A geotechnical anchor rod uplift bearing capacity detection data intelligent processing method

Abstract

The present application belongs to the technical field of data processing, and particularly relates to a kind of geotechnical anchor rod uplift bearing capacity detection data intelligent processing method, comprising: obtaining force-displacement-acoustic emission data, extracting mechanical, displacement and acoustic emission parameters to construct a standardized feature vector set and associate with bearing capacity label;Recursive construction of approximate nearest neighbor search tree: decompose the covariance matrix to obtain the first principal component direction and eigenvalue, calculate the dimension advantage measure, take the coordinate axis direction of the maximum dimension of the variance and sensitivity coefficient product, fuse to construct a hybrid split direction vector, take the local minimum value or mean value as the split point by kernel density estimation to construct a hyperplane, stop generating leaf nodes;The feature vector of the anchor rod to be analyzed is input into the approximate nearest neighbor search tree to obtain K nearest neighbors, and the bearing capacity state of the anchor rod to be analyzed is determined based on the distance reciprocal weighted voting. The present application effectively suppresses abnormal interference, and improves the accuracy and reliability of anchor rod bearing capacity determination.

Description

Technical Field

[0001] This invention relates to the field of data processing technology. More specifically, this invention relates to an intelligent processing method for testing the pull-out bearing capacity of soil and rock anchors. Background Technology

[0002] Rock and soil anchors are widely used support structures in geotechnical engineering. By anchoring the anchor rods into the rock and soil mass, they effectively improve the stability of the rock and soil mass and prevent engineering disasters such as slope slippage and foundation pit collapse. Pull-out bearing capacity is a core indicator for evaluating the effectiveness of anchor support and is directly related to the safety and reliability of the engineering structure. Currently, anchor pull-out bearing capacity testing is mainly achieved through on-site pull-out tests, during which multi-source data such as force, displacement, and acoustic emission are collected simultaneously. These data contain information on the stress evolution law, deformation response characteristics, and damage development state of the anchor. Through intelligent analysis of multi-source test data, accurate assessment of the anchor pull-out bearing capacity state can be achieved. The approximate nearest neighbor search algorithm is suitable for rapid querying in a multi-dimensional feature space. It can construct an approximate nearest neighbor search tree model based on spatial partitioning through a multi-dimensional feature set, and complete nearest neighbor retrieval and state matching based on the test data. Therefore, it is widely used in the field of intelligent processing of engineering test data.

[0003] However, the multi-source detection data generated by anchor pull-out tests are characterized by multi-source heterogeneity, high feature dimensionality, and significant dimensional differences. Data from different sources exhibit significant differences in physical meaning, numerical range, and variation patterns, and there are complex correlations among the various feature dimensions. For example, mechanical parameters reflect the anchor's load-bearing capacity, displacement parameters reflect its deformation characteristics, and acoustic emission parameters reflect the internal damage evolution process of the anchor; these three are interconnected but each has its own independent physical meaning. This complex data characteristic places higher demands on the spatial partitioning quality of the approximate nearest neighbor search tree, requiring the search tree to simultaneously consider the overall distribution trend of the data, the local sensitivity of key feature dimensions, and the density clustering characteristics of the data in the feature space.

[0004] Chinese patent application CN116257528A discloses a method for calculating the nearest neighbor sample points of a target point based on KDTree. This method summarizes the k-dimensional coordinates of all sample points, determines the first separation dimension based on the maximum variance, obtains the first left subtree sample point pool and the first right subtree sample point pool based on the first separation dimension, and continues to repeat the process of determining the separation dimension by the maximum variance in each sample point pool until there is only one node in the bottom sample point pool obtained by segmentation, thus obtaining the KDTree search structure. Finally, feature matching is performed in the constructed KDTree search structure to obtain the nearest neighbor sample point to the target point.

[0005] The aforementioned existing technologies, when constructing an approximate nearest neighbor search tree, recursively split only along the single coordinate axis with the largest variance. While this splitting method can ensure the balance of the tree to some extent, it is difficult to simultaneously consider the overall trend of data change and the local sensitivity of key feature dimensions. In multi-source anchor bolt detection data, some feature dimensions that are important indicators of bearing capacity status may have small variances in local data subsets, making them unsuitable for selection as splitting directions, thus weakening these key feature information. Furthermore, this method uses the median, the middle sample point on the first separating dimension, as the splitting point. This strategy only considers the data's ranking information and fails to reflect the density clustering characteristics and probability distribution differences of anchor bolt detection data in the feature space. This can easily fragment the natural clustering structure, causing samples with similar bearing capacity states to be assigned to different regions. These problems severely affect the spatial partitioning quality of the search tree, making it difficult for the model to stably obtain a set of highly similar nearest neighbor feature vectors when processing the anchor bolt data to be analyzed, thereby limiting the accuracy and reliability of anchor bolt pull-out bearing capacity status identification. Summary of the Invention

[0006] To address the technical problems of traditional approximate nearest neighbor search trees, which have a single splitting direction and rely solely on the median of the sorting, making it difficult to consider both the global distribution trend of the data and the local sensitivity of key dimensions, and also failing to reflect differences in density clustering, resulting in the fragmentation of samples with similar states and restricting the accuracy of bearing capacity identification, this invention provides an intelligent processing method for soil and rock anchor pull-out bearing capacity detection data.

[0007] This invention provides an intelligent processing method for testing the pull-out bearing capacity of soil and rock anchors, comprising: S1: acquiring the raw force-displacement-acoustic emission data of the anchor test, extracting mechanical parameters, displacement parameters, and acoustic emission parameters to construct an initial feature vector set, standardizing the initial feature vector set to obtain a feature vector set, and associating it with a bearing capacity status label; S2: recursively partitioning the feature vector set to construct an approximate nearest neighbor search tree, the construction of non-leaf nodes including: calculating the covariance matrix of the feature vector subset of the current node and performing eigenvalue decomposition, obtaining the direction of the first principal component and the eigenvalue; calculating the dimensionality advantage measure proportional to the variance explanation rate corresponding to the eigenvalue; calculating the product of the variance of each feature dimension and a preset sensitivity coefficient. The product is calculated, and the direction of the coordinate axis corresponding to the maximum value is taken as the weighted coordinate axis direction. Based on the dimensionality advantage metric, a weighted sum of the first principal component direction and the weighted coordinate axis direction is constructed and normalized as a mixed splitting direction vector. The feature vector subset is projected onto the mixed splitting direction vector, and the kernel density is estimated on the projected data. The local minimum or mean of the probability density function is taken as the splitting point, and a splitting hyperplane is constructed through the splitting point. Leaf nodes are generated when the preset stopping condition is met. S3: The feature vector of the anchor to be analyzed is input into the approximate nearest neighbor search tree to obtain K nearest neighbor feature vectors. The bearing capacity state of the anchor to be analyzed is determined by weighted voting based on the inverse distance of the bearing capacity state labels of the K nearest neighbor feature vectors.

