structure for labeling to obtain labeled data; based on the labeled data, perform outlier cleaning, standardization processing and adaptive weight fusion to obtain a registered data set; the geospatial data includes InSAR data, drone oblique photography data and 3D laser scanning data;

[0008] S2. Based on the registered data set, obtain the local feature set and its importance through probability tensor decomposition; generate the global feature matrix based on the local feature set and its importance; extract the time series feature vector based on the global feature matrix and the spatiotemporal index structure; perform feature fusion based on the local feature set, the global feature matrix and the time series feature vector to obtain the fused feature set;

[0009] S3. Based on the fusion feature set, a multi-level warning indicator set is constructed; based on the warning indicator set and pre-stored historical data, a stability indicator sequence is generated; based on the stability indicator sequence, a dynamic threshold network is constructed through adaptive updating; based on the stability indicator sequence and the dynamic threshold network, the warning level and credibility value are obtained through probability combination discrimination;

[0010] S4. Based on the warning level and credibility value, generate the verification matrix and consistency coefficient, calculate the robustness index and sensitivity matrix; based on the robustness index and sensitivity matrix, perform multi-objective optimization to obtain the optimized system parameters; obtain new data, and obtain new system parameters and a new knowledge base through dynamic updates based on the new data, the optimized system parameters and the pre-stored historical database.

[0011] The early identification system of cracks on the crest of high core rockfill dam includes:

[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 method for early identification of cracks on the crest of a high core rockfill dam.

[0015] Beneficial effects: The present invention ensures the quality and consistency of multi-source monitoring data, realizes multi-scale characterization and effective fusion of deformation characteristics, and improves the accuracy and reliability of early warning results; it not only solves the hysteresis and uncertainty problems existing in traditional methods in early crack identification, but also establishes a set of scalable and extensible intelligent early warning analysis framework, which provides strong technical support for the safety monitoring of high core wall rockfill dams. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the present invention.

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

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

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

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

[0021] like Figure 1 As shown, the present application proposes an early identification method for cracks on the crest of a high core rockfill dam, comprising the following steps:

[0022] S1. Obtain geospatial data and input it into a unified spatiotemporal index structure for labeling to obtain labeled data; based on the labeled data, perform outlier cleaning, standardization processing and adaptive weight fusion to obtain a registered data set; the geospatial data includes InSAR data, drone oblique photography data and 3D laser scanning data;

[0023] S2. Based on the registered data set, obtain the local feature set and its importance through probability tensor decomposition; generate the global feature matrix based on the local feature set and its importance; extract the time series feature vector based on the global feature matrix and the spatiotemporal index structure; perform feature fusion based on the local feature set, the global feature matrix and the time series feature vector to obtain the fused feature set;

[0024] S3. Based on the fusion feature set, a multi-level warning indicator set is constructed; based on the warning indicator set and pre-stored historical data, a stability indicator sequence is generated; based on the stability indicator sequence, a dynamic threshold network is constructed through adaptive updating; based on the stability indicator sequence and the dynamic threshold network, the warning level and credibility value are obtained through probability combination discrimination;

[0025] S4. Based on the warning level and credibility value, generate the verification matrix and consistency coefficient, calculate the robustness index and sensitivity matrix; based on the robustness index and sensitivity matrix, perform multi-objective optimization to obtain the optimized system parameters; obtain new data, and obtain new system parameters and a new knowledge base through dynamic updates based on the new data, the optimized system parameters and the pre-stored historical database.

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

[0027] S11, obtaining geographic spatial data from the dam monitoring system and inputting it into a spatiotemporal index processing unit; in the spatiotemporal index processing unit, establishing a time index table based on pre-stored timestamp information, establishing a spatial index table based on pre-stored spatial coordinate information, and combining the time index table and the spatial index table to generate a unified spatiotemporal index structure; based on the spatiotemporal index structure, marking the spatiotemporal position information of each data point in the geographic spatial data to obtain marked data;

[0028] S12, inputting the marked data into the dynamic window matrix, generating a local feature vector by calculating the local mean, standard deviation, skewness and kurtosis of the data within a predetermined range around each data point; inputting the generated local feature vector into the abnormality determination function, and calculating the abnormality degree using an adaptive weight coefficient; based on the abnormality degree, determining whether the data point is an abnormal value, and obtaining a determination result; inputting the determination result into the data cleaning unit, removing the abnormal value in the data point, and obtaining a cleaned data set;

[0029] S13, inputting the cleaned data set into the histogram analysis unit to calculate the distribution characteristics of the data; based on the distribution characteristics, determining the data segmentation boundary through the local extreme value points; inputting the data segmentation boundary and the cleaned data set into the segmentation standardization processing unit to calculate the statistical characteristics of each data segment, including the mean, standard deviation and skewness; based on the statistical characteristics, standardizing the data to generate a standardized data set;

[0030] S14. Input the standardized data set and the spatiotemporal index structure into the spatiotemporal weight calculation unit, calculate the time weight coefficient by the time interval, and calculate the spatial weight coefficient by the spatial distance; perform Hadamard product operation on the time weight coefficient and the spatial weight coefficient to obtain the fusion weight coefficient; input the fusion weight coefficient and the standardized data set into the data resampling unit, resample the data based on the fusion weight coefficient, and generate the aligned data set.

[0031] In one embodiment of the present application, original InSAR data D_insar, UAV oblique photography data D_uav and three-dimensional laser scanning data D_laser are obtained, the three types of data are indexed and marked according to a unified timestamp and spatial coordinates, and a unified spatiotemporal index structure T_index is output to establish. Based on the original InSAR data D_insar, UAV oblique photography data D_uav and 3D laser scanning data D_laser, a dynamic window matrix W(i, j) is constructed: W(i, j)={x(ik:i+k, jk:j+k)}, k is the adaptive window size; the local eigenvector is calculated: F(i, j)=[μ(i, j), σ(i, j), S(i, j), K(i, j)], where: μ is the mean, σ is the standard deviation, S is the skewness, and K is the kurtosis; a multidimensional anomaly judgment function is constructed: θ(x) =α·|x-μ| / σ+ β·|S| +γ·|K-3|, where: α, β, γ are adaptive weight coefficients; the cleaned data sets D_insar_c, D_uav_c, and D_laser_c are output. Based on the cleaned data set, calculate the data distribution characteristics: P(x)=hist(x, n_bins), where n_bins is the number of adaptive bins; identify data segments: Seg={[a1, b1], [a2, b2], ..., [ai, bi]}, determine the data segments based on the local extreme points of the data distribution characteristics P(x); perform segment normalization: x_norm=(x-μ k ) / (σ k ·ln(1+|S k |)), where k represents the data segment to which x belongs, S k The skewness of this segment is output, and the standardized data sets D_insar_n, D_uav_n, and D_laser_n are output. Based on the standardized data sets, the spatiotemporal weight matrix is ​​constructed: the time weight W_t=exp(-|t i -t j | / τ); spatial weight W_s = exp(-||x i -x j || / d), where τ and d are feature scale parameters; calculate the fusion weight: W_f=W_tΘW_s, where Θ is the Hadamard product; perform data resampling: D_aligned =∑(W_f·D) / ∑W_f; output the aligned data set D_aligned.

[0032] This embodiment achieves high-quality input of deformation monitoring data of high core rockfill dams through multi-source data acquisition and preprocessing. Specifically, a unified spatiotemporal index structure is used to mark InSAR data, drone oblique photography data and three-dimensional laser scanning data, so that monitoring data from different sources, different scales and different spatiotemporal resolutions can be organized and managed under a unified four-dimensional spatiotemporal framework; outlier cleaning is performed through a dynamic multi-dimensional threshold system, overcoming the limitation that the traditional single threshold method cannot adapt to data diversity, especially for local subtle deformation features such as dam top cracks, which can effectively retain the real deformation signal while eliminating interference noise; segmented adaptive standardization processing fully considers the nonlinear characteristics and multi-scale characteristics of rockfill dam deformation data, and ensures the comparability and consistency of data through adaptive segmentation and local statistical feature correction; spatiotemporal registration is achieved through adaptive weight fusion, solving the problem of non-correspondence of multi-source data in time and space dimensions, and laying a high-quality data foundation for subsequent feature extraction and early warning analysis; this embodiment improves the reliability and accuracy of subsequent analysis.

[0033] According to one aspect of the present application, step S12 is further:

[0034] S121, inputting the marked data into the data partition calculation unit respectively, and obtaining the initial partition size parameter by calculating the local data density distribution and the spatial autocorrelation coefficient; based on the initial partition size parameter, using the recursive quadtree partitioning algorithm, generating a multi-scale window partitioning result; based on the multi-scale window partitioning result, extracting the multi-scale features of each window to obtain an enhanced feature set; based on the enhanced feature set, calculating the mean, standard deviation, skewness and kurtosis, and generating a local feature vector set;

[0035] S122. Based on the local feature vector set, perform dimensionality reduction through kernel principal component analysis to obtain a reduced-dimensional feature matrix; based on the reduced-dimensional feature matrix, use the kernel density estimation method to calculate the multidimensional joint distribution of the features and generate a probability density function; based on the probability density function, determine the abnormality judgment threshold of each dimension through the maximum entropy principle and output a multidimensional threshold vector; based on the multidimensional threshold vector, perform correction through extreme value statistics theory to obtain an optimized threshold set;

[0036] S123, based on the local feature vector set and the optimized threshold set, using the improved DBSCAN algorithm to perform hierarchical clustering to obtain multi-level data clusters; based on the multi-level data clusters, using the local anomaly factor method to calculate the abnormality of each data point and generate an abnormality matrix; based on the abnormality matrix, determining the abnormality level through a fuzzy inference system, and outputting an abnormal labeling result; based on the abnormal labeling result, using a pre-configured Markov random field model for optimization to obtain a corrected abnormal label;

[0037] S124. Based on the corrected anomaly markers, extract the abnormal data points and their neighborhood information to generate a data set to be repaired; based on the data set to be repaired, extract the local manifold structure through tensor singular value decomposition to obtain a structural feature tensor; based on the structural feature tensor, use the local weighted regression method to reconstruct the abnormal data to generate a repair data set; based on the repair data set, verify the reliability of the repair result by calculating the reconstruction error and local consistency index, and output the cleaned data set.

[0038] In one embodiment of the present application, InSAR data, UAV oblique photography data and three-dimensional laser scanning data are respectively input into a data partition calculation unit, and an initial partition size parameter is obtained by calculating the local data density distribution and the spatial autocorrelation coefficient; the initial partition size parameter and the original data are input into an adaptive partition unit, and a multi-scale window partition result is generated using a recursive quadtree partitioning algorithm; the multi-scale window partition result is input into a feature enhancement unit, and a dual-tree complex wavelet transform is used to extract multi-scale features of each window to obtain an enhanced feature set; the enhanced feature set is input into a local statistical calculation unit, and statistics such as the mean, standard deviation, skewness, and kurtosis are calculated, and a local feature vector set is output.

