Pavement crack grouting continuity evaluation method and system based on ground penetrating radar signals

Abstract

The application provides a pavement crack grouting continuity evaluation method and system based on ground penetrating radar signals, and belongs to the technical field of road engineering pavement maintenance. The application performs time-frequency analysis on a ground penetrating radar data set sample to extract multiple features, adopts a strategy combining random forest importance sorting, Pearson correlation coefficient redundancy elimination and exhaustive calculation to screen out an optimal key feature set; an XGBoost classification model is trained using the key feature set, and the classification results of each detection point are output; the classification results are subjected to gridding processing and HSV color mapping to generate a visual classification result graph; hierarchical clustering and RANSAC algorithms are used to spatially cluster and fit grouting line segments for the "grouting" type detection points; and the grouting continuity is evaluated using the grouting continuity rate as a quantitative evaluation index. The application realizes rapid, continuous and non-destructive evaluation of pavement crack grouting effect, and overcomes the defects of low efficiency and inability to continuously evaluate of traditional core sampling.

Description

Technical Field

[0001] This invention belongs to the field of road engineering pavement maintenance technology, specifically relating to a method and system for evaluating the continuity of pavement crack grouting based on ground-penetrating radar signals. Background Technology

[0002] The semi-rigid base asphalt pavement is a commonly used structural type for high-grade highways. While possessing certain slab-like properties and stability, semi-rigid bases also have inherent drawbacks. They are highly susceptible to cracking under environmental factors such as wet-dry cycles, freeze-thaw cycles, and repeated loading. Cracks not only reduce the driving quality of vehicles but also disrupt the integrity and continuity of the pavement structure, weakening its structural strength to some extent. The appearance of cracks also negatively impacts pavement performance and durability, reducing the pavement system's waterproofing. Water entering the pavement structure through cracks often remains trapped between the surface layer and the base layer due to a lack of proper drainage channels. Under repeated traffic loads, this generates significant dynamic water pressure, eroding the base layer, widening the cracks, and causing edge chipping of the asphalt pavement on both sides of the cracks. This repeated process ultimately leads to the base layer losing its support and bond with the surface layer, resulting in network cracking and other damage to the asphalt surface layer, accelerating pavement deterioration and affecting the highway's service quality and lifespan.

[0003] Trenchless grouting, as a cost-effective, convenient, and mechanically superior semi-rigid base course crack treatment technology, can significantly reduce construction workload and save substantial maintenance funds. It not only effectively restores the integrity of the pavement structure and improves road service levels but also minimizes the impact of maintenance work on public travel, making it of significant practical importance. However, the evaluation method for trenchless grouting primarily relies on core sampling. This involves observing the fullness of the grout and conducting core sample strength tests to assess the grout strength recovery effect. Core sampling is inefficient and cannot provide continuous evaluation of grouted sections, lacking a rapid, comprehensive, and non-destructive method for evaluating the effectiveness of grouting in treating semi-rigid base course cracks.

[0004] Therefore, it is necessary to provide a method and system for evaluating the continuity of pavement crack grouting based on ground-penetrating radar signals to solve the above problems. Summary of the Invention

[0005] This invention provides a method and system for evaluating the continuity of grouting for road cracks based on ground-penetrating radar signals. By analyzing radar time-frequency domain signals through machine learning, an evaluation index for grouting continuity rate is proposed, enabling rapid and continuous evaluation of the grouting effect of road cracks based on non-destructive testing equipment. This can effectively solve at least one of the technical problems involved in the background art.