[0008] By employing the above technical solution, this invention constructs a hybrid splitting direction by fusing the data-driven first principal component direction with the domain knowledge-guided weighted coordinate axis direction. The splitting point is then determined by using the local density minima obtained from kernel density estimation along this direction. This allows the spatial partitioning of the approximate nearest neighbor search tree to simultaneously consider the global distribution trend of multi-source data and the local sensitivity of key feature dimensions. Furthermore, the splitting occurs at the natural clustering boundaries of the data, reducing the fragmentation of clusters of samples with similar bearing capacity states. Based on this, a weighted voting mechanism based on the inverse distance is used to determine the bearing capacity state, enabling highly similar nearest neighbor samples to contribute more significantly to the classification decision and reducing interference from distant abnormal neighbors near the classification boundary. This, in turn, improves the accuracy and reliability of anchor bolt pull-out bearing capacity state identification.

[0009] Preferably, the process of acquiring the raw force-displacement-acoustic emission data of the anchor bolt detection, extracting mechanical parameters, displacement parameters, and acoustic emission parameters to construct an initial feature vector set, standardizing the initial feature vector set to obtain a feature vector set, and associating it with a bearing capacity status label includes: extracting the maximum test load from the raw data as a mechanical parameter; extracting the displacement value recorded synchronously with the maximum test load as a displacement parameter; statistically analyzing the cumulative number of acoustic emission impacts within a preset time window as an acoustic emission parameter; splicing the mechanical parameters, displacement parameters, and acoustic emission parameters and summarizing them to obtain the initial feature vector set; subtracting the corresponding statistical mean from the raw values ​​of the initial feature vector set and dividing by the standard deviation to obtain the feature vector set; and reading the safety classification category identifier from a known detection database as a label field and associating it with the corresponding feature vector for storage, thus completing the association of the bearing capacity status label.

[0010] By adopting the above technical solution, this invention constructs a three-parameter feature vector using the maximum test load, synchronous displacement value, and cumulative acoustic emission impact count. This vector characterizes the anchor bolt's load-bearing characteristics through three physical dimensions: mechanics, deformation, and damage. Z-Score standardization eliminates numerical scale differences between different physical dimensions, ensuring that each feature dimension is balanced in subsequent covariance calculations and distance measurements. This prevents large-scale parameters from dominating distance calculations and overshadowing the discriminative information of other dimensions.

[0011] Preferably, the step of calculating the covariance matrix and performing eigenvalue decomposition of the current node's feature vector subset to obtain the first principal component direction and eigenvalue includes: constructing a decentralized matrix of the feature vector subset; generating a covariance matrix based on the decentralized matrix and its transpose matrix using the sample number minus 1 as a normalization factor; performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding unit eigenvectors; and extracting the maximum eigenvalue and its corresponding unit eigenvector as the first principal component direction and eigenvalue.

[0012] By employing the above technical solution, this invention eliminates the spatial shift of data clusters through decentralization, constructs an unbiased estimate of the covariance matrix using a normalization factor of 1 minus the number of samples, and extracts the direction of the first principal component using eigenvalue decomposition. This processing method can accurately capture the maximum data scattering axis of the feature vector subset within the current node, providing a reliable data-driven basis for the construction of subsequent hybrid splitting directions.

[0013] Preferably, the step of kernel density estimation of the projected data, taking the local minimum point or mean of the probability density function as the split point, includes: continuously estimating the probability density function of the projected data using a Gaussian kernel function; calculating the first and second derivatives of the probability density function; taking the projected scalar coordinates corresponding to the local minimum point that satisfies the condition that the first derivative is equal to zero and the second derivative is greater than zero as the split point; if no local minimum point exists, taking the arithmetic mean of the projected data as the split point.

[0014] By employing the above technical solution, this invention utilizes a Gaussian kernel function to perform nonparametric probability density estimation on projected data. It accurately identifies local minima at density troughs as splitting points using first and second derivative conditions, and backtracks to the mean in unimodal distributions. This splitting strategy based on data density distribution allows the splitting hyperplane to be divided along the natural sparse bands of samples in the feature space, effectively protecting the cluster integrity of samples with similar bearing capacity states.

[0015] Preferably, the step of inputting the feature vector of the anchor to be analyzed into an approximate nearest neighbor search tree to obtain K nearest neighbor feature vectors includes: traversing the feature vector of the anchor to be analyzed downwards from the root node of the approximate nearest neighbor search tree, calculating the projection result of the feature vector of the anchor to be analyzed and the mixed splitting direction vector of the current node, comparing the projection result with the splitting point to enter the corresponding child node; recording the side branch nodes that have not yet been visited and their corresponding distances based on a priority queue; calculating the Euclidean distance to update the candidate nearest neighbor set after reaching the leaf node; popping the node from the priority queue and backtracking the search to extract the K feature vectors with the smallest distance.

[0016] Preferably, the step of determining the bearing capacity state of the anchor to be analyzed by weighted voting based on the reciprocal of the distance between the bearing capacity state labels of the K nearest neighbor feature vectors includes: calculating the Euclidean distance between the feature vector of the anchor to be analyzed and each nearest neighbor feature vector; weighting and summing the nearest neighbor feature vectors belonging to the same category according to the reciprocal of the Euclidean distance to obtain a weight score; and taking the category corresponding to the maximum weight score as the bearing capacity state of the anchor to be analyzed.

[0017] Preferably, the step of calculating the product of the variance of each feature dimension and the preset sensitivity coefficient, and taking the coordinate axis direction corresponding to the maximum value as the weighted coordinate axis direction, includes: selecting the feature dimension with the largest product result as the target label dimension; constructing a zero basis vector with the same dimension as the feature vector, and setting the component corresponding to the target label dimension to 1 as the weighted coordinate axis direction.

[0018] By employing the above technical solution, this invention amplifies the influence of key feature dimensions, which are of significant indicative importance to bearing capacity status, in determining the splitting direction by multiplying the sample variance of each feature dimension with a preset engineering sensitivity coefficient. Even if the variance of a certain core detection index is small in a local data subset, it can still dominate the splitting direction of the search tree in that local space with the help of the pre-calibrated sensitivity coefficient, thus avoiding the dilution of key engineering feature information during the recursive partitioning process.