[0039] The local feature vector set is input into the feature space transformation unit, and the dimension is reduced by kernel principal component analysis to obtain the reduced dimension feature matrix; the reduced dimension feature matrix is ​​input into the probability distribution estimation unit, and the multidimensional joint distribution of the features is calculated by the kernel density estimation method to generate the probability density function; the probability density function is input into the adaptive threshold calculation unit, and the abnormal judgment threshold of each dimension is determined based on the maximum entropy principle, and the multidimensional threshold vector is output; the multidimensional threshold vector is input into the threshold optimization unit, and the threshold is corrected by extreme value statistics theory to obtain the optimized threshold set;

[0040] The local feature vector set and the optimized threshold set are input into the hierarchical clustering unit, and the improved DBSCAN algorithm is used to perform hierarchical clustering on the data to obtain multi-level data clusters. Specifically, the feature vector is constructed as a k-nearest neighbor graph, and the local density and relative density of each data point are calculated; based on the density calculation, an adaptive neighborhood radius is introduced, and the size of the neighborhood range is determined by the local density gradient; the neighborhood range parameters and local density are input into the core point identification unit, and the core point is determined based on the density ratio and density gradient; for the identified core points, a density reachable graph is constructed, and a hierarchical traversal is performed on the density reachable graph to achieve hierarchical clustering and obtain multi-level data clusters. The multi-level data clusters are input into the anomaly calculation unit, and the anomaly of each data point is calculated based on the local anomaly factor method to generate an anomaly matrix; the anomaly matrix is ​​input into the hierarchical discrimination unit, and the anomaly level is determined through the fuzzy inference system, and the anomaly labeling result is output; the anomaly labeling result is input into the spatial consistency verification unit, and optimized based on the Markov random field model to obtain the corrected anomaly labeling;

[0041] The corrected abnormal mark and the original data are input into the abnormal data location unit, the data points marked as abnormal and their neighborhood information are extracted, and the data set to be repaired is generated; the data set to be repaired is input into the local structure analysis unit, and the local manifold structure of the data is extracted by tensor singular value decomposition to obtain the structural feature tensor; the structural feature tensor is input into the data reconstruction unit, and the abnormal data is reconstructed by the local weighted regression method to generate the repaired data set; the repaired data set is input into the quality assessment unit, and the reliability of the repair result is verified by calculating the reconstruction error and the local consistency index, and the final cleaned data set is output.

[0042] The specific implementation of the improved DBSCAN algorithm is as follows:

[0043] (1) Improve density calculation: introduce adaptive neighborhood radius: r(p) = r_base * (1 +α*σ_local(p)), where r_base is the base radius, σ_local(p) is the local standard deviation of point p, and α is the adjustment coefficient; calculate local density: density(p) = Σ(gaussian_kernel(dist(p, q)) / r(p)), q∈N(p), where gaussian_kernel is the Gaussian kernel function and N(p) is the neighborhood of point p;

[0044] (2) Improvement of core point identification: Calculate relative density: rel_density(p) = density(p) / avg_density(N(p)), introduce density gradient: grad_density(p) = ▽density(p); core point determination condition is: density(p)>min_pts && rel_density(p) > density_ratio && ||grad_density(p)|| <grad_threshold;

[0045] (3) Hierarchical clustering process: Construct a density-reachable graph: G = (V, E), where V is the data point set and E is the density-reachable edge; define a hierarchical threshold sequence: [h1, h2, ..., hk]; for each threshold hi: extract the subgraph Gi that satisfies density(p)>hi; perform connected component analysis on the subgraph Gi; and record the clustering results at each level. Where r(p) is the adaptive neighborhood radius of point p; r_base is the base neighborhood radius, which specifies the initial search radius; α is the neighborhood radius adjustment coefficient, which is used to control the degree of adaptation; σ_local(p) is the standard deviation of the local area of ​​point p; density(p) is the local density value of point p; dist(p, q) is the Euclidean distance between point p and point q; N(p) is the set of neighborhood points of point p; rel_density(p) is the relative density value of point p; grad_density(p) is the density gradient value at point p; density_ratio is the relative density judgment threshold; grad_threshold is the density gradient judgment threshold.

[0046] This embodiment uses an improved DBSCAN algorithm to implement a multi-level data anomaly detection scheme, which is specifically designed for the early identification of cracks on the crest of a high core rockfill dam and has achieved significant technical results. By combining an adaptive partitioning mechanism with a dual-tree complex wavelet transform, the deformation characteristics at different scales are effectively captured, especially with a high sensitivity to local deformation anomalies caused by tiny cracks; a multi-dimensional threshold system optimized by kernel density estimation and maximum entropy principle is adopted to overcome the limitations of the traditional single threshold method in processing multi-source heterogeneous data, and the discrimination criteria can be adaptively adjusted according to the local statistical characteristics of the data; hierarchical clustering analysis and local anomaly factor methods are combined to achieve hierarchical identification of abnormal patterns, effectively solving the multi-scale and multi-modal abnormal characteristics that appear in the development of dam crest cracks; local structure analysis based on tensor singular value decomposition and local weighted regression methods are adopted to ensure that noise interference can be removed during data cleaning, while retaining the true deformation information to the greatest extent. This embodiment improves the accuracy and reliability of subsequent dam crest crack identification.

[0047] According to one aspect of the present application, step S13 is further:

[0048] S131. Based on the cleaned data set, the minimum description length criterion is used to determine the optimal number of histogram intervals; based on the optimal number of histogram intervals, a Gaussian kernel function with a variable bandwidth is used to calculate the probability density distribution; based on the probability density distribution, local extreme value features are extracted by continuous wavelet transform to obtain a data distribution characteristic curve;

[0049] S132, based on the data distribution characteristic curve, using a peak detection algorithm based on duration to identify local maximum and minimum points, and generate an extreme point sequence; based on the extreme point sequence, using a density peak clustering algorithm to group the extreme points to obtain a grouping result; based on the grouping result, using a maximum inter-class variance criterion to determine the final data segmentation boundary;

[0050] S133, based on the data segmentation boundary and the cleaned data set, using a robust statistical method to calculate the statistical characteristics of each data segment; based on the statistical characteristics, using the Huber estimator for processing to obtain the corrected segment statistical characteristics;

[0051] S134. Based on the corrected segmented statistical features and the cleaned data set, a standardized transformation function based on skewness correction is constructed, and standardized transformation is performed on each data segment to obtain preliminary standardized results; based on the preliminary standardized results, a local polynomial regression method is used to smooth the transition area between segments to obtain a standardized data set.

[0052] This embodiment aims at the non-stationary and multi-scale characteristics of the monitoring data of the cracks on the dam crest of the high core rockfill dam, and realizes the high-quality standardization of the data. Specifically, the Gaussian kernel function with adaptive bandwidth is used for local density estimation, combined with the continuous wavelet transform method, so that the data segmentation process can adaptively match the spatial non-uniform characteristics of the rockfill dam deformation, avoiding the problem of smoothing the local mutation characteristics of the traditional fixed segmentation method; the segmentation boundary is determined by density peak clustering and the maximum inter-class variance criterion, while ensuring the smooth transition of the segmentation, the mutation characteristics of the deformation are retained, which is of great significance for the early identification of the cracks on the dam crest; by constructing a standardized transformation function based on skewness correction, the skewed distribution problem of the deformation data is effectively handled, so that the standardized result better maintains the statistical characteristics of the original data; the boundary smoothing processing is performed by the local polynomial regression method, which solves the boundary effect problem caused by the segmentation processing and ensures the continuity and reliability of the standardized data. This embodiment realizes the accurate characterization and effective conversion of data distribution characteristics, laying a reliable foundation for the subsequent multi-source data fusion.

[0053] According to one aspect of the present application, step S13 may also be:

[0054] S13a, input the cleaned InSAR data, UAV oblique photography data and 3D laser scanning data into the data distribution feature analysis unit respectively, first construct a 3D probability distribution matrix, and obtain the initial distribution feature by calculating the local probability density of each dimension; input the initial distribution feature into the multi-peak detection unit, and use the improved mean shift algorithm for peak detection; the improved mean shift algorithm dynamically adjusts the kernel function bandwidth according to the local density of the data by introducing an adaptive bandwidth selection mechanism, and uses the local covariance matrix to adjust the shape of the search window to achieve accurate peak location; input the peak detection result into the distribution pattern recognition unit, determine the distribution type of the data by high-order moment analysis, and output the distribution feature descriptor;

[0055] S13b, input the distribution feature descriptor and the original data into the boundary point identification unit, first estimate the density distribution of the data by constructing a local density function; input the density distribution result into the improved watershed segmentation algorithm for boundary identification, the algorithm introduces an adaptive immersion step and a region merging criterion, wherein the adaptive immersion step is dynamically adjusted based on the size of the local density gradient, and the region merging criterion comprehensively considers the density difference and spatial continuity between regions; input the initially identified boundary points into the boundary optimization unit, use the level set evolution method to finely adjust the boundary position, and output the optimized segmented boundary point set;

[0056] S13c, input the segment boundary point set and the original data into the segment statistical feature calculation unit, calculate the mean, variance, and skewness coefficient of each data segment; input the statistical features into the improved z-score standardization unit, the improved algorithm introduces quantile-based robust estimation, replaces the mean and standard deviation with the median and interquartile range, and introduces a skewness correction factor to dynamically adjust the standardization coefficient according to the skewness of the data; input the preliminary standardization result into the nonlinear mapping unit, use the segmented hyperbolic tangent function for nonlinear transformation, realize adaptive standardization of the data, and output a standardized data set;

[0057] S13d. Input the standardized data set into the distribution consistency test unit, and use the improved KS test method to evaluate the standardization effect; this improved method gives different importance to different data areas by introducing local density weights, while considering the spatiotemporal correlation of the data; input the test results into the anomaly detection unit, and identify possible standardized anomalies based on local density estimation; input the anomaly information into the adaptive correction unit, correct the anomalies through local linear regression, construct a standardized quality evaluation index, and output the final standardized verification report and the corrected standardized data set.

[0058] According to one aspect of the present application, step S14 is further:

[0059] S141. Based on the standardized data set and spatiotemporal index structure, a dynamic time warping algorithm is used to calculate the time correspondence between data sequences; based on the time correspondence, an optimal time window size is determined by an information entropy criterion to obtain a time weight coefficient;

[0060] S142. Based on the standardized data set and spatiotemporal index structure, an iterative closest point algorithm is used to calculate the spatial correspondence between different data sources; based on the spatial correspondence, a heterogeneous kernel function is used to construct an adaptive bandwidth matrix to obtain a spatial weight coefficient;

[0061] S143, constructing a sparse constraint term based on the L1 norm based on the time weight coefficient and the space weight coefficient; obtaining a fusion weight coefficient through iterative optimization based on the sparse constraint term;

[0062] S144. Based on the fusion weight coefficient and the standardized data set, the optimal affine transformation parameters of each data point are calculated; based on the optimal affine transformation parameters, a variational reconstruction method is used to resample the data to obtain the aligned data set.