[0006] To solve the above-mentioned technical problems, the present invention is implemented as follows: A method for evaluating the continuity of pavement crack grouting based on ground-penetrating radar signals includes the following steps: Step S1: Construct a ground-penetrating radar dataset containing three categories of labels: "grouting", "cracks", and "intact". Step S2 involves performing time-frequency analysis on the samples in the ground-penetrating radar dataset to obtain various features. The importance score of each feature is calculated based on the random forest model, and the cumulative importance index is used to initially screen all features. After the initial screening, redundant features are removed using the correlation coefficient between features. Finally, an exhaustive calculation method is used to determine the optimal combination of various features as the key feature set. Step S3: Construct a classification model, train the classification model using the key feature set, collect ground-penetrating radar data for the road surface crack grouting section to be evaluated to form a prediction dataset, extract the key feature set of the prediction dataset and send it into the trained classification model, and output the classification result. Step S4: The classification results from step S3 are processed into a grid. Based on the HSV color space segmentation technology, different color detection points are configured for each classification category to obtain a visualized classification result map. Step S5: The hierarchical clustering method is introduced to perform spatial clustering on the detection points classified as "grouting" to form multiple candidate clusters. The RANSAC algorithm is then used to fit a grouting line segment to each candidate cluster. Step S6: Use the grouting continuity rate as a quantitative evaluation index to assess the grouting continuity, wherein the grouting continuity rate represents the ratio of the grouting segment to the total length of the crack.

[0007] As a preferred improvement, in step S2, the features obtained from the time-frequency analysis include time-domain features and frequency-domain features. The time-domain features include: mean, absolute mean, maximum value, squared difference, cubic difference, quartic difference, root mean square, standard deviation, peak-to-peak value, squared mean square, peak-to-RMS, skewness, kurtosis, max-to-mean, RMS-to-abs, max-to-abs, energy, power, waveform factor, peak factor, impulse factor, zero-crossing rate, and average amplitude. Value difference, weighted average, dynamic range, instantaneous mean difference, short-time power change rate, amplitude non-stationarity, autocorrelation, number of extrema; frequency domain characteristics include: spectral mean, spectral standard deviation, spectral skewness, spectral kurtosis, weighted frequency average, weighted root mean square deviation, weighted variance, fourth-order center distance, peak factor, skewness ratio, third-order kurtosis, fourth-order kurtosis, low-frequency energy ratio, mid-frequency energy ratio, high-frequency energy ratio, spectral centroid, spectral entropy, spectral flatness, frequency variance, bandwidth, harmonic distortion rate, dominant frequency, spectral skewness, signal-to-noise ratio.

[0008] As a preferred improvement, step S2, "calculating the importance score of each feature based on the random forest model and using the cumulative importance index to initially screen all features," specifically includes the following process: The importance of a statistical feature is measured by the amount of impurity reduction it produces during the splitting of all decision tree nodes, resulting in an importance score. Sort all features in descending order of importance score; The importance scores of each feature are accumulated sequentially according to their sorting order, before defining the preceding... The cumulative importance index is the proportion of the sum of the importance scores of each feature to the sum of the importance scores of all features. A cumulative importance threshold is set, and accumulation stops when the cumulative importance index reaches the threshold, and the previous values ​​are taken. One feature is used as the initial screening feature, and other features are eliminated.

[0009] As a preferred improvement, step S2 involves eliminating redundant features using the correlation coefficient between features, specifically including: The Pearson correlation coefficient is used to measure the degree of linear correlation between two features; Set a correlation coefficient threshold. If the correlation coefficient between two features is greater than the threshold, it is determined that there is a strong linear correlation between the two features. The feature with lower importance is removed, and only the feature with higher importance is retained. Otherwise, both features are retained.

[0010] As a preferred improvement, step S2 employs an exhaustive calculation method to determine the optimal combination of multiple features, specifically including: Perform a full combination enumeration on the feature set retained after removing redundant features, traversing all possible feature subsets; For each feature combination, the K-fold cross-validation method is used to evaluate its classification performance, and the average value of the evaluation index of each fold is taken as the performance evaluation result of the feature combination. Select the feature combination with the highest evaluation index score as the key feature set.

[0011] As a preferred improvement, in step S3, the classification model is the XGBoost model.