[0019] Preferably, the construction of the weighted sum of the first principal component direction and the weighted coordinate axis direction based on the dimensional advantage metric includes: if the inner product of the first principal component direction and the weighted coordinate axis direction is less than 0, then the sign of the weighted coordinate axis direction component is reversed; the first principal component direction and the sign-processed weighted coordinate axis direction are weighted and summed according to the dimensional advantage metric and the remaining weight obtained by subtracting the dimensional advantage metric from 1.

[0020] Preferably, the calculation of the dimensionality advantage measure, which is proportional to the variance explained by the eigenvalue, includes: dividing the largest eigenvalue by the sum of all eigenvalues ​​of the covariance matrix to obtain the dimensionality advantage measure.

[0021] Preferably, the acquisition of the raw force-displacement-acoustic emission data of the anchor bolt detection includes: continuously acquiring mechanical and displacement signals at low frequency using a tension sensor and a linear variable differential transformer; acquiring acoustic emission signals using a high-frequency transient capture instrument; and generating raw data by synchronously aligning each signal based on a timestamp.

[0022] The technical solution of the present invention has the following beneficial technical effects: This invention extracts mechanical parameters, displacement parameters, and acoustic emission parameters from force-displacement-acoustic emission multi-source detection data to construct standardized feature vectors, eliminating numerical scale differences between different physical dimensions and providing a unified data representation basis for nearest neighbor retrieval. During the construction of the approximate nearest neighbor search tree, a hybrid splitting direction is formed by dynamically fusing the direction of the first principal component with the weighted coordinate axis direction amplified by the engineering sensitivity coefficient through a dimensionality advantage metric. This takes into account both the global distribution trend of the data and the local sensitivity of key feature dimensions, avoiding the search tree ignoring a key dimension due to its small variance in a local subset. Simultaneously, a continuous probability density distribution of the projected data is constructed along the hybrid splitting direction using Gaussian kernel density estimation. Splitting points are selected at density troughs, causing the splitting hyperplane to divide along the natural clustering boundary of the data, reducing direct fragmentation of sample clusters with similar bearing capacity states.

[0023] Furthermore, in the bearing capacity state determination stage, the present invention adopts a weighted voting mechanism based on the reciprocal of distance, which makes the highly similar neighbor samples that are closer to the feature vector space of the anchor to be analyzed contribute more to the classification decision, reduces the interference of distant abnormal neighbors near the classification boundary, and improves the accuracy and reliability of anchor pull-out bearing capacity state identification. Attached Figure Description

[0024] Figure 1 This is a flowchart of an intelligent processing method for testing the pull-out bearing capacity of rock and soil anchors according to the present invention; Figure 2 This is a diagram illustrating the weighted evaluation of feature dimension variance and sensitivity coefficient; Figure 3This is a schematic diagram illustrating nuclear density estimation and splitting point selection; Figure 4 This is a schematic diagram comparing the effects of ablation experiments. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses an intelligent processing method for testing the pull-out bearing capacity of soil and rock anchors, referring to... Figure 1 This includes steps S1-S3: S1: Data preprocessing and feature construction.

[0027] This step is used to extract multi-source parameters from the raw force-displacement-acoustic emission data collected from the anchor pull-out test and construct a standardized set of feature vectors to eliminate the dimensional differences between different physical quantities. At the same time, it associates the bearing capacity state label to provide a unified data representation basis for the subsequent construction of the approximate nearest neighbor search tree and nearest neighbor retrieval.

[0028] The input to this step is the raw force-displacement-acoustic emission time series data synchronously collected by sensors, and the output is a set of standardized feature vectors and their associated bearing capacity state labels, which are used for the recursive partitioning construction of the approximate nearest neighbor search tree in S2.

[0029] First, time-series signals during the pull-out test are acquired in real time using a tensile sensor, a linear variable differential transformer, and an acoustic emission instrument. During this process, differentiated sampling frequencies are configured according to the sensor type; for example, the tensile sensor and the linear variable differential transformer employ... to The sampling frequency is used for continuous low-frequency acquisition, while the acoustic emission instrument uses... to High-frequency transient detection is performed using a specific sampling frequency. Each acquisition channel is synchronized via a unified timestamp, allowing the reading of the structured raw data file. Mechanical parameters, displacement parameters, and acoustic emission parameters are extracted from the raw data. The mechanical parameters reflect the anchor bolt's load-bearing capacity, the displacement parameters reflect its deformation characteristics, and the acoustic emission parameters reflect the internal damage evolution process of the anchor bolt. These parameters within the same sampling time window are concatenated and combined in a fixed order to construct an initial feature vector, which is then aggregated to obtain an initial feature vector set.

[0030] In one embodiment, the specific method for extracting the above parameters is as follows: extract the maximum test load from the original data as a mechanical parameter; extract the displacement value recorded synchronously with the maximum test load as a displacement parameter; and statistically analyze the cumulative number of acoustic emission impacts within a preset time window as an acoustic emission parameter. The mechanical parameter, displacement parameter, and acoustic emission parameter are concatenated in a preset order to form an initial feature vector, and the initial feature vector set is obtained by summarizing these parameters. For the mechanical parameter, the load curve of the entire pull-out test cycle is analyzed, and the maximum test load corresponding to the peak value of the curve is extracted; for example, if the maximum test load under the extreme state is identified as... This value is then used as the benchmark for evaluating pull-out resistance. Simultaneously, using timestamp alignment technology, displacement values ​​recorded synchronously with the maximum test load are extracted as displacement parameters. For acoustic emission parameters, a preset time window is set, preferably with a width ranging from 0.1 s to 1 s; for example, the window width can be taken as... The number of acoustic emission impacts exceeding a preset amplitude threshold within the window is counted using an acoustic emission acquisition card. This number is used as the acoustic emission parameter; for example, the preset amplitude threshold is 40 dB. The extracted parameters are then sequentially concatenated according to a pre-written data structure, such as [maximum load, displacement value, cumulative number of acoustic emission impacts], to form the initial three-dimensional feature vector of a single anchor rod. The summation of multiple test anchor rods forms a scale... The initial set of feature vectors.

[0031] In another embodiment, mechanical parameters can be expanded to extract maximum tensile force and initial stiffness parameters, displacement parameters can be expanded to extract residual deformation and plastic strain parameters, and acoustic emission parameters can be expanded to extract ring count, cumulative energy, and peak frequency parameters. These parameters within the same sampling time window are then concatenated in a fixed order to construct an initial feature vector. The selection of parameters in different embodiments can be determined based on actual engineering testing requirements and sensor configuration. As long as the extracted parameters belong to the three categories of mechanics, displacement, and acoustic emission, they fall within the protection scope of this method.