[0063] This embodiment solves the problem of inconsistency in the time and space dimensions of multi-source monitoring data of high core rockfill dams. The dynamic time warping algorithm is used to achieve time alignment of monitoring data with different sampling frequencies. The time window size is optimized by the information entropy criterion, which effectively avoids the problem of over-smoothing or information loss caused by the traditional fixed window method; the spatial registration scheme combining the iterative nearest point algorithm and the anisotropic kernel function is introduced, and the adaptive bandwidth matrix is ​​constructed by considering the local principal curvature direction, which accurately characterizes the spatial heterogeneity characteristics of the dam body deformation and improves the registration accuracy; the alternating direction multiplier method based on the L1 norm constraint is used to optimize the fusion weight, so that the weight distribution maintains sparsity and fully considers the reliability differences of different data sources; through the coordinated use of local affine transformation and variational reconstruction methods, high-precision data resampling is achieved, especially in local areas of severe deformation such as cracks on the dam top, and important structural details are maintained. This embodiment forms a complete processing chain of time alignment, spatial registration and data reconstruction, which improves the accuracy and reliability of multi-source data fusion.

[0064] According to one aspect of the present application, step S14 may also be:

[0065] S14a, input the standardized data set into the time series feature extraction unit, and obtain the time series feature matrix by calculating the autocorrelation coefficient and the mutual correlation coefficient; input the time series feature matrix into the missing point analysis unit, and use the improved dynamic time warping algorithm to align the data; the algorithm introduces an adaptive window mechanism to dynamically adjust the comparison window size according to the local change rate of the data, and at the same time uses the local morphological features to construct an improved distance metric, in which the distance metric comprehensively considers the amplitude difference, trend similarity and local morphological features; input the alignment result into the time series interpolation unit, generate a continuous time series based on the multidimensional interpolation method of tensor decomposition, and output the time-aligned data set;

[0066] S14b, input the time-aligned data set into the spatial transformation parameter estimation unit to construct a multi-view spatial feature descriptor; the feature descriptor is optimized by an improved RANSAC algorithm, and the improvements include: introducing an adaptive sampling strategy to dynamically adjust the sampling probability based on the reliability of feature points; constructing a multi-scale consistency evaluation criterion to verify the validity of transformation parameters at different spatial scales; introducing a local deformation model to handle non-rigid transformation situations; inputting the optimized feature matching results into the coordinate transformation unit, using a piecewise affine transformation model to calculate the transformation matrix, and outputting a spatial registration data set in a unified coordinate system;

[0067] S14c, input the spatial registration data set into the data quality assessment unit, calculate the signal-to-noise ratio, spatial resolution and coverage of each data source; input the quality assessment index into the improved multi-source weight calculation unit, the algorithm first constructs a local reliability index based on information entropy, then modulates the weight distribution by introducing a spatial distance decay function, and finally optimizes the weight distribution using local consistency constraints; input the initial weight result into the weight optimization unit, adopt a global optimization method based on graph cuts to ensure the spatial continuity of the weight distribution, and output a fusion weight matrix;

[0068] S14d. Input the fusion weight matrix and the spatial registration data set into the data fusion unit, and use the improved local weighted regression algorithm for data fusion. The algorithm dynamically adjusts the smoothing parameters according to the local distribution characteristics of the data by introducing an adaptive kernel function, and uses the local structure tensor to guide the fusion process to maintain the geometric characteristics of the data. The fusion result is input into the quality control unit, and the fusion effect is evaluated based on the improved cross-validation method. This method constructs a multi-scale evaluation index, comprehensively considers spatial consistency, temporal continuity and physical validity, generates a fusion quality evaluation report, and outputs the final spatiotemporal registration data set.

[0069] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0070] S21, input the registered data set into the local feature tensor construction unit, calculate the deformation, curvature and gradient value of each data point, and construct the local feature tensor; input the local feature tensor into the probability tensor decomposition unit, calculate the eigenvalue and eigenvector, and generate the tensor decomposition result; input the tensor decomposition result into the feature importance calculation unit, calculate the importance of each feature by the ratio of the eigenvalue and the residual size, and output the local feature set and its corresponding importance value;

[0071] S22, input the local feature set and its corresponding importance value into the adjacency matrix construction unit, calculate the connection strength between nodes through the Euclidean distance and importance value between local features, and generate an adjacency matrix; input the adjacency matrix into the graph entropy calculation unit, calculate the graph entropy value of each node; input the graph entropy value into the entropy matrix construction unit, and generate a global feature matrix through the combination operation and gradient correction of the entropy value;

[0072] S23, inputting the global feature matrix and the spatiotemporal index structure into the time series probability matrix construction unit, calculating the conditional probability between adjacent time points, and generating a time series probability matrix; inputting the time series probability matrix into the feature decomposition unit, and obtaining the main components of the time series features through basis function decomposition; inputting the main components into the stability analysis unit, and obtaining the time series feature vector by calculating the gradient ratio;

[0073] S24. Input the local feature set, the global feature matrix and the time series feature vector into the probability combination matrix construction unit, and combine them row by row to form a probability combination matrix; input the probability combination matrix and the pre-stored historical tag data into the weight optimization unit, and obtain the combination weight by constructing and solving the least squares problem with L1 norm constraint; input the combination weight and the probability combination matrix into the feature reconstruction unit, and obtain the fused feature set through weighted combination and gradient correction.

[0074] In one embodiment of the present application, based on the aligned data set D_aligned, a local feature tensor T(i, j, k) is constructed: T={v_ij, c_ij, g_ij}, where: vi_ij is the deformation, c_ij is the curvature, and g_ij is the gradient; tensor probability decomposition is performed: P(T)=∑λ k (u k Θv k Θw k )+R, where: k is the eigenvalue, u, v, w are the eigenvectors, and R is the residual; calculate the feature importance: I(k) = λ k / ∑λ i (1-||R|| / ||T||), output the local feature set F_local and its importance I_local.

[0075] Based on the local feature set F_local and its importance I_local, construct the adjacency matrix A: A_ij= exp(-||F_i- F_j|| 2 / σ 2 )·I_i·I_j; Calculate the graph entropy: H(i)=-∑(A_ij·log(A_ij)) / log(n); Construct the entropy matrix: E =HΘH T +β·▽ 2 H, where β is the smoothing coefficient, and outputs the global feature matrix G_feature.

[0076] Based on the global feature matrix G_feature and the spatiotemporal index structure T_index, construct the time series probability matrix: M(t)={P(x_t|x_{tk},...,x_{t-1})}; perform feature decomposition: Φ(t)=∑α i φ i (t)+ε(t), where φ i (t) is the basis function, α i is the coefficient; perform stability analysis: S(t)=|▽Φ(t)| / |Φ(t)|, and output the time series feature vector T_feature.

[0077] Based on the local feature set F_local, the global feature matrix G_feature and the temporal feature vector T_feature, a probability combination matrix is ​​constructed: C = [F_local; G_feature; T_feature]; and the combination weight is calculated: W = argmin ||C·wy|| 2 +α·||w||1, where y is the historical labeled data; perform feature reconstruction: F_final=C·W+γ▽W, where γ is the smoothing factor, and output the fused feature set F_final.

[0078] This embodiment realizes the all-round characterization and in-depth mining of the deformation characteristics of the high core rockfill dam through the multi-dimensional feature extraction and combination mechanism. In the local feature extraction link, the probability tensor decomposition method is used to construct a three-dimensional feature tensor, and the physical quantities such as deformation, curvature and gradient are uniformly expressed. Through tensor decomposition and feature importance calculation, the local deformation mode caused by the dam top crack is effectively captured; in the global feature construction link, the association network between features is established based on Euclidean distance and adjacency matrix, and the overall complexity of the deformation field is characterized by graph entropy calculation, especially for the spatial propagation characteristics of the crack development process. It has good expression ability; in the time series feature extraction link, the probability time series decomposition model is introduced, and the dynamic characteristics of deformation evolution are accurately characterized by constructing conditional probability matrix and basis function decomposition; through multi-dimensional probability combination, the effective fusion of features is realized, which not only maintains the independent expression ability of various features, but also establishes the synergistic relationship between features, providing rich and reliable feature representation for subsequent early warning analysis. The feature extraction process of this embodiment forms a complete feature spectrum from local to global, from static to dynamic, and improves the ability to identify early signs of dam top cracks.

[0079] According to one aspect of the present application, step S22 is further:

[0080] S221, inputting the local feature set and the feature importance value into the feature distance calculation unit, using Wasserstein distance to measure the similarity between features, inputting the similarity matrix into the adaptive threshold unit, determining the connection threshold by the OTSU algorithm, and obtaining the initial adjacency matrix;

[0081] S222, inputting the initial adjacency matrix into the spectral clustering unit, extracting the feature subspace by normalized Laplace matrix decomposition, inputting the feature vector into the manifold learning unit, reconstructing the node relationship by using the local linear embedding algorithm, and obtaining the optimized adjacency matrix;

[0082] S223, inputting the optimized adjacency matrix into the random walk model unit, calculating the transition probability matrix between nodes, inputting the transition probability into the entropy rate calculation unit, and estimating the graph entropy value of each node through Kullback-Leibler divergence;

[0083] S224. Input the entropy value of the graph into a matrix decomposition unit, use non-negative matrix decomposition to extract the main mode of the entropy value, input the decomposition result into a gradient adjustment unit, perform smoothing through a total variation regularization method, and obtain a global feature matrix.

[0084] This embodiment realizes the multi-scale characterization of the deformation characteristics of the high core rockfill dam through the global feature construction mechanism. The Wasserstein distance is used to measure feature similarity, which can better reflect the distribution characteristics of deformation data than the traditional Euclidean distance, especially for the local deformation mode caused by the dam top cracks. It has better ability to distinguish; through the combined application of normalized Laplace matrix decomposition and local linear embedding, the local topological structure of the deformation data is effectively maintained, so that the extracted features can better reflect the spatial correlation of crack development; the introduction of the random walk model and the Kullback-Leibler divergence estimation of the graph entropy calculation method not only captures the local complexity of the deformation characteristics, but also establishes the dynamic correlation between different observation points; through the combination of non-negative matrix decomposition and total variation regularization method, the sparse representation and smoothing of features are realized, which not only reduces data redundancy, but also maintains the continuity of deformation evolution. This embodiment forms a complete framework from local feature extraction to global feature organization, providing rich and reliable feature representation for subsequent early warning analysis.