[0012] As a preferred improvement, in step S4, the classification results of step S3 are subjected to gridding processing, specifically including: The annotation boxes in the classification results are located based on HSV color space segmentation technology; The vertical direction of the labeled box is divided equally according to the number of channels of the ground penetrating radar, and the horizontal direction is dynamically divided equally according to the channel number to form a set of grid cells. Each grid cell corresponds to the classification result of a detection point. The detection points are color-coded according to the classification category: green is marked for those predicted as "intact", red is marked for those predicted as "cracks", and blue is marked for those predicted as "grouting".

[0013] As a preferred improvement, step S5 specifically includes: Extract the set of all marker points representing "grouting" from the visualization classification result image; The set of marked points is spatially clustered using a hierarchical clustering algorithm based on the Ward link criterion, with Euclidean distance as the metric. Clustering is terminated when the minimum inter-cluster distance exceeds a preset threshold, thus forming candidate clusters. The candidate clusters are filtered for validity, and clusters with fewer than a preset threshold number of points or a vertical span less than a preset proportion of the total height of the annotation box are removed. For each candidate cluster, apply the RANSAC algorithm to fit a straight line to obtain the direction vector of the cluster, and retain the fitted straight line segments whose angle with the direction vector is within a preset range. Project all points within the retained cluster onto their respective direction vectors, take the minimum and maximum values ​​of the projection interval as the two endpoints of the grouting segment, and output all valid grouting segments.

[0014] A system for performing the above-described method for evaluating the continuity of pavement crack grouting based on ground-penetrating radar signals, comprising: The dataset construction module builds a ground-penetrating radar dataset with three categories of labels: "grouting", "cracks", and "intact". The feature selection module performs time-frequency analysis on samples in the ground penetrating radar dataset to obtain various features. It calculates the importance of each feature based on the random forest model and sorts them. It uses the correlation coefficient between features to remove redundant features and uses an exhaustive calculation method to determine the optimal combination of multiple features as the key feature set. The classification module constructs a classification model, trains the classification model using the key feature set, and after training, collects ground-penetrating radar data for the grouting section of the road surface crack to be evaluated to form a prediction dataset, extracts the key feature set of the prediction dataset and sends it into the trained classification model, and outputs the classification result. The visualization module performs gridding on the classification results and configures different color detection points for each category based on HSV color space segmentation technology to obtain a visualized classification result map. The clustering module introduces a hierarchical clustering method to spatially cluster the detection points classified as "grouting" to form multiple candidate clusters. The RANSAC algorithm is then used to fit a grouting line segment to each candidate cluster. The evaluation module uses the grouting continuity rate as a quantitative evaluation index to assess the continuity of grouting, wherein the grouting continuity rate represents the ratio of the grouting segment to the total length of the crack.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention combines ground-penetrating radar time-frequency domain signals with machine learning to achieve rapid, continuous and non-destructive evaluation of the grouting effect of road cracks, overcoming the shortcomings of traditional core drilling sampling methods, such as low efficiency, strong destructiveness and inability to conduct continuous evaluation.

[0016] (2) The present invention adopts a cascaded feature selection strategy that combines random forest importance ranking, Pearson correlation coefficient elimination and exhaustive calculation, which effectively reduces feature redundancy and improves the classification accuracy and generalization ability of the model. (3) By eliminating interlayer signals, this invention effectively avoids interference from interlayer grout signals in high-pressure grouting on the evaluation of grouting effect inside cracks, and improves the authenticity of the evaluation results. (4) This invention uses hierarchical clustering and RANSAC algorithm to perform spatial clustering and line fitting on grouting prediction results, which can effectively remove discrete noise points and accurately extract grouting line segments, providing a reliable data basis for quantitative evaluation. (5) This invention proposes a quantitative evaluation index for grouting continuity rate, realizes the quantitative characterization of crack grouting continuity, and provides an intuitive and scientific basis for road maintenance decision-making. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A diagram illustrating the annotation of the dataset; Figure 2 A graph showing the cumulative importance of features; Figure 3 A heatmap showing the correlation between features; Figure 4 A graph representing the ranking of feature importance; Figure 5 Represents a key feature combination diagram; Figure 6 This diagram shows the effect of core sample grouting. Figure 7 This represents a visual classification result chart; Figure 8 This represents another type of visual classification result chart; Figure 9 express Figure 7 The corresponding clustering result diagram of the visualization classification results; Figure 10 express Figure 8 The corresponding clustering result diagram of the visualization classification results; Figure 11 A diagram showing the grouting location markings; Figure 12 This indicates the on-site layout drawing; Figure 13 This represents a core sampling diagram. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figures 1-13 As shown, this embodiment provides a method for evaluating the continuity of pavement crack grouting based on ground-penetrating radar signals, including the following steps: Step S1: Construct a ground-penetrating radar dataset containing three categories of labels: “grouting”, “cracks”, and “intact”.