[0032] Subsequently, the feature dimensions in the initial feature vector set are standardized to eliminate dimensional differences. This standardization is necessary because mechanical parameters, displacement parameters, and acoustic emission parameters each have different physical dimensions and value ranges. For example, the maximum test load is on the order of hundreds. The displacement values ​​are on the order of millimeters, and the cumulative number of acoustic emission impacts is a dimensionless count. Direct splicing will lead to an imbalance in the numerical scale of different feature dimensions, affecting the rationality of subsequent covariance calculations and distance measurements. When calling the Z-Score normalization algorithm, the statistical mean and standard deviation of each feature dimension in the initial feature vector set are calculated. The original values ​​are subtracted from the statistical mean and then divided by the standard deviation, thereby transforming the data of each dimension into a mean. variance is The standard dimensionless eigenvectors are used to obtain the standardized set of eigenvectors.

[0033] Next, the safety classification category identifier is read from the known detection database and associated with the corresponding bearing capacity status label through a unique identification code. In this embodiment, the classification criteria for the bearing capacity status label are specifically set as follows: the ratio of the cumulative displacement increment of the anchor bolt under each load level to the load increment is defined as the compliance coefficient; when the compliance coefficient remains in a linear stable period and no sudden high-energy acoustic emission impact occurs, it is determined to be in a safe state; when the load reaches more than 80% of the design limit load, and the compliance coefficient increase under two adjacent load levels exceeds 50%, or the acoustic emission cumulative energy curve shows a significant inflection point, it is determined to be in a critical state; when the pull-out force drops by more than or equal to 10% of the peak load, or the residual displacement exceeds the preset engineering limit allowable value, it is determined to be in a failure state. The above quantitative judgment results are converted into enumerated type values ​​and stored as a label field associated with the corresponding feature vector. It should be noted that this label field does not participate in feature standardization, covariance calculation, and distance calculation, but is only used for node data organization in the supervised tree construction of S2 and weighted voting classification in S3. This enables the association between multi-source detection data and bearing capacity status labels, generating a training dataset that can be input into the search tree model.

[0034] Thus, this step transforms the raw force-displacement-acoustic emission multi-source data into a standardized set of feature vectors and associates them with bearing capacity state labels, providing standardized input data for the construction of the search tree in S2.

[0035] S2: Construction of an approximate nearest neighbor search tree.

[0036] This step is used to recursively partition the standardized feature vector set output by S1 to construct an approximate nearest neighbor search tree. By fusing the direction of the first principal component with the direction of the weighted coordinate axis to form a mixed splitting direction, and using kernel density estimation to determine the splitting point, the spatial partitioning of the search tree takes into account both the overall distribution trend of the data and the local sensitivity of key feature dimensions. At the same time, it splits at the natural clustering boundary according to the density distribution characteristics of the projected data, reducing the direct cutting of adjacent sample clusters.

[0037] The input to this step is the set of standardized feature vectors produced by S1, and the output is an approximate nearest neighbor search tree that can be used for fast nearest neighbor retrieval, which is then used in S3 to analyze the anchor feature vectors. Used for nearest neighbor retrieval.

[0038] Define a recursive tree construction function based on a priority strategy, starting from the root node containing all normalized feature vectors, and perform the following construction process for each non-leaf node.

[0039] 1. Covariance matrix calculation and first principal component direction extraction.

[0040] This sub-step is used to analyze the direction of the maximum divergence of the data distribution in the current node, i.e. the direction of the first principal component, to provide a data-driven basis for determining the subsequent splitting direction.

[0041] For the standardized feature vector subset passed to the current node, accumulate the sample values ​​along each column dimension and divide by the total number of samples. The arithmetic mean of each dimension is calculated and combined to form a mean vector. Then, the mean vector is subtracted element-wise from each feature vector in this subset of feature vectors to achieve zero-mean normalization, thus constructing a decentralized matrix. This operation removes the spatial translation of the data clusters, aligning the centroid of the dataset to the origin of the coordinate system.

[0042] Based on this, the covariance matrix between features is calculated to capture the linear correlation structure between each dimension.

[0043] Specifically, the covariance matrix Satisfying the relation:

[0044] In the formula, Let the covariance matrix be denoted as . A symmetric positive semidefinite matrix; This represents the number of samples in the feature vector subset of the current node. ; Represents a decentralized matrix with size . ,in The dimension of the feature space; express The transpose of the matrix. This formula uses the sample size minus 1 as a normalization factor to obtain an unbiased estimate of the covariance.

[0045] For example, when the normalized feature vector subset of the current node contains Strips processed by S1 standardization When constructing a dimensional feature vector, the arithmetic mean is calculated for each column dimension. The mean vector, after being zero-meaned, is used to construct a vector of size. Decentralized real matrix Then, a formula is generated using the above relational expression. Symmetric positive semidefinite covariance matrix .

[0046] Next, the eigenvalue decomposition algorithm is used to analyze the covariance matrix. Processing can be performed, or a singular value decomposition algorithm equivalent to the eigenvalue decomposition of a symmetric positive semidefinite matrix can be used to extract the... The non-negative eigenvalues ​​and their corresponding values A set of mutually orthogonal unit eigenvectors. All eigenvalues ​​are sorted in descending order of numerical value. The largest eigenvalue is extracted as the eigenvalue corresponding to the first principal component direction, and the unit eigenvector corresponding to the largest eigenvalue is determined as the first principal component direction. This direction represents the axis with the largest data dispersion in the current eigenvector subset.

[0047] 2. Dimensional advantage measurement calculation.

[0048] This sub-step is used to quantitatively measure the dominance of the first principal component direction in the overall data distribution, providing a basis for the weight allocation between the first principal component direction and the weighted coordinate axis direction in the subsequent mixing and splitting directions.

[0049] After completing the eigenvalue decomposition of the covariance matrix, all values ​​are sequentially retrieved from the memory array. Each feature value. Dimensional advantage measure. The variance explained is proportional to the eigenvalues ​​corresponding to the first principal component direction, satisfying the following relationship:

[0050] In the formula, This represents a dimensional advantage metric, with its value range limited to... Within the range; This represents the largest eigenvalue corresponding to the direction of the first principal component; Indicates the index number of the feature value; The first term represents the result of covariance matrix decomposition. One eigenvalue; The dimension of the eigenvector; Indicates all The summation of the eigenvalues ​​represents the total degree of dispersion of the current node's dataset across all dimensions.

[0051] when A high value indicates that the subset of feature vectors exhibits a narrow linear distribution, with the first principal component playing a dominant role; when the data is spherical or multi-directionally dispersed, The value will decrease accordingly, at which point the variance of each dimension tends to be uniform, and the weighted coordinate axis direction should obtain a higher fusion weight. For example, in a three-dimensional feature space, if the feature values ​​arranged in descending order are... Values , Values , Values Then the total explained variance is Calculated Value This indicates that the direction of the first principal component explains approximately The total variance should have a dominant weight in the splitting direction. This metric generation mechanism, based on variance proportion calculation, allows for the evaluation of data distribution characteristics without introducing additional artificial hyperparameters.