[0085] According to one aspect of the present application, step S22 may also be:

[0086] S22a, inputting the local feature set and the feature importance value into the multi-level adjacency relationship analysis unit, and establishing the initial connection relationship through the improved k-nearest neighbor graph construction algorithm; the improved algorithm optimizes the neighborhood construction in the following ways: firstly, introducing an adaptive distance metric, and dynamically adjusting the distance calculation method according to the local density distribution of the feature space; then constructing a multi-scale connection threshold, and determining the optimal connection range based on the statistics of each order of the local feature distribution; finally, introducing a feature importance weighting mechanism, and modulating the connection strength through the feature importance value; inputting the initial connection relationship into the topology optimization unit, and using the connection pruning method based on spectral analysis to eliminate redundant connections, and outputting the optimized multi-level adjacency matrix;

[0087] S22b, input the multi-level adjacency matrix into the graph feature extraction unit, and use the improved graph convolution algorithm for feature propagation; the improvements of the algorithm include: constructing an adaptive convolution kernel to dynamically adjust the convolution parameters according to the local topological structure of the node; introducing an attention mechanism to assign different importance weights to different types of feature information; designing a multi-head feature aggregation strategy to capture the interaction between nodes from multiple perspectives; inputting the propagated features into the entropy calculation unit, and quantifying the uncertainty of the features through the improved information entropy calculation method; this method realizes the adaptability of entropy calculation by introducing local probability density estimation, and outputs the graph entropy feature matrix;

[0088] S22c, input the graph entropy feature matrix into the feature propagation optimization unit, and use the improved message passing algorithm to update the features; the algorithm is improved in the following ways: construct a hierarchical message passing mechanism to transmit feature information at different abstract levels; introduce a gated update mechanism to dynamically adjust the update intensity according to the importance of the feature; design a residual connection structure to maintain the original feature information; input the updated features into the consistency constraint unit, ensure the stability of feature propagation by constructing a global consistency loss function, and output the optimized feature propagation results;

[0089] S22d. Input the feature propagation results into the global feature optimization unit, and use the improved iterative optimization algorithm to refine the features. The algorithm first constructs a multi-objective optimization function, while considering the discriminability, compactness and stability of the features. Then, by introducing an adaptive learning rate strategy, the update step size is dynamically adjusted according to the gradient changes in the optimization process. Finally, an early stopping mechanism is designed to automatically determine the optimization rounds based on the validation set performance. The optimized features are input into the feature verification unit, and the effectiveness of the features is comprehensively evaluated by constructing a multi-dimensional quality evaluation index. The verification results are input into the feature correction unit, and the abnormal features are corrected by the local linear embedding method, and the final global feature matrix is ​​output.

[0090] According to one aspect of the present application, step S24 is further:

[0091] S241, based on the local feature set, the global feature matrix and the time series feature vector, a power law transformation method is used to perform scale standardization to obtain standardized features; based on the standardized features, a mutual information matrix between features is calculated by a kernel density estimation method to obtain an initial probability combination matrix;

[0092] S242, based on the initial probability combination matrix and the pre-stored historical marking data, construct a regularization term based on the elastic network to obtain an optimization problem; based on the optimization problem, determine the optimal regularization parameter through cross-validation to obtain a combination weight coefficient;

[0093] S243, based on the combined weight coefficient, dynamically update the weight distribution using the Hedge algorithm to obtain an updated weight; based on the updated weight, evaluate the stability of the weight using the bootstrap method to obtain a revised combined weight;

[0094] S244. Based on the modified combination weights and the initial probability combination matrix, Tucker decomposition is used to extract the high-order correlation pattern of the features to obtain the decomposition result; based on the decomposition result, feature reconstruction is performed through a variational autoencoder to obtain the final fusion feature set.

[0095] This embodiment realizes the comprehensive characterization and effective integration of deformation characteristics of high core rockfill dam. The power law transformation is used for feature standardization to overcome the differences in dimensions and distribution of different types of features (such as local deformation features, global structural features and temporal evolution features), making feature fusion more reasonable; kernel density estimation is introduced to calculate the mutual information between features, and by constructing a probability combination matrix, the nonlinear dependency between different features is effectively captured, especially for the feature co-evolution mode in the process of dam top crack development, which has good expression ability; elastic network regularization and proximal gradient descent method are used to optimize the combination weights, while maintaining feature sparsity, the feature group effect is considered, avoiding the problem that traditional single L1 regularization easily ignores related features; the weight distribution is dynamically updated by the Hedge algorithm, and the weight stability is evaluated by the bootstrap method, so that the feature fusion process can adaptively respond to the dynamic changes of the deformation state; the combination of Tucker tensor decomposition and variational autoencoder is used to extract the high-order correlation mode of the features, and realize the nonlinear dimension reduction and reconstruction of the features, improving the expression ability and robustness of the fusion features.

[0096] According to one aspect of the present application, step S24 may also be:

[0097] S24a, input the local feature set, the global feature matrix and the time series feature vector into the feature correlation analysis unit, and use the improved mutual information calculation method to evaluate the correlation between the features; the method is improved in the following ways: first, an adaptive binning strategy is constructed to dynamically determine the binning boundaries according to the distribution characteristics of the features; then, a local linear approximation technique is introduced to establish an accurate probability density estimate in each bin; finally, a smoothing mechanism based on a kernel function is designed to reduce the information loss caused by discretization; the preliminarily calculated mutual information matrix is ​​input into the correlation optimization unit, and the nonlinear dependency between the features is captured through high-order mutual information analysis, and the feature correlation matrix is ​​output;

[0098] S24b, input the feature correlation matrix into the weight initialization unit, and use the improved maximum entropy weight allocation algorithm to calculate the initial weight; the algorithm first constructs an importance measure based on information gain, then ensures the sparsity of weight allocation by introducing regularization constraints, and finally designs an iterative optimization strategy to dynamically adjust the weight distribution; the initial weight is input into the adaptive adjustment unit, and the improved gradient boosting algorithm is used to optimize the weight allocation; the algorithm constructs a multi-level weak learner combination and introduces adaptive step size control, dynamically adjusts the learning rate according to the performance of each iteration, and outputs the optimized weight vector;

[0099] S24c, input the weight vector and the original features into the feature combination unit, and adopt the improved hierarchical combination strategy to perform feature fusion; this strategy is implemented in the following ways: first, construct a bottom-up feature aggregation mechanism to gradually fuse feature information at different abstract levels; then introduce an attention gating mechanism to dynamically adjust the information flow according to the importance of the features; finally, design a residual connection structure to ensure the effective transmission of low-level feature information; input the combination result into the nonlinear mapping unit, enhance the feature expression ability through the improved kernel mapping method, and output the preliminary fusion feature;

[0100] S24d. Input the preliminary fusion features into the fusion optimization unit, and use the improved multi-objective optimization algorithm to refine the features. The improvements of the algorithm include: constructing multi-dimensional optimization goals while considering the discriminability, redundancy and stability of the features; introducing an adaptive weight mechanism to dynamically adjust the importance of each goal according to the convergence of the goals during the optimization process; designing a solution selection strategy based on the Pareto frontier to ensure the optimal trade-off solution; inputting the optimized features into the quality assessment unit, and comprehensively evaluating the fusion effect by constructing a multi-angle evaluation index system; inputting the evaluation results into the feature correction unit, using local sensitive hashing technology to locate and correct abnormal features, and outputting the final fusion feature set.

[0101] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0102] S31, input the fused feature set into the feature stratification unit, calculate the micro-change index, the meso-evolution index and the macro-stability index respectively, and generate a feature stratification matrix of at least three layers; input the feature stratification matrix into the inter-layer correlation analysis unit, calculate the inter-layer correlation coefficient through matrix operation and gradient correction; construct and solve the optimization problem with double sparse constraints based on the inter-layer correlation coefficient, the feature stratification matrix and the pre-stored historical warning data, and obtain the warning indicator set; wherein the micro-change index includes deformation gradient and strain rate; the meso-evolution index includes cumulative displacement and deformation acceleration; the macro-stability index includes overall stiffness and energy dissipation;

[0103] S32, input the early warning indicator set and the pre-stored historical database into the state transition probability calculation unit, calculate the transition probability between different states, and generate a state transition probability matrix; input the state transition probability matrix at adjacent moments into the stability matrix calculation unit, and obtain the stability matrix through matrix inversion and product operations; input the stability matrix into the eigenvalue analysis unit, calculate the matrix eigenvalue, and construct a stability indicator sequence;

[0104] S33, inputting the warning indicator set and the stability indicator sequence into the threshold matrix initialization unit, constructing an initial threshold matrix for each warning indicator and warning level; inputting the initial threshold matrix into the adaptive update unit, performing threshold update by calculating the gradient and second-order gradient of the loss function, and obtaining an updated threshold matrix; inputting the updated threshold matrix into the threshold network construction unit, calculating the association weights between threshold nodes, and constructing a dynamic threshold network; wherein the loss function includes a prediction error term, a sparse constraint term, and a total variation constraint term;

[0105] S34, input the warning indicator set, stability indicator sequence and dynamic threshold network into the warning probability calculation unit, calculate the probability vector of each warning level; input the probability vector into the combination discrimination unit, obtain the combination discrimination result through multi-order probability combination and gradient correction; input the combination discrimination result into the warning level determination unit, determine the warning level through maximum probability selection, and calculate the credibility value of the warning result at the same time;

[0106] S35. Generate a warning parameter set based on the warning indicator set, the stability indicator sequence, the dynamic threshold network, the warning level and the credibility value.

[0107] In one embodiment of the present application, based on the fusion feature set F_final, a feature hierarchical matrix is ​​constructed: L={L1, L2, L3}; where L1 is the microscopic change indicator matrix (deformation gradient, strain rate); L2 is the mesoscopic evolution indicator matrix (cumulative displacement, deformation acceleration); L3 is the macroscopic stability indicator matrix (overall stiffness, energy dissipation); inter-layer correlation analysis is performed: R_ij=(Li·Lj T ) / (||Li||·||Lj||)+μ·▽ 2 (Li·Lj T ), where μ is the smoothing coefficient; the reconstruction index is sparse: I = argmin ||L·w - y|| 2 + α1||w||1+ α2||▽w||1, where y is the historical warning data, and the output warning indicator set I_warning.

[0108] Based on the warning indicator set I_warning and the historical database H_data, the state transition probability matrix is ​​constructed: P(t)={p_ij(t)}, where p_ij(t) represents the probability of transitioning from state i to state j at time t; the stability matrix is ​​calculated: S(t) =P(t)·P(t-1) -1 ; Perform eigenvalue analysis: λ(t)=eig(S(t)); Construct stability index: η(t)=min|Re(λ)| / max|Re(λ)|, and output stability index sequence η_series.

[0109] Based on the warning indicator set I_warning and the stability indicator sequence η_series, a threshold matrix is ​​constructed: ξ={θ_ij}, where i represents the indicator and j represents the warning level; the threshold is adaptively updated: ξ(t+1) =ξ(t) +α▽L(ξ)+β▽ 2 L(ξ), where L(ξ) is the loss function: L(ξ) = ||Y - f(X,ξ)|| 2 + γ1||ξ||1+ γ2TV(ξ), TV is the total variation term; construct a threshold network: N={V, E}, V is the threshold node, E is the threshold associated edge; perform weight update: w_ij=exp(-||θ_i-θ_j|| 2 / σ 2 ), output dynamic threshold network N_threshold.

[0110] Based on the warning indicator set I_warning, the stability indicator sequence η_series and the dynamic threshold network N_threshold, a warning probability vector is constructed: P_w={p1, p2, p3, p4}, where pi represents the probability of the i-th level warning; combined discrimination is performed: W=∑w_k·P_w k + λ·▽P_w, where k is the combination order; determine the warning level: Level=argmax(W), and calculate the credibility: C= max(W) / ∑W, and output the warning level Level and credibility C.