[0020] The ground-penetrating radar (GPR) dataset includes several GPR data maps, each with a resolution of 320×320 pixels. GPR transmits high-frequency electromagnetic waves into the road surface structure via a transmitting antenna. These waves are reflected when they encounter interfaces with varying dielectric constants (such as cracks or grouting materials). The receiving antenna records the reflected waveform signals. Based on the differences in waveform signals at different locations, the Labelimg annotation tool can be used to label different locations on the map, providing clear classification targets for subsequent machine learning.

[0021] Step S2 involves performing time-frequency analysis on the samples in the ground-penetrating radar dataset to obtain various features. The importance score of each feature is calculated based on the random forest model, and all features are initially screened using the cumulative importance index. After the initial screening, redundant features are removed using the correlation coefficient between features. Finally, an exhaustive calculation method is used to determine the optimal combination of various features as the key feature set.

[0022] By performing a Fourier transform on the raw time-domain signal (GPR data map) of ground penetrating radar, frequency-domain information can be obtained, forming joint time-frequency domain data. This joint time-frequency domain data contains rich road structure information and can extract multi-dimensional and highly complex features. However, due to the large data scale, significant nonlinear interactions between features, severe noise interference, and relatively discrete sample distribution, this invention adopts a systematic feature selection strategy for the joint time-frequency domain data to maximize the retention of effective information while minimizing feature redundancy, thereby improving the training efficiency and classification accuracy of subsequent models.

[0023] Because crack damage is severely affected by noise and the sample distribution is relatively discrete, in order to select the characteristics reflecting crack damage as comprehensively as possible, a total of 54 time-frequency features were selected in this embodiment, including 30 time-domain features (as shown in Table 1) and 24 frequency-domain features (as shown in Table 2).

[0024] Table 1 Temporal Characteristics Table 2 Frequency Domain Characteristics The feature selection strategy provided by this invention includes the following three stages: (1) Ranking of feature importance based on random forest model Random forest models offer advantages such as high prediction accuracy and strong resistance to overfitting, making them well-suited for constructing mappings between multidimensional features and labels. This particular random forest model employs a conventional architecture in the field, taking 54 extracted features as input and outputting the model's predicted labels. During training, the importance scores of each feature are calculated and sorted from highest to lowest.

[0025] Specifically, the importance score of a feature is calculated as follows: The importance of a decision tree node is measured by the amount of impurity reduction (MDI) generated during all node splits in the decision tree. The importance of each feature is defined as: In the formula, For the first The importance score of each feature; This represents the number of decision trees in the random forest model. For the first In the decision tree, the first A set of nodes whose features serve as splitting variables; For nodes The decrease in impurity resulting from splitting.

[0026] A higher feature importance score indicates a more significant contribution of that feature to the model's prediction results. To identify key features that characterize the model's main decision information, after obtaining the importance scores of all features, they are sorted in descending order of score, and the importance scores of each feature are accumulated sequentially according to the sorting order. (Before definition...) The cumulative importance index is the proportion of the sum of the importance scores of each feature to the sum of the importance scores of all features. A cumulative importance threshold is set, and accumulation stops when the cumulative importance index reaches (is greater than or equal to) the cumulative importance threshold, and the previous values ​​are taken. One feature is used as the initial screening feature, and other features are eliminated. The initial screening is carried out by a proposed cumulative importance index, which effectively retains the core features on which the model's main decisions depend, while focusing on reducing the number of features and improving generalization ability.