[0052] 3. Determine the direction of the weighted coordinate axes.

[0053] This sub-step is used to introduce domain knowledge, that is, to amplify the influence of key feature dimensions that have important indicative significance for bearing capacity status in the splitting direction by setting a sensitivity coefficient, so as to avoid the search tree ignoring a certain key dimension because the variance of a certain key dimension is small in the current local data subset.

[0054] First, calculate the sample variance of the standardized feature vector subset on each independent feature dimension. For example, the sample variances for the three dimensions—maximum test load, displacement value, and cumulative number of acoustic emission impacts—were calculated as follows: , and At the same time, a lookup table method is used to retrieve the preset sensitivity coefficients bound to the three feature dimensions. This coefficient is calibrated based on engineering experience, and its exemplary range of values ​​can be found in... to Here they are respectively taken as , and .

[0055] The specific calibration logic is as follows: The maximum test load is the absolute core indicator for evaluating the pull-out bearing capacity of the anchor bolt, directly determining whether the anchor bolt has reached its ultimate failure state. Therefore, it is assigned a high sensitivity coefficient, which is 1.8 in this embodiment. The acoustic emission impact accumulation number can keenly capture early hidden damage characteristics such as the propagation of microcracks inside the anchor body and slippage and debonding at the mortar-soil interface. As a precursor to failure, it has high early warning value, therefore, it is assigned a second-highest sensitivity coefficient, which is 1.5 in this embodiment. The displacement parameter directly reflects the multi-stage complex deformation characteristics of the anchor bolt during the pull-out process, covering different physical evolution stages such as elastic deformation, local creep, and plastic slippage. Its data characteristics... In local high-dimensional spaces, features often exhibit large natural fluctuations and distribution variances. If the search tree split is dominated solely by the variance of pure mathematics, it is easy to mask the true mechanical damage signals. In order to balance the influence of multi-source features with different physical dimensions and scales of change on spatial partitioning and to prevent deformation features with high natural variance from masking other core bearing capacity indicators, the system configures an appropriate sensitivity coefficient for the displacement parameters. In this embodiment, it is set to 1.2. This strategy ensures that parameters with higher engineering sensitivity can occupy a reasonable dominant weight in the search tree splitting direction in the local space while retaining displacement deformation evolution information, thereby realizing the complementary advantages of multi-source detection data.

[0056] Subsequently, the product of the sample variance and the corresponding sensitivity coefficient for each feature dimension is calculated, and the feature dimension with the largest product is selected as the target label dimension. The relationship is as follows:

[0057] In the formula, Indicates the dimension of the target label; The index variable represents the feature dimension, and its value range is... ,in The total dimension of the feature space; Indicates the first The sample variance values ​​for each feature dimension; Indicates the first Preset sensitivity coefficients corresponding to each feature dimension; This indicates the operation of retrieving the index that maximizes the value of the expression within the parentheses. When the product of multiple dimensions is of equal and equal magnitude, the index of the dimension with the smallest value is selected. Substituting the example data, the product of the first dimension is calculated as follows: The second dimension is The third dimension is After comparison, extract the feature dimension label of the term with the largest product. This strategy successfully suppressed the interference of high-variance but weakly indicative displacement dimensions on spatial partitioning, ensuring that parameters with high engineering sensitivity dominate the search tree splitting direction in the local space.

[0058] Finally, construct an all-zero basis vector with the same dimension as the feature vector, and label the dimension. The corresponding component is set to This is used as the unit vector for the specified weighted coordinate axis direction. This step ensures that even if a core detection metric is not the dimension with the largest variance in the child nodes, it can still participate in the determination of the search tree splitting direction in that local space with the help of a pre-amplified sensitivity coefficient.

[0059] 4. Construction of the hybrid splitting direction vector.

[0060] This sub-step is used to merge the direction of the first principal component, which represents the overall distribution characteristics of the data, with the direction of the weighted coordinate axis, which represents the domain knowledge, into a unified hybrid splitting direction vector. This allows the search tree to take into account both the dimensionality reduction advantages of data-driven approaches and the sensitivity of key dimensions determined by engineering experience when splitting.

[0061] Before fusion, calculate the unit vector of the first principal component direction. relative to the weighted coordinate axis direction The inner product of the two vectors, if less than 0, indicates that the directions of the high-dimensional half-space indicated by the two vectors are obtusely opposite. The signs of all components are inverted to ensure directional consistency and reduce the feature cancellation effect generated during linear superposition.

[0062] After completing the orientation alignment verification, the dimensional advantage metric is used. Construct a weighted sum of the two directions, and normalize it to obtain the hybrid splitting direction vector. The following relation is satisfied:

[0063] In the formula, This represents the normalized unit vector of the mixed splitting direction; The unit vector representing the direction of the first principal component; This represents the unit vector of the weighted coordinate axis directions after direction consistency processing; This represents a dimensional advantage metric; the denominator is a vector. of The norm is used to normalize the composite vector to a unit vector. Since the orientation consistency processing has already controlled the angle between the two vectors to be within a certain range... within, and The sum of vectors cannot be zero, and the normalization operation does not carry the risk of division by zero.

[0064] For example, take the unit vector of the first principal component direction. The component is Weighted coordinate axis direction The component is The inner product is Since it is greater than zero, no inversion is needed. Substitute it into the dimensional advantage measure. Values The weight of the weighted axis is calculated as follows: The synthesized intermediate vector components are , and After calculating the Euclidean magnitude of the intermediate vector, execute... Normalization process to generate unit hybrid splitting direction vector .

[0065] 5. Nuclear density estimation and splitting point determination.

[0066] This sub-step is used to project high-dimensional feature vectors onto a one-dimensional line along the mixing splitting direction, and to find local minima of the probability density function on the one-dimensional projected data through kernel density estimation as splitting points, so that the splitting points fall at the density trough of the data distribution, thereby reducing the direct splitting of natural sample clusters.

[0067] First, traverse the current node. Each feature vector is calculated individually, along with the normalized mixed splitting direction vector. The dot product of these values ​​generates a one-dimensional projected dataset. This operation is equivalent to mapping the data in the multi-dimensional feature space onto a one-dimensional straight line according to the orthogonal projection rule. The scalar coordinates in the direction represent the position of each feature vector.