[0111] This embodiment realizes accurate assessment and early warning of the deformation state of high core rockfill dam through early warning index construction and early warning level discrimination mechanism. In the early warning index construction link, a hierarchical sparse representation method is implemented, and the early warning index system is constructed from three levels: microscopic changes (deformation gradient, strain rate), mesoscopic evolution (cumulative displacement, deformation acceleration) and macroscopic stability (overall stiffness, energy dissipation). Through inter-layer correlation analysis and sparse reconstruction, the completeness and effectiveness of the index system are ensured; in the critical state identification link, the quantitative evaluation of the deformation development trend is realized through state transition probability analysis and stability matrix calculation; in the early warning threshold construction link, a sensitive and reliable early warning judgment system is established through dynamic update and network optimization of multi-dimensional early warning thresholds; in the early warning level assessment link, the accurate division and credibility assessment of the early warning level are realized through the discrimination method of multi-order probability combination. This embodiment establishes a complete early warning chain from indicator construction to level discrimination, which is particularly suitable for the early identification and early warning of progressive damage such as dam top cracks.

[0112] According to one aspect of the present application, step S32 is further:

[0113] S321, inputting the early warning indicator set and historical data into the state space partitioning unit, using the Gaussian mixture model to adaptively partition the state space, inputting the partitioning result into the Markov chain construction unit, calculating the state transition probability by maximum likelihood estimation, and obtaining an initial state transition probability matrix;

[0114] S322, inputting the initial state transition probability matrix into the time series correlation analysis unit, using the conditional random field model to capture the long-range dependency of the state sequence, inputting the sequence characteristics into the probability propagation unit, updating the state transition probability through the belief propagation algorithm, and obtaining the optimized state transition probability matrix;

[0115] S323, inputting the optimized state transition probability matrix into the stability measurement unit, constructing a rank-based matrix decomposition model, inputting the decomposition result into the eigenvector analysis unit, evaluating the system stability by a spectral radius calculation method, and obtaining a stability matrix;

[0116] S324. Input the stability matrix into the spectrum analysis unit, use singular value decomposition to extract the main stable mode, input the mode characteristics into the indicator construction unit, determine the combination weight of the eigenvalues ​​by the entropy weight method, and obtain the stability indicator sequence.

[0117] This embodiment achieves accurate characterization and prediction of the deformation development trend of high core rockfill dam. The Gaussian mixture model is used for adaptive state space partitioning, which better adapts to the nonlinear evolution characteristics of the deformation state than the traditional fixed threshold partitioning method; the conditional random field model is introduced to capture the long-range dependency of the state sequence, and the state transition probability is optimized through the belief propagation algorithm, which effectively solves the limitation of the traditional Markov chain model ignoring historical information, especially for the prediction of the gradual development process such as the dam top crack. The rank-based matrix decomposition and spectral radius calculation method are used to evaluate the system stability, which not only considers the probabilistic characteristics of state transition, but also introduces the theory of dynamic system stability, providing more reliable state assessment results; the main stable mode is extracted through singular value decomposition, and the eigenvalue combination weight is determined by combining the entropy weight method, realizing the multi-dimensional quantitative expression of the deformation evolution trend. This embodiment establishes a complete analysis chain from state space division to stability assessment, providing a reliable theoretical basis for the judgment of warning level.

[0118] According to one aspect of the present application, step S32 may also be:

[0119] S32a, input the warning indicator set and the historical database into the multi-scale state sequence construction unit, and use the improved time series decomposition algorithm to extract state features; the algorithm is improved in the following ways: first, an adaptive window mechanism is constructed to dynamically adjust the analysis window size according to the local change characteristics of the data; then, the wavelet transform basis function is introduced to realize the time series decomposition under multi-scale; finally, a trend extraction method based on singular spectrum analysis is designed to separate the state change characteristics of different time scales; the decomposition result is input into the state coding unit, and the continuous state sequence is converted into a discrete state code through the improved symbolic dynamics method, and the multi-scale state sequence matrix is ​​output;

[0120] S32b, input the multi-scale state sequence matrix into the state transition probability calculation unit, and adopt the improved Markov migration probability estimation method; the improvements of this method include: constructing an adaptive state partitioning criterion, dynamically determining the state boundary according to the distribution characteristics of the state space; introducing a time-varying transfer matrix, and capturing the dynamic changes of the transfer probability through a sliding time window; designing a smoothing strategy based on kernel density estimation to reduce the volatility of the state transition probability estimation; inputting the initial transfer probability matrix into the probability optimization unit, and outputting the optimized state transition probability matrix by introducing a Bayesian update mechanism, integrating prior knowledge and observation data;

[0121] S32c, input the state transition probability matrix into the stability matrix construction unit, and use the improved matrix decomposition algorithm to perform stability analysis; the algorithm first constructs a time-varying eigenvalue decomposition model to evaluate the system stability by tracking the evolution path of the eigenvalue; then introduces the singular value decomposition technology to extract the main mode of the state transition matrix; finally, designs a stability criterion based on the spectral radius to quantify the overall stability of the system; inputs the analysis results into the stability evaluation unit, and comprehensively evaluates the stability of the system state by constructing a multi-dimensional stability index, and outputs a stability matrix;

[0122] S32d. Input the stability matrix into the critical feature extraction unit and adopt an improved critical slowing-down index calculation method. This method is improved in the following ways: construct a multidimensional critical index system, including recovery rate, volatility and autocorrelation, etc.; introduce an adaptive threshold mechanism to dynamically adjust the critical judgment standard according to the historical changes of the system state; design a verification mechanism based on statistical significance test to evaluate the reliability of critical features; input the critical features into the verification optimization unit, evaluate the predictive ability of the features through cross-validation technology, and adopt an integrated learning method to improve the discrimination accuracy, and output the final critical state feature set.

[0123] According to one aspect of the present application, step S33 is further:

[0124] S331, inputting the early warning indicator set and the stability indicator sequence into the quantile analysis unit, using the kernel density weighted adaptive quantile estimation method to establish an initial threshold distribution for each indicator, inputting the distribution parameters into the extreme value statistics unit, fitting the critical threshold through the generalized extreme value distribution, and obtaining the initial threshold matrix;

[0125] S332, input the initial threshold matrix into the gradient boosting unit, build a threshold optimization model based on XGBoost, input the historical warning data into the objective function construction unit, balance the false alarm rate and the missed alarm rate through a multi-objective optimization method, and obtain an optimized threshold matrix;

[0126] S333, inputting the optimized threshold matrix into the correlation analysis unit, using the partial correlation analysis method to calculate the conditional independence relationship between the thresholds, inputting the association structure into the graph network construction unit, learning the dynamic weights between nodes through the graph attention mechanism, and obtaining the threshold association network;

[0127] S334, input the threshold association network into the network optimization unit, build an importance scoring model based on PageRank, input the scoring result into the weight update unit, fine-tune the network parameters through the back propagation algorithm, and obtain a dynamic threshold network.

[0128] This embodiment realizes the dynamic optimization and precise control of deformation warning criteria of high core rockfill dams through a multi-dimensional warning threshold construction mechanism. The kernel density weighted adaptive quantile estimation method is used, combined with the generalized extreme value distribution fitting critical threshold, to overcome the problem that the traditional fixed threshold method is difficult to adapt to the non-stationarity of deformation data, especially for the early warning of the gradual development process of dam top cracks. It is more sensitive; the threshold optimization model based on XGBoost is introduced, and the false alarm rate and missed alarm rate are balanced through a multi-objective optimization method, thereby improving the accuracy and reliability of the early warning; the threshold association network is constructed by partial correlation analysis and graph attention mechanism, which not only considers the conditional independence relationship between different early warning indicators, but also realizes the dynamic adjustment of the threshold, so that the early warning system can respond adaptively to changes in deformation state; the network parameters are optimized by PageRank scoring model and back propagation algorithm, and the importance rating mechanism of the early warning threshold is established to ensure the leading role of key early warning indicators in the judgment process. This embodiment forms a complete framework from initial threshold estimation to dynamic optimization, which improves the adaptability and accuracy of the early warning system.

[0129] According to one aspect of the present application, step S33 may also be:

[0130] S33a, input the early warning indicator set and stability indicator sequence into the historical data statistical analysis unit, and use the improved multidimensional data analysis method to extract features; this method first constructs an adaptive binning strategy to determine the optimal binning scheme based on the multidimensional distribution characteristics of the data; then introduces the kernel density estimation technology to achieve probability density estimation in multidimensional space; finally, designs a feature selection mechanism based on mutual information to identify the most discriminative statistical features; inputs the analysis results into the statistical model construction unit, and establishes a probability distribution model of the early warning indicator through the improved extreme value statistical method, and outputs a statistical feature matrix;

[0131] S33b, inputting the statistical feature matrix into the threshold initialization unit, and using the improved clustering analysis method to perform initial threshold division; the method is improved in the following ways: constructing a density peak clustering algorithm to automatically determine the optimal number of clusters; introducing a fuzzy membership function to describe the degree of belonging of data points to different threshold intervals; designing an entropy-based clustering effectiveness evaluation index to optimize the clustering results; inputting the clustering results into the threshold mapping unit, establishing a corresponding relationship between the warning level and the threshold by constructing a multi-level threshold mapping relationship, and outputting the initial threshold matrix;

[0132] S33c, input the initial threshold matrix into the adaptive update unit, and use the improved online learning algorithm to optimize the threshold; the improvement of the algorithm includes: constructing an incremental update mechanism to dynamically adjust the threshold value according to the newly added data; introducing a forgetting factor to control the influence weight of historical data; designing a feedback mechanism based on the early warning effect to evaluate and adjust the effectiveness of the threshold in real time; inputting the update result into the consistency check unit, ensuring the rationality of the threshold update by constructing a multi-dimensional consistency constraint, and outputting the optimized threshold matrix;

[0133] S33d. Input the optimized threshold matrix into the threshold network construction unit, and use the improved graph network construction algorithm to establish the threshold association relationship. The algorithm first constructs a multi-level network structure to describe the transformation relationship between different warning levels. Then, a dynamic weight mechanism is introduced to adjust the network connection strength according to historical warning data. Finally, stability constraints are designed to ensure the temporal consistency of the network structure. The network structure is input into the dynamic optimization unit, and the topological characteristics of the network are evaluated by the graph analysis method. The network structure is optimized based on the random walk algorithm to output the final dynamic threshold network.

[0134] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0135] S41, input the warning level, credibility value and pre-stored historical database into the verification feature construction unit, calculate the spatial consistency feature, time continuity feature and physical rationality feature respectively, and combine them to generate a verification feature matrix; input the verification feature matrix into the verification score calculation unit to calculate the weight coefficient; based on the weight coefficient, obtain the verification score through weighted combination; input the verification score into the consistency verification unit, and obtain the consistency coefficient by calculating the difference between the observed consistency and the random consistency;

[0136] S42. Based on the warning parameter set, the verification feature matrix and the consistency coefficient, the central moment and the cross moment of each parameter are calculated to construct a probability moment tensor; the probability moment tensor is input into the robustness analysis unit, and the robustness index is obtained by calculating the ratio of the Frobenius norm and the trace of the probability moment tensor; the warning parameter set and the warning level are input into the sensitivity calculation unit, and the sensitivity index of each parameter is obtained by calculating the standardized partial derivative to form a sensitivity matrix;

[0137] S43, input the robustness index, the sensitivity matrix and the pre-stored system parameter set into the objective function construction unit, and respectively construct the warning accuracy function, the system complexity function and the calculation efficiency function; input the warning accuracy function, the system complexity function, the calculation efficiency function and the boundary constraints, the physical meaning constraints and the balance constraints of the system parameters into the multi-objective optimization unit, construct the dominance relationship matrix of the solution, and calculate the congestion distance of the solution; sort the solutions based on the dominance relationship matrix and the congestion distance to obtain the optimized system parameters;

[0138] S44. Obtain new data, input the optimized system parameters, historical database and new data into the incremental learning matrix construction unit, and use the forgetting factor to calculate the retention weight of historical knowledge; construct the incremental learning matrix based on the retention weight and the gradient information of the new data; input the incremental learning matrix and the new data into the knowledge base update unit, integrate the new knowledge into the existing knowledge base through the knowledge fusion operator to obtain a new knowledge base; input the new knowledge base into the parameter adjustment unit, and dynamically adjust the optimized system parameters according to the gradient information of the new knowledge base to obtain new system parameters.