[0027] (2) Redundant features are removed based on Pearson correlation coefficient After the first stage of importance ranking, some features were identified as contributing little to the classification. Based on this, for the top-ranked feature set, redundant features carrying highly similar information need to be further identified and removed. Specifically, this stage uses the Pearson correlation coefficient to measure the degree of linear correlation between two features, and its calculation formula is as follows: In the formula, Representation of features and characteristics The Pearson correlation coefficient; , Representing features respectively and characteristics The Each sample value; , Representing features respectively and characteristics The average of all sample values.

[0028] Set correlation coefficient threshold ,like , then represents the feature and characteristics There is a strong linear correlation between them, i.e., features and characteristics Retaining features that carry the same or similar performance information not only fails to add extra information to the model but may also introduce collinearity problems, leading to overfitting. Therefore, feature removal is necessary. and characteristics For features ranked lower in importance, only the features ranked higher in importance are retained, thus simplifying the feature set; conversely, if... , then represents the feature and characteristics There is no strong linear correlation between them, i.e., the features and characteristics Both types of performance information are retained, carrying different data.

[0029] (3) Determine the optimal feature combination based on exhaustive calculation After the first two stages of screening, features with low importance and high linear redundancy were removed from the feature set. However, nonlinear interactions may still exist among the remaining features: some features have mediocre classification ability when used alone, but can significantly improve classification performance when combined with other features; or some features perform well individually, but may interfere with each other when combined. Such nonlinear interactions cannot be determined through the first two stages of screening. To capture these combination effects, this invention employs an exhaustive calculation method to enumerate all possible combinations of the currently retained feature set, traversing all possible feature subsets. For each feature combination, a unified evaluation process is used for verification: based on the K-fold cross-validation method, the dataset under this feature combination is divided into five roughly equal-sized subsets. Each time, one subset is used as the validation set, and the remaining four subsets are used as the training set to train the XGBoost model. This process is repeated five times, ensuring that each subset is used as a validation set once. The average of the evaluation metrics from the five experiments is taken as the final model performance evaluation result. Specifically, in this embodiment, the evaluation metric is the coefficient of determination. The calculation process is as follows: In the formula, Indicates the first The true results of each sample; Indicates the first The classification results of each sample; This represents the average of the true results for all test samples.

[0030] After traversing all feature subsets, the evaluation metrics of all feature subsets are compared, and the feature subset with the highest score is selected as the final key feature combination. This stage ensures that the features finally selected into the model are not only effective at the individual level, but also achieve optimal configuration at the collaborative level through global traversal search.

[0031] Through the cascaded screening of the above three stages, a set of key features with low information redundancy, high classification contribution, and strong synergistic effect is finally obtained for subsequent model training.

[0032] Figure 2 The cumulative importance curve of features in the crack grouting classification model is shown. From Figure 2It can be seen that the cumulative importance gradually increases with the number of features. When there are few features, key features contribute significantly, resulting in a steep curve; as the number of features gradually increases to a certain extent, the gain on importance from newly added features diminishes, and the curve flattens, exhibiting the "diminishing marginal returns" law of feature importance. To avoid redundant and excessive noisy features, this paper uses cumulative importance as a threshold, effectively retaining the core features upon which the model's main decisions depend while focusing on compressing the number of features and improving generalization ability. Specifically, when the cumulative importance of a feature reaches the 70% threshold, the curve enters a "dense region," the growth of cumulative importance slows significantly, and the contribution rate of newly added features to overall importance decreases substantially. This indicates that the key influence of core features has been fully captured, and the supplementary value of subsequent features tends to weaken. Therefore, 70% cumulative importance is determined as the threshold for feature selection in the model.