[0068] Subsequently, the continuous probability density function of this one-dimensional projected dataset is constructed using the Gaussian kernel function nonparametric estimation method. The relationship is as follows:

[0069] In the formula, Represents the probability density function; Represents the projected scalar coordinate variable; This represents the total number of feature vectors within a node; The smoothing bandwidth parameter is determined using the Silverman rule of thumb or the Scott rule of thumb, and its preferred range is typically set to 0.1 to 0.8, exemplarily taken as [missing value]. ; The sequence number represents the observation sequence number of the feature vector sample, and its value ranges from... arrive Integers; Indicates the first One-dimensional projection value of a feature vector; Represented by natural constant An exponential function with base 0; Pi is a constant. This indicates that all nodes within the current node... The kernel function values ​​are summed by accumulating the values ​​of each projected sample. This is achieved by discretely sampling within a specified range of extreme data values. The curve can be used to obtain the probability density distribution shape of the projected data in the direction of mixing and splitting.

[0070] Next, the probability density function is calculated using the discrete finite difference algorithm. The search uses the first and second derivative sequences to find local minima of the function curve that satisfy the condition that the first derivative is zero and the second derivative is greater than zero. If multiple local minima are found, the probability density values ​​at each point are compared, and the projected scalar coordinates corresponding to the local minima with the smallest density value are taken as the split point. For example, if the projected coordinates of the local minima with the smallest density value are found... Values If the search fails to find a local minimum point that meets the conditions due to factors such as unimodal distribution, then the mean of the one-dimensional projected dataset is taken as the split point.

[0071] Finally, in the original multidimensional feature space, with the selected split point as the spatial intercept, and with... Given geometric normal vectors, instantiate the analytical equation of an orthogonal splitting hyperplane in memory. Assign eigenvectors with projection values ​​less than the split point to the left child node, and eigenvectors with projection values ​​greater than or equal to the split point to the right child node.

[0072] 6. Recursion termination and leaf node generation.

[0073] Before each recursive partitioning is triggered, it is determined whether the current node meets the preset stopping conditions: whether the total number of feature vectors contained in the current node is less than the set minimum number of leaf node samples, or whether the sum of the variances of the data in each dimension is less than the set minimum floating-point tolerance. If either of the above stopping conditions is met, the recursion stops, the current node is instantiated as a leaf node, and the feature vector sample identifier index assigned to that node is saved; if not, the above non-leaf node construction process continues, and the tree construction function is recursively called for the left and right child nodes respectively.

[0074] For example, the minimum number of leaf node samples can be taken as: The data variance and tolerance can be taken as follows: Order of magnitude. When the number of feature vectors within a node is less than this value or the data variance is below this tolerance, it indicates that the samples within that node are sufficiently clustered or the data has hardly changed, and there is no need to continue splitting.

[0075] Thus, this step produces a complete near nearest neighbor search tree, whose spatial partitioning takes into account both the overall distribution trend of the data and the local sensitivity of key feature dimensions. Furthermore, the split point is located at the density trough of the projected data, reducing the risk of fragmentation where samples with similar bearing capacity are assigned to different regions.

[0076] S3: Nearest neighbor search and bearing capacity status determination.

[0077] This step involves inputting the standardized feature vector of the anchor to be analyzed into the approximate nearest neighbor search tree constructed by S2, and obtaining the results through traversal and backtracking. The system uses the nearest neighbor feature vectors and employs a weighted voting mechanism based on the inverse distance to determine the bearing capacity state, thereby improving the accuracy and reliability of the state determination.

[0078] The inputs to this step are the approximate nearest neighbor search tree produced by S2 and the feature vector of the anchor to be analyzed after S1 standardization. The output is the bearing capacity state determination result of the anchor to be analyzed.

[0079] First, the original detection data of the anchor bolt to be analyzed is standardized in the same way as the S1 training data. That is, according to the mean and standard deviation of each feature dimension determined in the S1 stage, the mean and standard deviation are subtracted respectively to obtain the feature vector of the anchor bolt to be analyzed.

[0080] Subsequently, the feature vector of the anchor to be analyzed is input starting from the root node of the search tree, and a top-down traversal search is performed. During the traversal, the feature vector of the anchor to be analyzed and the mixed splitting direction vector stored in the current traversed node are calculated. The inner product is calculated, and the projection result of the inner product is compared with the split point stored in the current node. If the projection value is less than the split point, the node proceeds to the left child node; if the projection value is greater than or equal to the split point, the node proceeds to the right child node. This process is repeated layer by layer downwards until a leaf node is reached.

[0081] During the downward traversal, a priority queue is used to record unvisited side branch nodes on the traversal path, and the distance between the splitting hyperplane corresponding to the side branch node and the feature vector of the anchor to be analyzed is saved for subsequent backtracking approximate search. When the initial leaf node is reached, the Euclidean distance between the feature vector of the anchor to be analyzed and the feature vectors of each sample in the leaf node is calculated, and the corresponding sample is added to the candidate nearest neighbor set.

[0082] Next, the node closest to the query boundary is popped from the priority queue, and the backtracking search continues. The Euclidean distance between the samples within that node and the feature vector of the anchor to be analyzed is calculated, and the candidate set is updated, until the search process satisfies either an empty priority queue or a preset early stopping limit is reached. The candidate set is then truncated to the preset integer with the smallest distance. For example, the nearest neighbor feature vectors, Can be taken as .

[0083] Finally, the retrieved The nearest neighbor feature vectors are used to perform a weighted voting process based on the inverse distance to determine the bearing capacity state. The feature vectors of the anchor bolt to be analyzed are then calculated. With the The nearest neighbor feature vectors Euclidean distance between That is, through the loop structure in all Calculate the squares of the differences on each feature component, sum them, and perform a square root operation to generate a distance scalar value. Read The bearing capacity state label carried by each of the nearest neighbor feature vectors The nearest neighbor samples belonging to the same label category are weighted and accumulated based on the inverse distance. The weight scores of each category satisfy the following relationship:

[0084] In the formula, This indicates that the anchor bolt to be analyzed belongs to the bearing capacity state category. The weighted score; Indicates the retrieved The sample index of the nearest neighbor feature vector; This indicates that the eigenvector of the anchor bolt to be analyzed is related to the first... Euclidean distance between the nearest neighbor eigenvectors; This represents a preset, tiny positive real constant, used to prevent division by zero when the distance is zero; Indicates the first The load-bearing capacity status label of the nearest neighbor feature vectors; Indicates the preset safety classification category, such as safe, critical, or compromised; This means that only those labels that are equal to or higher in the specified category are eligible for this rule. The nearest neighbor samples are weighted and accumulated. The weight scores of all categories are compared, and the category with the highest weight score is selected as the bearing capacity state of the anchor bolt to be analyzed.