[0139] In one embodiment of the present application, based on the warning level Level, credibility C and historical database H_data, a verification feature matrix is ​​constructed: V=[V_space; V_time; V_phys], where: V_space is the spatial consistency verification submatrix, V_time is the time continuity verification submatrix, and V_phys is the physical rationality verification submatrix; the verification score is calculated: Score(i, j) =∑w_k·v_k(i, j), where v_k is the score of the kth verification indicator, w_k=exp(entropy(v_k)) / ∑exp(entropy(v_k)); a consistency check is performed: κ= (P_o-P_e) / (1 - P_e), where P_o is the observation consistency and P_e is the random consistency; the verification matrix V_matrix and the consistency coefficient κ are output.

[0140] Based on the verification matrix V_matrix and the warning parameter set P_set, construct the probability moment tensor: M(i, j, k)=E[(X_i-μ_i)(X_j-μ_j)(X_k -μ_k)], where X is the warning parameter vector; calculate the robustness index: R=||M|| F / tr(M), where ||·|| F is the Frobenius norm; perform sensitivity analysis: S_i=ΨLevel / ΨP_i·σ(P_i) / σ(Level), where Ψ represents the partial derivative; construct the sensitivity matrix S={S_i}, and output the robustness index R and the sensitivity matrix S.

[0141] Based on the robustness index R, the sensitivity matrix S and the system parameter set Θ_sys, the objective function is constructed: minF(Θ)=[f1(Θ), f2(Θ), f3(Θ)], where f1 is the warning accuracy, f2 is the system complexity, and f3 is the computational efficiency; the constraint conditions are constructed: G(Θ)={g1(Θ)≤0, g2(Θ)≤0, h(Θ)= 0}, where g1 is the parameter boundary constraint, g2 is the physical meaning constraint, and h is the balance constraint; the optimization solution is performed: Θ*=argmin{F(Θ)|G(Θ)}, and the improved Pareto optimization algorithm is used: a) construct the dominance relationship matrix D; b) calculate the congestion distance cd_i; c) sort the solutions based on the dominance relationship matrix D and the congestion distance cd_i, and output the optimized system parameters Θ_opt.

[0142] Based on the optimized system parameters Θ_opt, the historical database H_data and the newly added data D_new, an incremental learning matrix is ​​constructed: L(t) = α·L(t-1) + (1-α)·▽L(D_new), where α is the forgetting factor; the knowledge base is updated: K(t)=K(t-1)⊕f(D_new, L(t)), where ⊕ is the knowledge fusion operator; the parameters are dynamically adjusted: Θ(t)=Θ_opt+β·▽K(t), where β is the learning rate; and the updated system parameters Θ_new and knowledge base K_new are output.

[0143] This embodiment realizes the reliability assurance and continuous improvement of the deformation early warning system of high core rockfill dam through early warning result verification and system optimization mechanism. In the early warning result verification link, multi-dimensional verification is carried out from three dimensions: spatial consistency, temporal continuity and physical rationality. The reliability of the early warning result is ensured through information entropy weighting and consistency test; in the uncertainty analysis link, based on the probability moment tensor method, the uncertainty level of the early warning result is quantitatively evaluated through robustness index calculation and sensitivity analysis; in the system adaptive optimization link, the three goals of early warning accuracy, system complexity and computational efficiency are considered at the same time, and the optimal configuration of system parameters is achieved; in the feedback correction link, the incremental learning matrix mechanism is introduced, and the continuous optimization of system performance is achieved through dynamic update of knowledge base and adaptive adjustment of parameters. This embodiment forms a closed-loop quality assurance system, which not only ensures the reliability of early warning results, but also realizes the dynamic improvement of system performance, which is particularly suitable for the needs of long-term monitoring and early warning tasks; by establishing a complete verification-optimization-feedback mechanism, the applicability and reliability of the early warning system in actual engineering are improved.

[0144] According to one aspect of the present application, step S41 is further:

[0145] S411, input the warning level, credibility value and historical database into the spatial feature extraction unit, use Ripley's K function to analyze the spatial aggregation characteristics, input the aggregation index into the local Moran index calculation unit, evaluate the spatial consistency through spatial autocorrelation analysis, and obtain the spatial verification feature;

[0146] S412, inputting the warning level sequence and historical data into the time series pattern analysis unit, using the dynamic time warping algorithm to calculate the similarity matrix of the warning sequence, inputting the similarity feature into the periodicity detection unit, analyzing the time continuity through wavelet transform, and obtaining the time verification feature;

[0147] S413, inputting the warning results and monitoring data into the physical constraint inspection unit, constructing a rationality discrimination model based on the stress-strain relationship, inputting the discrimination results into the physical rule verification unit, evaluating the physical rationality through the fuzzy reasoning system, and obtaining the physical verification feature;

[0148] S414, inputting the spatial verification feature, the temporal verification feature and the physical verification feature into the feature fusion unit, calculating the feature credibility using the Dempster-Shafer evidence theory, inputting the credibility weight into the comprehensive scoring unit, and generating the final verification score through the evidence combination rule;

[0149] S415. Input the verification score into the consistency verification unit, and obtain the consistency coefficient by calculating the difference between the observed consistency and the random consistency.

[0150] This embodiment realizes the reliability evaluation and quality assurance of the deformation early warning system of high core rockfill dam through the multi-dimensional verification mechanism of early warning results. The spatial aggregation characteristics are analyzed by Ripley's K function, and the spatial autocorrelation is evaluated by combining the local Moran index, which effectively solves the problem that the traditional verification method ignores the spatial correlation. In particular, for the deformation mode with obvious spatial propagation characteristics such as dam top cracks, the spatial consistency of the early warning results can be more accurately evaluated. The time series verification scheme combining the dynamic time warping algorithm and wavelet transform is introduced to accurately characterize the time continuity and evolution law of the early warning results by calculating the similarity matrix and periodic characteristics of the early warning sequence. Based on the stress-strain relationship, a physical rationality discrimination model is constructed, and the rule verification is carried out in combination with the fuzzy reasoning system. The correlation evaluation system between the early warning results and the physical mechanism is established to ensure that the early warning results conform to the basic mechanical laws of rockfill dam deformation. The reliable fusion of multi-dimensional verification features is realized through the Dempster-Shafer evidence theory, which not only considers the uncertainty of each verification indicator, but also generates a comprehensive score with credibility quantification through the evidence combination rule, which improves the credibility and persuasiveness of the verification results.

[0151] According to one aspect of the present application, step S41 may also be:

[0152] S41a, input the warning level, credibility value and historical database into the spatial consistency analysis unit, and use the improved spatial autocorrelation analysis method for verification; this method is improved in the following ways: construct an adaptive spatial weight matrix, dynamically adjust the weight according to the spatial relationship between monitoring points and data similarity; introduce a local indicator calculation mechanism, and calculate the spatial correlation characteristics at different scales by the moving window method; design anisotropy-based spatial relationship description, considering the directional characteristics of spatial correlation; input the analysis results into the spatial clustering test unit, evaluate the local spatial correlation pattern by the improved LISA statistic, and output the spatial consistency indicator matrix;

[0153] S41b, input the spatial consistency index matrix and the original time series data into the time continuity analysis unit, and use the improved dynamic time warping algorithm to perform sequence alignment; the algorithm first constructs a multi-scale feature extraction mechanism to extract sequence features at different time resolutions; then introduces an adaptive penalty function to dynamically adjust the warping cost according to the local change characteristics of the sequence; finally, an entropy-based path selection strategy is designed to optimize the sequence alignment results; the warping results are input into the mutation detection unit, and by constructing multiple test statistics, abnormal change points in the time series are identified, and the time continuity evaluation index is output;

[0154] S41c, input the warning level, time continuity evaluation index and historical physical parameters into the physical rationality analysis unit, and adopt the improved physical model verification method; this method is realized through the following improvements: construct a multi-physical field coupling model to describe the interaction relationship between deformation, stress and environmental factors; introduce a parameter sensitivity analysis mechanism to identify physical parameters that have a significant impact on the warning results; design constraints based on energy conservation to verify the physical feasibility of the warning results; input the verification results into the consistency evaluation unit, quantify the degree of consistency between the warning results and the physical model by constructing a physical law compliance index, and output a physical rationality evaluation matrix;

[0155] S41d. Input the spatial consistency index matrix, temporal continuity evaluation index and physical rationality evaluation matrix into the comprehensive evaluation unit, and integrate them using an improved multi-criteria decision-making method. The improvements of this method include: constructing a hierarchical analysis structure to organize the verification indicators of each dimension in a hierarchical manner according to importance; introducing a fuzzy comprehensive evaluation mechanism to deal with the uncertainty of indicators of different dimensions; designing an information fusion strategy based on evidence theory to reasonably deal with the conflicts between indicators of various dimensions; inputting the evaluation results into the credibility calculation unit, and realizing the quantitative evaluation of the credibility of the warning results by constructing a dynamic weight system, and outputting the final verification evaluation report.

[0156] According to one aspect of the present application, step S43 is further:

[0157] S431, input the robustness index, sensitivity matrix and system parameter set into the target decomposition unit, use the principal component analysis method to extract the main influencing mode of the parameters, input the mode characteristics into the target function construction unit, construct the comprehensive performance index through the adaptive weight method, and obtain the multi-objective optimization function;

[0158] S432, inputting the multi-objective optimization function into the constraint generation unit, constructing a parameter boundary constraint model using support vector regression, inputting the constraint conditions into the feasible domain analysis unit, determining the parameter search space using the convex hull algorithm, and obtaining a complete description of the optimization problem;

[0159] S433, inputting the optimization problem description into the evolutionary algorithm unit, using the improved NSGA-III algorithm to generate a candidate solution set, inputting the candidate solution into the Pareto sorting unit, determining the priority of the solution by calculating the congestion distance, and obtaining a non-dominated solution set;

[0160] S434. Input the non-dominated solution set into the decision analysis unit, build a solution evaluation model based on TOPSIS, input the evaluation results into the parameter selection unit, determine the optimal parameter combination through the fuzzy hierarchical analysis method, and obtain the optimized system parameters.