[0033] Twenty-one key features were identified using a 70% cumulative importance threshold. Importance and correlation analyses were then performed on these selected features. Figure 3 , Figure 4 As shown in the figure, among the top 21 features affecting crack grouting classification, P15 and P23, P3 and P8, P9 and P33, P14 and P15, and P4 and P11 show strong correlations. To reduce the impact of correlated features on the model's predictive performance, features with R > 0.95 were selected. The feature importance calculated by the random forest was obtained under the condition that all features participated in the modeling simultaneously. When highly correlated features exist, these features will express similar information, thus distributing their importance in the model. After removing redundant features based on Pearson correlation (|R| > 0.95), the collinearity effect of some features was eliminated, and their relative contributions changed. Therefore, after removing correlated features, the remaining features were re-ranked using the random forest model to more accurately reflect the contribution of each feature in crack grouting classification. The ranking results are shown in the figure. Figure 4 As shown. Through correlation analysis and feature importance ranking, 14 features with no strong linear correlation and close to the top were identified, with P22 having the highest importance and P11 having the lowest.

[0034] Exhaustive calculations are performed on the 14 feature vectors, and the combination of feature vectors with the highest prediction accuracy that affects performance is selected as the key feature vector, such as... Figure 5 As shown in the figure, among all the feature factor combinations for crack grouting classification, the classification model with the full combination of features P22, P29, P25, P12, P23, P52, P3, P2, P28, P53, P10, P21, P13, and P11 as input variables has an AUC of 91.73%, which is the best performing feature combination. The sample set of this combination is the key feature set.

[0035] Step S3: Construct a classification model and train the classification model using the key feature set. After training, for any road surface crack grouting continuity evaluation task, first collect ground-penetrating radar data of the road surface to form a prediction dataset, then extract the key feature set of the prediction dataset and send it into the trained classification model to output the classification result.

[0036] The classification model is the XGBoost model. During the training process, in order to avoid training bias caused by class imbalance, a stratified random sampling strategy is adopted to construct the key feature set. Specifically, the key feature set contains 999 samples, of which 333 samples each contain the labels "grouting", "crack" and "intact".

[0037] After training, the XGBoost model can be used to evaluate ground-penetrating radar data to automatically classify the road surface condition at each detection point. The classification results include bounding boxes, classification categories, and confidence scores.

[0038] It should be noted that during grouting construction, high-pressure grouting can easily cause the grout to diffuse along the weak surfaces between layers, forming an interlayer connectivity effect. This results in a large amount of grout accumulating between layers rather than inside the cracks. This interlayer grout will also generate significant radar reflection signals. Classification models may misjudge the interlayer grouting signals as effective grouting within the cracks, thus severely overestimating the actual grouting effect. Therefore, in the data preprocessing process of this invention, the interlayer portion of the radar electromagnetic signal is directly deleted, without performing time-frequency domain analysis or subsequent model prediction. For example, a single electromagnetic wave has 256 electromagnetic points from top to bottom; the electromagnetic data at the interlayer location (range of 50-60 points) is directly deleted.

[0039] Step S4: The classification results from step S3 are processed into a grid. Based on the HSV color space segmentation technology, different color detection points are configured for each classification category to obtain a visualized classification result map.

[0040] After the classification prediction in step S3, each valid detection point has obtained one of the three labels: "crack," "grouting," or "intact." To make the evaluation results intuitive and readable, this invention adopts a visualization processing method: based on HSV color space segmentation technology, the annotation boxes in the classification results are located, and the vertical direction of the annotation box (green box) is evenly divided according to the number of channels of the ground penetrating radar. In this embodiment, the vertical direction is evenly divided into 14 parts, corresponding to the 14 channels of the ground penetrating radar; the horizontal direction is dynamically evenly divided according to the number of channel numbers. Thus, a set of grid cells defined by the depth layer and the horizontal channel number is formed within the annotation box. The channel number range of each annotation box is calculated according to the image resolution and the number of channel numbers, and the classification result of the channel number is mapped to the corresponding position in the image. Then, each grid is color-coded according to the predicted category: green for "intact," red for "crack," and blue for "grouting." After color mapping, a visualized classification result image is generated. In the classification result image, the spatial distribution of blue detection points directly reflects the filling status of grout at different depths and horizontal positions of the crack. To highlight the grout distribution, you can choose to display only the blue detection points and set other types of detection points to the background color or make them semi-transparent.