[0085] This weighted voting mechanism ensures that samples closer to the feature vector of the anchor being analyzed contribute more to the classification result. Since the inverse distance of closer samples has a larger reciprocal value and higher cumulative weight, while the inverse distance of farther samples has a smaller reciprocal value and lower cumulative weight, this mechanism reduces the interference of distant abnormal neighbors near the classification boundary on the judgment result.

[0086] For example, if retrieved from the search tree Among the nearest neighbor feature vectors, those belonging to the safe category are... For samples with smaller distances, the inverse of the distance has a higher cumulative weight; those belonging to the destruction category... Individual samples that are relatively far apart have lower cumulative weights; those belonging to the borderline category The samples are centered. After comparing the total weight scores of the three categories, the safety category has the highest weight score, and therefore safety is output as the bearing capacity state of the anchor bolt to be analyzed.

[0087] Thus, this step utilizes the search tree constructed by S2 to complete the nearest neighbor search of the anchor bolt to be analyzed, and outputs the bearing capacity status determination result through distance inverse weighted voting.

[0088] A total of historical anchor pull-out and acoustic emission monitoring datasets from a certain engineering project were selected. The samples underwent ablation experiments, and the data labels included three load-bearing states: safe, critical, and destructive. The dataset was divided into training and test sets. The experimental hardware platform was configured with an eight-core processor and... Memory, nearest neighbor search parameters Unified settings Density estimation smoothing bandwidth parameter Set as To reduce the impact of random errors, each control group was run ten times under the same hardware and software environment, and the average result was taken as the experimental data.

[0089] Four experimental groups were set up for comparison. The first group served as the control group using the traditional method. Pure principal component analysis was used to determine the splitting direction with the maximum variance, the mean was used as the splitting point, and classification was completed through majority voting. The test accuracy was [missing value]. The average search time is The second group is the feature tree ablation group. Building upon the first group, it introduces a hybrid splitting direction based on dimensionality advantage and kernel density estimation of the minimum splitting point, improving test accuracy to [percentage missing]. The third group is the classifier ablation group. Based on the first group, only the classification method was replaced with a weighted voting mechanism based on the inverse of distance, improving the test accuracy to [percentage missing]. The fourth group was the complete experimental group, which comprehensively employed a hybrid splitting direction, nuclear density estimation of splitting points, and a distance-inverse weighted voting mechanism, achieving a test accuracy of [percentage missing]. The average search time is .

[0090] The complete protocol experimental group showed improved classification accuracy compared to the traditional method control group. The percentage point increase only slightly increased the time cost per inference step. It still meets the needs of real-time detection applications. Comparison of the results from different ablation groups shows that: the mixed splitting direction reduces the weakening of key mechanical feature information caused by solely relying on the principal component direction; kernel density estimation of splitting points improves the local partitioning structure of the multi-source feature space based on the density troughs of the projection data; and the distance-inverse weighted voting mechanism reduces the impact of distant abnormal neighbors near the classification boundary on the judgment results. These modules work together to improve the quality of nearest neighbor retrieval and enhance the accuracy and stability of anchor bolt bearing capacity status assessment.

[0091] Figure 2This diagram illustrates the weighted evaluation of feature dimension variance and sensitivity coefficient. The diagram shows the numerical distribution of three key dimensions: maximum test load, displacement value, and cumulative acoustic emission impacts. The first bar on the left represents the sample variance, the second bar in the middle represents the preset sensitivity coefficient, and the bar on the right represents the weighted product of the two. The bar chart shows that although the original sample variance is highest for the displacement value dimension, after introducing the preset sensitivity coefficient guided by domain knowledge, the weighted product of the maximum test load dimension ultimately takes precedence. This proves that the algorithm can effectively correct the bias caused by purely data-driven approaches, preventing highly engineering-sensitive mechanical features from being overwhelmed during the partitioning process. Observing the changing trend of the weighted product, it can be found that the sensitivity coefficient plays a crucial amplification role, making the final selected weighted coordinate axis direction more consistent with the physical essence of the soil and rock anchor pull-out bearing capacity assessment. This corresponds to the technical feature of guiding the search tree splitting direction and strengthening the weight of key engineering feature information in the specific implementation.

[0092] Figure 3 This diagram illustrates kernel density estimation and split point selection. The solid line in the background represents the continuous probability density distribution curve generated after projection onto the mixed split direction vector; the vertical dashed line with an arrow marks the location of the local minimum point determined by the algorithm, i.e., the local minimum split point. The probability density curve shows a clear bimodal distribution in the projection space, with the two peaks representing sample clusters with different bearing capacity response characteristics; the split point with the arrow precisely falls in the trough region between the two peaks. This proves that the algorithm successfully identified the natural sparse band in the feature space, achieving splitting at the weak density areas of the data distribution, rather than mechanically performing median segmentation; observing the curve shape on both sides of the split point reveals that the split surface effectively avoids the high-density clustering areas of the samples, reducing direct fragmentation of sample clusters with similar bearing capacity states. This corresponds to the physical characteristics described in the specific implementation method of using Gaussian kernel density estimation to achieve data-driven natural partitioning and improve the quality of spatial division.

[0093] Figure 4This diagram illustrates the comparison of ablation experiment results. The bars at the bottom of the graph represent the test accuracy of each experimental group, with values ​​corresponding to the percentage on the left vertical axis. The broken lines with solid dots represent the average retrieval time for each group, with values ​​corresponding to the time magnitude on the right vertical axis. The performance matrix shows that, compared to the first group (traditional method control group), the second group (feature tree ablation group) independently introduced a feature tree improvement module, and the third group (classifier ablation group) independently introduced a weighted classifier module. Both of these improvements to a single module effectively improved the test accuracy. In the fourth group (complete solution experimental group), the test accuracy reached its highest value after comprehensively applying all the technical modules. Furthermore, the average retrieval time shown in the broken line graph did not fluctuate drastically with the increase in algorithm complexity, remaining consistently within a low latency range of milliseconds. This demonstrates that while the hybrid splitting direction and weighted voting mechanism improve state recognition accuracy, they still maintain the efficient retrieval advantage of an approximate nearest neighbor search tree. Observing the difference between the ablation experimental group and the complete solution group reveals that the hybrid splitting axis and natural boundary segmentation make a decisive contribution to the reliability of recognition. This corresponds to the verification of the technical effect of improving the accuracy and real-time performance of anchor bolt pull-out bearing capacity state identification in the specific implementation method.