[0161] In one embodiment of the present application, the specific implementation of the improved NSGA-III algorithm is to improve the special requirements of early warning system parameter optimization:

[0162] (1) Improvement of reference point generation: Adopt adaptive grid method to generate reference points in target space. The density of reference points is proportional to the gradient of objective function: density(r) = base_density * (1 + λ * ||▽f(r)||);

[0163] (2) Improvement of correlation distance calculation: Introduce target weight: w_i = exp(-μ*σ_i 2 ), where σ_i is the variance of the i-th target; weighted Euclidean distance: d(x, r) = sqrt(Σ(w_i * (fi(x)-ri) 2 ));

[0164] (3) Improved environmental selection: Construct a hierarchical Pareto frontier and calculate crowding: crowd(x) = min_r(d(x, r)) * Π(1 + ρ * ni), where ni is the number of neighbors of x on the i-th target; the selection strategy is: give priority to high-level individuals; sort by crowding within the same level; maintain the balance of reference point association;

[0165] (4) Improvement of genetic operators: adaptive crossover probability: pc = pc_base * (1 + η * (f_max - f_avg) / (f_max - f_min)); adaptive mutation intensity: σ_m = σ_base * exp(-κ * generation / max_generation). Where density(r) is the density value at the reference point r, base_density is the base reference point density, λ is the density adjustment coefficient, ▽f(r) is the gradient of the objective function at r, w_i is the weight value of the i-th target, μ is the weight calculation coefficient, σ_i is the standard deviation of the i-th target, d(x, r) is the weighted distance from the solution x to the reference point r, crowd(x) is the crowding value of the solution x, ρ is the crowding adjustment coefficient, pc is the adaptive crossover probability, pc_base is the baseline crossover probability, η is the crossover probability adjustment coefficient, f_max is the maximum value of the objective function, f_min is the minimum value of the objective function, f_avg is the average value of the objective function, σ_m is the adaptive mutation strength, σ_base is the baseline mutation strength, and κ is the mutation strength attenuation coefficient.

[0166] This embodiment realizes the performance optimization and parameter adjustment of the deformation early warning system of high core rockfill dam. The principal component analysis method is used to extract the main influencing mode of the parameters, and the adaptive weight method is combined to construct a comprehensive performance index, which overcomes the limitation of the traditional single-objective optimization method that it is difficult to balance multiple performance requirements; the support vector regression is introduced to construct the parameter boundary constraint model, and the parameter search space is determined by the convex hull algorithm, which effectively avoids the problem of the optimization process falling into the local optimum; the improved NSGA-III algorithm is used to generate the candidate solution set, and the priority of the solution is determined by combining the congestion distance calculation, which not only improves the optimization efficiency, but also ensures the diversity of the solution set; the optimal parameter combination is determined by the TOPSIS scheme evaluation model and the fuzzy hierarchical analysis method, and a decision-making mechanism that takes into account multi-objective trade-offs is established, making the optimization process of system parameters more scientific and explainable.

[0167] According to one aspect of the present application, step S44 is further:

[0168] S441. Input the optimized system parameters, historical database and newly added data into the incremental data preprocessing unit, and use the improved data quality assessment method to screen them; the method first builds a multidimensional anomaly detection model to identify anomalies in the newly added data; then introduces spatiotemporal consistency constraints to verify the continuity and rationality of the data; finally, designs an importance assessment mechanism based on information gain to screen data samples with update value; inputs the screening results into the data annotation unit, automatically annotates the newly added data through the improved semi-supervised learning method, and outputs the processed valid data set;

[0169] S442. Input the valid data set into the knowledge base structure update unit, and adopt the improved dynamic knowledge graph construction algorithm; the algorithm is improved in the following ways: construct a hierarchical knowledge representation framework to support multi-granularity knowledge storage; introduce a knowledge redundancy detection mechanism to avoid the accumulation of duplicate information; design a knowledge relevance analysis method to establish logical connections between knowledge units; input the updated knowledge structure into the consistency verification unit, ensure the integrity and consistency of the knowledge base by constructing knowledge graph verification rules, and output the optimized knowledge base structure;

[0170] S443, input the knowledge base structure and system parameters into the parameter adaptive adjustment unit, and use the improved incremental learning algorithm to update the parameters; the algorithm first builds a parameter importance evaluation model to identify the parameters that need to be updated; then introduces a progressive learning strategy to gradually adjust the parameter values; finally, designs a stability constraint mechanism to prevent drastic fluctuations during the parameter update process; inputs the update results into the parameter verification unit, and evaluates the effectiveness of the parameter adjustment by building multi-scenario test cases, and outputs a new system parameter set;

[0171] S444. Input the new system parameter set into the performance evaluation unit and adopt an improved system performance measurement method; this method is improved in the following ways: construct a multi-dimensional performance indicator system to comprehensively evaluate the early warning effect of the system; introduce a time-varying performance tracking mechanism to monitor the dynamic changes of system performance; design a reliability evaluation method based on incremental verification to ensure the stability of system updates; input the evaluation results into the optimization feedback unit, continuously improve the system performance by building a closed-loop optimization mechanism, and output the final system performance evaluation report and optimization suggestions.

[0172] This embodiment can effectively identify and clean anomalies in newly added data through a multi-dimensional anomaly detection model, spatiotemporal consistency constraints, and an importance evaluation mechanism based on information gain, ensuring the continuity and rationality of the data; the improved semi-supervised learning method further improves the accuracy of data annotation, generates high-quality valid data sets, and provides a solid foundation for subsequent analysis. The dynamic knowledge graph construction algorithm supports multi-granularity knowledge storage and avoids knowledge redundancy; through the knowledge relevance analysis method, a logical connection between knowledge units is established to ensure the integrity and consistency of the knowledge base; the introduction of knowledge graph verification rules further ensures the optimization and reliability of the knowledge base structure. The improved incremental learning algorithm gradually adjusts system parameters through a parameter importance evaluation model and a progressive learning strategy to ensure the stability of parameter updates; the stability constraint mechanism prevents drastic fluctuations during parameter updates, and multi-scenario test cases verify the effectiveness of parameter adjustments and output optimized system parameter sets. Construct a multi-dimensional performance indicator system to comprehensively evaluate the early warning effect of the system; introduce a time-varying performance tracking mechanism to monitor the dynamic changes of system performance in real time; use a reliability evaluation method based on incremental verification to ensure the stability of system updates; continuously improve system performance through a closed-loop optimization mechanism, and finally output a system performance evaluation report and optimization suggestions. This embodiment improves the reliability and accuracy of the deformation early warning system for high core wall rockfill dams through improvements in data quality, knowledge base optimization, parameter adaptive adjustment, and system performance evaluation, ensuring the credibility and persuasiveness of the early warning results.

[0173] According to one aspect of the present application, an early identification system for cracks on the crest of a high core rockfill dam comprises:

[0174] at least one processor; and,

[0175] a memory communicatively connected to at least one of the processors; wherein,

[0176] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the early identification method for cracks on the crest of a high core rockfill dam described in any one of the above embodiments.

[0177] The present invention realizes the intelligent processing of the whole process from data acquisition, feature extraction to early warning analysis by constructing a complete method system for early identification of cracks on the crest of a high core rockfill dam. In the data preprocessing stage, the quality and consistency of multi-source monitoring data are ensured through a unified spatiotemporal index structure and a multi-dimensional anomaly detection strategy; in the feature extraction stage, methods such as probability tensor decomposition and graph entropy matrix are used to realize the multi-scale characterization and effective fusion of deformation features; in the early warning analysis stage, the accuracy and reliability of the early warning results are improved by establishing a dynamic threshold network and a multi-dimensional verification mechanism. The present invention is particularly aimed at the early identification needs of cracks on the crest of the dam, and has significant advantages in processing massive multi-source heterogeneous data, capturing weak deformation signals, and evaluating system stability; through adaptive optimization and feedback correction mechanisms, the system can continuously learn and update the knowledge base to achieve dynamic improvement of early warning performance. The present invention not only solves the hysteresis and uncertainty problems existing in the traditional method in early crack identification, but also establishes a set of generalizable and scalable intelligent early warning analysis framework, which provides a strong technical support for the safety monitoring of high core rockfill dams.