[0041] Step S5: A hierarchical clustering method is introduced to spatially cluster the detection points classified as "grouting" to form multiple candidate clusters. The RANSAC algorithm is then used to fit a grouting line segment to each candidate cluster.

[0042] In the generated visualization image, the blue detection points may be discretely distributed. This invention further introduces a hierarchical clustering method based on the Ward link criterion to perform spatial clustering of the blue detection point set, specifically including the following process: First, the set of all blue marker points representing "grouting" in the visualization image is accurately extracted by HSV color space segmentation, and the boundary information of the annotation box is obtained as a spatial constraint. Then, a hierarchical clustering algorithm is used to spatially cluster the set of blue-marked points: initially, each point is treated as an independent cluster; subsequently, in each iteration, the Euclidean distance between all clusters is calculated, and the two closest clusters are merged into a new cluster. The inter-cluster distance is calculated using the Ward link criterion, which aims to minimize the increase in the sum of squares within the merged cluster, tending to generate compact and uniformly sized clusters. The merging process terminates when the minimum inter-cluster distance exceeds a preset spatial distance threshold. This threshold is preset based on the spatial resolution of the radar image and the typical scale of the grouting section.

[0043] Hierarchical clustering automatically generates several candidate clusters, each representing a group of spatially contiguous or neighboring grouting points. Then, the candidate clusters are filtered for effectiveness. The filtering criteria include: (1) If the number of points in a cluster is less than the preset threshold, it is considered as isolated noise and is removed; (2) Clusters whose vertical span is less than a preset proportion (e.g., 30%) of the total height of the labeling frame are considered as interference signals in non-crack directions and are removed. This filtering step eliminates most of the discrete noise.

[0044] Next, the RANSAC algorithm is applied to fit a straight line to each retained candidate cluster. RANSAC is a robust parameter estimation method with natural resistance to outliers. Specifically, it is implemented as follows: two points are randomly selected from the cluster to determine a candidate straight line; the number of inliers within the cluster whose geometric distance to the straight line is less than a preset residual threshold (e.g., 5 pixels) is counted; the random sampling and statistical process is repeated a preset number of times (e.g., 1000 times); the candidate straight line with the largest number of inliers is selected as the optimal fitted straight line for that cluster, and its direction vector is the extension direction of the grouting section.

[0045] Cracks in radar images typically exhibit a near-vertical orientation. Therefore, the angle between the direction vector and the horizontal axis is calculated, and only the fitted straight line segments that are approximately vertical (angle range 70~110 degrees) are retained, while horizontal or inclined abnormal line segments are discarded.

[0046] Finally, for all grouting clusters that passed the above verification, all their points are projected onto their corresponding direction vectors. The minimum and maximum values ​​of the projection interval are taken as the two endpoints of the grouting segment, and all valid grouting line segments within the same annotation frame are output. This method transforms discrete grouting pixels into a set of grouting line segments with definite spatial locations and extension lengths, thus comprehensively depicting multiple discontinuous segments in grouting construction and providing structured data for subsequent quantitative evaluation. Experiments show that this scheme is insensitive to outlier noise and can effectively reflect real-world conditions such as grouting discontinuities.

[0047] Step S6: Use the grouting continuity rate as a quantitative evaluation index to assess the grouting continuity, wherein the grouting continuity rate represents the ratio of the grouting segment to the total length of the crack.