[0094] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for intelligent processing of data from testing the pull-out bearing capacity of soil and rock anchors, characterized in that, include: S1: Obtain the raw force-displacement-acoustic emission data of the anchor bolt detection, extract the mechanical parameters, displacement parameters, and acoustic emission parameters to construct an initial feature vector set, standardize the initial feature vector set to obtain a new feature vector set, and associate it with the bearing capacity state label; S2: Recursively partition the feature vector set to construct an approximate nearest neighbor search tree. The construction of non-leaf nodes includes: calculating the covariance matrix of the current node's feature vector subset and performing eigenvalue decomposition to obtain the first principal component direction and eigenvalue; calculating the dimensionality advantage measure proportional to the variance explanation rate corresponding to the eigenvalue; calculating the product of the variance of each feature dimension and the preset sensitivity coefficient, and taking the coordinate axis direction corresponding to the maximum value as the coordinate axis direction. Weighted coordinate axis direction; construct a weighted sum of the first principal component direction and the weighted coordinate axis direction based on the dimensionality advantage metric, and normalize it as a mixed splitting direction vector; project the feature vector subset onto the mixed splitting direction vector, perform kernel density estimation on the projected data, take the local minimum point or mean of the probability density function as the splitting point, and construct a splitting hyperplane through the splitting point; generate leaf nodes when the preset stopping condition is met; S3: input the feature vector of the anchor to be analyzed into the approximate nearest neighbor search tree to obtain K nearest neighbor feature vectors, and determine the bearing capacity state of the anchor to be analyzed by weighted voting based on the inverse distance of the bearing capacity state labels of the K nearest neighbor feature vectors.

2. The intelligent processing method for testing the pull-out bearing capacity of soil and rock anchors according to claim 1, characterized in that, The process involves acquiring raw force-displacement-acoustic emission data from anchor bolt testing, extracting mechanical parameters, displacement parameters, and acoustic emission parameters to construct an initial feature vector set, standardizing the initial feature vector set to obtain a new feature vector set, and associating it with a bearing capacity status label. This includes: extracting the maximum test load from the raw data as a mechanical parameter; extracting the displacement value recorded synchronously with the maximum test load as a displacement parameter; statistically analyzing the cumulative number of acoustic emission impacts within a preset time window as an acoustic emission parameter; concatenating the mechanical parameters, displacement parameters, and acoustic emission parameters and summarizing them to obtain the initial feature vector set; subtracting the corresponding statistical mean from the raw values ​​of the initial feature vector set and dividing by the standard deviation to obtain the new feature vector set; and reading the safety classification category identifier from a known testing database as a label field and associating it with the corresponding feature vector to complete the association of the bearing capacity status label.

3. The intelligent processing method for testing the pull-out bearing capacity of rock and soil anchors according to claim 1, characterized in that, The step of calculating the covariance matrix and performing eigenvalue decomposition of the current node's feature vector subset to obtain the first principal component direction and eigenvalues ​​includes: constructing a decentralized matrix of the feature vector subset; generating a covariance matrix based on the decentralized matrix and its transpose matrix using the sample number minus 1 as a normalization factor; performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding unit eigenvectors; and extracting the maximum eigenvalue and its corresponding unit eigenvector as the first principal component direction and eigenvalue.

4. The intelligent processing method for testing the pull-out bearing capacity of rock and soil anchors according to claim 1, characterized in that, The step of kernel density estimation of the projected data, taking the local minimum or mean of the probability density function as the split point, includes: continuously estimating the probability density function of the projected data using a Gaussian kernel function; calculating the first and second derivatives of the probability density function; taking the projected scalar coordinates corresponding to the local minimum point that satisfies the condition that the first derivative is equal to zero and the second derivative is greater than zero as the split point; if no local minimum point exists, taking the arithmetic mean of the projected data as the split point.

5. The intelligent processing method for testing the pull-out bearing capacity of rock and soil anchors according to claim 1, characterized in that, The step of inputting the feature vector of the anchor to be analyzed into an approximate nearest neighbor search tree to obtain K nearest neighbor feature vectors includes: traversing the feature vector of the anchor to be analyzed downwards from the root node of the approximate nearest neighbor search tree, calculating the projection result of the feature vector of the anchor to be analyzed and the mixed splitting direction vector of the current node, comparing the projection result with the splitting point to enter the corresponding child node; recording the side branch nodes that have not yet been visited and their corresponding distances based on a priority queue; calculating the Euclidean distance to update the candidate nearest neighbor set after reaching the leaf node; popping the node from the priority queue and backtracking the search to extract the K feature vectors with the smallest distance.

6. The intelligent processing method for testing the pull-out bearing capacity of soil and rock anchors according to claim 1, characterized in that, The method of determining the bearing capacity state of the anchor to be analyzed by weighted voting based on the reciprocal of the distance between the bearing capacity state labels of the K nearest neighbor feature vectors includes: calculating the Euclidean distance between the feature vector of the anchor to be analyzed and each nearest neighbor feature vector; weighting and summing the nearest neighbor feature vectors belonging to the same category according to the reciprocal of the Euclidean distance to obtain a weight score; and taking the category corresponding to the maximum weight score as the bearing capacity state of the anchor to be analyzed.

7. The intelligent processing method for testing the pull-out bearing capacity of soil and rock anchors according to claim 1, characterized in that, The step of calculating the product of the variance of each feature dimension and the preset sensitivity coefficient, and taking the coordinate axis direction corresponding to the maximum value as the weighted coordinate axis direction, includes: selecting the feature dimension with the largest product result as the target label dimension; constructing a zero basis vector with the same dimension as the feature vector, and setting the component corresponding to the target label dimension to 1 as the weighted coordinate axis direction.

8. The intelligent processing method for testing the pull-out bearing capacity of rock and soil anchors according to claim 7, characterized in that, The construction of the weighted sum of the first principal component direction and the weighted coordinate axis direction based on the dimensional advantage metric includes: if the inner product of the first principal component direction and the weighted coordinate axis direction is less than 0, then the sign of the weighted coordinate axis direction component is reversed; the first principal component direction and the sign-processed weighted coordinate axis direction are weighted and summed according to the dimensional advantage metric and the remaining weight obtained by subtracting the dimensional advantage metric from 1.

9. The intelligent processing method for testing the pull-out bearing capacity of soil and rock anchors according to claim 3, characterized in that, The calculation of the dimensionality advantage measure, which is proportional to the variance explained by the eigenvalue, includes: dividing the largest eigenvalue by the sum of all eigenvalues ​​of the covariance matrix to obtain the dimensionality advantage measure.

10. The intelligent processing method for testing the pull-out bearing capacity of soil and rock anchors according to claim 2, characterized in that, The process of acquiring the raw force-displacement-acoustic emission data of the anchor bolt detection includes: continuously acquiring mechanical and displacement signals at low frequency using a tension sensor and a linear variable differential transformer; acquiring acoustic emission signals using a high-frequency transient capture instrument; and generating raw data by synchronously aligning each signal based on a timestamp.

Images