[0178] It should be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. An early identification method for cracks on the crest of a high core rockfill dam, characterized in that: The steps include: S1. Obtain geographic spatial data and input it into a unified spatiotemporal index structure for labeling to obtain labeled data; based on the labeled data, perform outlier cleaning, standardization processing and adaptive weight fusion to obtain a registered data set; The geospatial data include InSAR data, drone oblique photography data and 3D laser scanning data; S2, based on the registered data set, obtain the local feature set and its importance through probability tensor decomposition; Generate a global feature matrix based on the local feature set and its importance; Extract the time series feature vector based on the global feature matrix and the spatiotemporal index structure; perform feature fusion based on the local feature set, the global feature matrix and the time series feature vector to obtain a fused feature set; S3. Based on the fusion feature set, a multi-level early warning indicator set is constructed; Generate stability indicator sequence based on early warning indicator set and pre-stored historical data; Based on the stability indicator sequence, a dynamic threshold network is constructed through adaptive updating; Based on the stability index sequence and dynamic threshold network, the warning level and credibility value are obtained through probability combination judgment; S4. Based on the warning level and credibility value, generate the verification matrix and consistency coefficient, and calculate the robustness index and sensitivity matrix; Based on the robustness index and sensitivity matrix, multi-objective optimization is performed to obtain the optimized system parameters; new data is obtained, and new system parameters and a new knowledge base are obtained through dynamic updates based on the new data, the optimized system parameters and the pre-stored historical database; Step S1 is further as follows: S11, obtaining geographic spatial data from the dam monitoring system and inputting it into a spatiotemporal index processing unit; in the spatiotemporal index processing unit, establishing a time index table based on pre-stored timestamp information, establishing a spatial index table based on pre-stored spatial coordinate information, and combining the time index table and the spatial index table to generate a unified spatiotemporal index structure; Based on the spatiotemporal index structure, the spatiotemporal location information of each data point in the geographic spatial data is marked to obtain the marked data; S12, inputting the marked data into the dynamic window matrix, generating a local feature vector by calculating the local mean, standard deviation, skewness and kurtosis of the data within a predetermined range around each data point; inputting the generated local feature vector into the abnormality determination function, and calculating the abnormality degree using an adaptive weight coefficient; based on the abnormality degree, determining whether the data point is an abnormal value, and obtaining a determination result; inputting the determination result into the data cleaning unit, removing the abnormal value in the data point, and obtaining a cleaned data set; S13, input the cleaned data set into the histogram analysis unit to calculate the distribution characteristics of the data; based on the distribution characteristics, determine the data segmentation boundaries through local extreme points; input the data segmentation boundaries and the cleaned data set into the segmentation standardization processing unit to calculate the statistical characteristics of each data segment, including the mean, standard deviation and skewness; based on Statistical features, standardize the data and generate a standardized data set; S14, input the standardized data set and the spatiotemporal index structure into the spatiotemporal weight calculation unit, calculate the time weight coefficient by the time interval, and calculate the spatial weight coefficient by the spatial distance; perform Hadamard product operation on the time weight coefficient and the spatial weight coefficient to obtain the fusion weight coefficient; input the fusion weight coefficient and the standardized data set into the data resampling unit, resample the data based on the fusion weight coefficient, and generate the registered data set; Step S2 is further as follows: S21, input the registered data set into the local feature tensor construction unit, calculate the deformation, curvature and gradient value of each data point, and construct the local feature tensor; input the local feature tensor into the probability tensor decomposition unit, calculate the eigenvalue and eigenvector, and generate the tensor decomposition result; input the tensor decomposition result into the feature importance calculation unit, calculate the importance of each feature by the ratio of the eigenvalue and the residual size, and output the local feature set and its corresponding importance value; S22, inputting the local feature set and its corresponding importance value into the adjacency matrix construction unit, calculating the connection strength between nodes through the Euclidean distance and importance value between the local features, and generating an adjacency matrix; Input the adjacency matrix into the graph entropy calculation unit to calculate the graph entropy value of each node; Input the entropy value of the graph into the entropy matrix construction unit, and generate the global feature matrix through the combination operation and gradient correction of the entropy value; S23, inputting the global feature matrix and the spatiotemporal index structure into the time series probability matrix construction unit, calculating the conditional probability between adjacent time points, and generating a time series probability matrix; inputting the time series probability matrix into the feature decomposition unit, and obtaining the main components of the time series features through basis function decomposition; inputting the main components into the stability analysis unit, and obtaining the time series feature vector by calculating the gradient ratio; S24, inputting the local feature set, the global feature matrix and the time series feature vector into the probability combination matrix construction unit, and forming the probability combination matrix by row combination; inputting the probability combination matrix and the pre-stored historical tag data into the weight optimization unit, and obtaining the combination weight by constructing and solving the least squares problem with L1 norm constraint; inputting the combination weight and the probability combination matrix into the feature reconstruction unit, and obtaining the fused feature set by weighted combination and gradient correction; Step S3 is further as follows: S31, inputting the fused feature set into a feature stratification unit, respectively calculating a microscopic change index, a mesoscopic evolution index, and a macroscopic stability index, and generating a feature stratification matrix of at least three layers; Input the feature hierarchical matrix into the inter-layer correlation analysis unit, and calculate the inter-layer correlation coefficient through matrix operation and gradient correction; Based on the inter-layer correlation coefficient, characteristic hierarchical matrix and pre-stored historical warning data, an optimization problem with double sparse constraints is constructed and solved to obtain a set of warning indicators; the microscopic change indicators include deformation gradient and strain rate; the mesoscopic evolution indicators include cumulative displacement and deformation acceleration; the macroscopic stability indicators include overall stiffness and energy dissipation; S32, inputting the early warning indicator set and the pre-stored historical database into the state transition probability calculation unit, calculating the transition probability between different states, and generating a state transition probability matrix; The state transition probability matrix at adjacent moments is input into the stability matrix calculation unit, and the stability matrix is ​​obtained through matrix inversion and product operations; Input the stability matrix into the eigenvalue analysis unit, calculate the matrix eigenvalues, and construct the stability index sequence; S33, inputting the early warning indicator set and the stability indicator sequence into the threshold matrix initialization unit, and constructing an initial threshold matrix for each early warning indicator and early warning level; The initial threshold matrix is ​​input into the adaptive update unit, and the threshold is updated by calculating the gradient and second-order gradient of the loss function to obtain an updated threshold matrix; the updated threshold matrix is ​​input into the threshold network construction unit, and the association weights between threshold nodes are calculated to construct a dynamic threshold network; The loss function includes prediction error term, sparse constraint term and total variation constraint term; S34, inputting the warning indicator set, stability indicator sequence and dynamic threshold network into the warning probability calculation unit to calculate the probability vector of each warning level; The probability vector is input into the combined discrimination unit, and the combined discrimination result is obtained through multi-order probability combination and gradient correction; the combined discrimination result is input into the warning level determination unit, and the warning level is determined by the maximum probability selection, and the credibility value of the warning result is calculated at the same time; S35, generating a warning parameter set based on the warning indicator set, the stability indicator sequence, the dynamic threshold network, the warning level and the credibility value; Step S4 is further as follows: S41, inputting the warning level, the credibility value and the pre-stored historical database into the verification feature construction unit, respectively calculating the spatial consistency feature, the temporal continuity feature and the physical rationality feature, and combining them to generate a verification feature matrix; Input the verification feature matrix into the verification score calculation unit to calculate the weight coefficient; Based on the weight coefficient, the verification score is obtained through weighted combination; The verification score is input into the consistency test unit, and the consistency coefficient is obtained by calculating the difference between the observed consistency and the random consistency; S42. Based on the warning parameter set, the verification feature matrix and the consistency coefficient, the central moment and the cross moment of each parameter are calculated to construct a probability moment tensor; The probability moment tensor is input into the robustness analysis unit, and the robustness index is obtained by calculating the ratio of the Frobenius norm and the trace of the probability moment tensor; the warning parameter set and the warning level are input into the sensitivity calculation unit, and the sensitivity index of each parameter is obtained by calculating the standardized partial derivative to form a sensitivity matrix; S43, input the robustness index, the sensitivity matrix and the pre-stored system parameter set into the objective function construction unit, respectively construct the warning accuracy function, the system complexity function and the calculation efficiency function; input the warning accuracy function, the system complexity function, the calculation efficiency function and the boundary constraints, the physical meaning constraints and the balance constraints of the system parameters into the multi-objective optimization unit, construct the dominance relationship matrix of the solution, and calculate the congestion distance of the solution; sort the solutions based on the dominance relationship matrix and the congestion distance to obtain the optimized system parameters; S44, obtaining new data, inputting the optimized system parameters, the historical database and the new data into the incremental learning matrix construction unit, and using the forgetting factor to calculate the retention weight of the historical knowledge; Based on the retained weights and the gradient information of the newly added data, an incremental learning matrix is ​​constructed; the incremental learning matrix and the newly added data are input into the knowledge base update unit, and the new knowledge is integrated into the existing knowledge base through the knowledge fusion operator to obtain a new knowledge base; The new knowledge base is input into the parameter adjustment unit, and the optimized system parameters are dynamically adjusted according to the gradient information of the new knowledge base to obtain new system parameters; Step S12 is further as follows: S121, inputting the marked data into the data partition calculation unit respectively, and obtaining the initial partition size parameter by calculating the local data density distribution and the spatial autocorrelation coefficient; based on the initial partition size parameter, using the recursive quadtree partitioning algorithm to generate a multi-scale window partitioning result; Based on the multi-scale window division results, the multi-scale features of each window are extracted to obtain an enhanced feature set; based on the enhanced feature set, the mean, standard deviation, skewness and kurtosis are calculated to generate a local feature vector set; S122, based on the local feature vector set, perform dimensionality reduction by kernel principal component analysis to obtain a reduced-dimensional feature matrix; based on the reduced-dimensional feature matrix, use a kernel density estimation method to calculate the multidimensional joint distribution of features and generate a probability density function; Based on the probability density function, the anomaly judgment threshold of each dimension is determined through the maximum entropy principle, and a multi-dimensional threshold vector is output; based on the multi-dimensional threshold vector, it is corrected through extreme value statistical theory to obtain an optimized threshold set; S123, based on the local feature vector set and the optimized threshold set, using the improved DBSCAN algorithm to perform hierarchical clustering to obtain multi-level data clusters; Based on multi-level data clusters, the local anomaly factor method is used to calculate the anomaly degree of each data point and generate an anomaly degree matrix; Based on the abnormality degree matrix, the abnormality level is determined through the fuzzy inference system, and the abnormality marking result is output; Based on the anomaly marking results, a pre-configured Markov random field model is used for optimization to obtain the corrected anomaly marking; S124, based on the corrected abnormal markers, extract the abnormal data points and their neighborhood information to generate a data set to be repaired; based on the data set to be repaired, extract the local manifold structure through tensor singular value decomposition to obtain a structural feature tensor; Based on the structural feature tensor, the local weighted regression method is used to reconstruct the abnormal data and generate a repaired data set. Based on the repaired data set, the reliability of the repair result is verified by calculating the reconstruction error and local consistency index, and the cleaned data set is output. Step S13 is further as follows: S131, based on the cleaned data set, using the minimum description length criterion to determine the optimal number of histogram intervals; based on the optimal number of histogram intervals, using a Gaussian kernel function with a variable bandwidth to calculate the probability density distribution; Based on the probability density distribution, the local extreme value features are extracted through continuous wavelet transform to obtain the data distribution characteristic curve; S132, based on the data distribution characteristic curve, using a peak detection algorithm based on duration, identifying local maximum and minimum points, and generating an extreme point sequence; Based on the extreme point sequence, the extreme points are grouped by the density peak clustering algorithm to obtain the grouping results; Based on the grouping results, the maximum inter-class variance criterion is used to determine the final data segmentation boundary; S133, based on the data segmentation boundaries and the cleaned data set, using a robust statistical method to calculate the statistical characteristics of each data segment; based on The statistical features are processed using the Huber estimator to obtain the corrected segmented statistical features; S134, based on the corrected segmented statistical features and the cleaned data set, construct a standardized transformation function based on skewness correction, perform standardized transformation on each data segment, and obtain a preliminary standardized result; based on the preliminary standardized result, use a local polynomial regression method to smooth the transition area between segments, and obtain a standardized data set; Step S14 is further as follows: S141. Based on the standardized data set and spatiotemporal index structure, a dynamic time warping algorithm is used to calculate the time correspondence between data sequences; based on the time correspondence, an optimal time window size is determined by an information entropy criterion to obtain a time weight coefficient; S142, based on the standardized data set and spatiotemporal index structure, using an iterative closest point algorithm to calculate the spatial correspondence between different data sources; Based on the spatial correspondence, an adaptive bandwidth matrix is ​​constructed according to the direction of the principal curvature to obtain the spatial weight coefficient; S143, constructing a sparse constraint term based on the L1 norm based on the time weight coefficient and the space weight coefficient; Based on the sparse constraint term, the fusion weight coefficient is obtained through iterative optimization; S144, based on the fusion weight coefficient and the standardized data set, calculating the optimal affine transformation parameters of each data point; based on the optimal affine transformation parameters, using a variational reconstruction method to resample the data to obtain a registered data set; Step S24 is further as follows: S241, based on the local feature set, the global feature matrix and the time series feature vector, a power law transformation method is used to perform scale standardization to obtain standardized features; based on the standardized features, a mutual information matrix between features is calculated by a kernel density estimation method to obtain an initial probability combination matrix; S242, based on the initial probability combination matrix and the pre-stored historical label data, construct a regularization term based on the elastic network to obtain an optimization problem; Based on the optimization problem, the optimal regularization parameter is determined through cross-validation to obtain the combined weight coefficient; S243, based on the combined weight coefficient, dynamically update the weight distribution using the Hedge algorithm to obtain an updated weight; Based on the updated weights, the stability of the weights is evaluated by the bootstrap method to obtain the revised combined weights; S244, based on the modified combination weights and the initial probability combination matrix, using Tucker decomposition to extract the high-order association pattern of the features to obtain a decomposition result; Based on the decomposition results, feature reconstruction is performed through a variational autoencoder to obtain the final fusion feature set.

2. An early identification system for cracks on the crest of a high core rockfill dam, 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 that can be executed by the processor, and the instructions are used to be executed by the processor to implement the early identification method of cracks on the crest of a high core rockfill dam according to claim 1.

Images