[0048] The calculation process for the grouting continuity rate is expressed as follows: In the formula, (Grouting interruption rate) represents the percentage of the total crack depth that is not effectively filled with grout. A smaller value indicates better grouting continuity; a larger value indicates more severe grouting interruptions, requiring supplementary grouting. (GroutingLength) refers to the grouting length, which is the length of the grouting segment in step S5. The length of each grouting segment is determined by the vertical projection distance between its two endpoints, and the unit is consistent with the depth scale of the radar image; (CrackLengtht) is the total length of the crack, in... Figure 10 For example, the radar's scanning width is 1.7m, and it is divided into 14 channels from top to bottom. The total length of the crack is calculated based on the extension length of the crack in the radar image.

[0049] To ensure the reliability of the evaluation results, the evaluation conclusions in step S6 need to be verified by sampling. The verification uses core sampling, and the specific implementation process is as follows: ① Grouting location acquisition: Mark the location using radar post-processing software and export the latitude and longitude information of the grouting location; ② On-site layout: GPS is used to lay out and mark the grouting location of the cracks to guide the core sampling personnel to accurately locate the grouting location; ③ Core sampling: Samples are taken from the GPS-marked points using a core drilling machine; ④ Result Comparison: Compare the actual observation results of the core sample (presence or absence of slurry filling, degree of filling) with the model classification results. If the core sample shows that the location is effectively filled with slurry, the accuracy of the model prediction is verified; otherwise, further analysis of the source of model error is needed, and the classification model or feature selection strategy should be adjusted and optimized.

[0050] Through the verification steps, a closed-loop evaluation system of "non-destructive testing prediction - core sampling verification" is formed to ensure the reliability of the evaluation method in practical engineering applications.

[0051] This embodiment also provides a system for performing the above-described method for evaluating the continuity of pavement crack grouting based on ground-penetrating radar signals, comprising: The dataset construction module builds a ground-penetrating radar dataset with three categories of labels: "grouting", "cracks", and "intact". The feature selection module performs time-frequency analysis on samples in the ground penetrating radar dataset to obtain various features. It calculates the importance of each feature based on the random forest model and sorts them. It uses the correlation coefficient between features to remove redundant features and uses an exhaustive calculation method to determine the optimal combination of multiple features as the key feature set. The classification module constructs a classification model, trains the classification model using the key feature set, and after training, collects ground-penetrating radar data for the grouting section of the road surface crack to be evaluated to form a prediction dataset, extracts the key feature set of the prediction dataset and sends it into the trained classification model, and outputs the classification result. The visualization module performs gridding on the classification results and configures different color detection points for each category based on HSV color space segmentation technology to obtain a visualized classification result map. The clustering module introduces a hierarchical clustering method to spatially cluster the detection points classified as "grouting" to form multiple candidate clusters. The RANSAC algorithm is then used to fit a grouting line segment to each candidate cluster. The evaluation module uses the grouting continuity rate as a quantitative evaluation index to assess the continuity of grouting, wherein the grouting continuity rate represents the ratio of the grouting segment to the total length of the crack.

[0052] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for evaluating the continuity of pavement crack grouting based on ground-penetrating radar signals, characterized in that, Includes the following steps: Step S1: Construct a ground-penetrating radar dataset containing three categories of labels: "grouting", "cracks", and "intact". Step S2: Perform time-frequency analysis on the samples in the ground penetrating radar dataset to obtain various features, calculate the importance score of each feature based on the random forest model, and use the cumulative importance index to perform initial screening of all features; After the initial screening is completed, redundant features are removed using the correlation coefficients between features. Finally, an exhaustive calculation method is used to determine the optimal combination of multiple features, which serves as the key feature set. Step S3: Construct a classification model, train the classification model using the key feature set, collect ground-penetrating radar data for the road surface crack grouting section to be evaluated to form a prediction dataset, extract the key feature set of the prediction dataset and send it into the trained classification model, and output the classification result. Step S4: The classification results from step S3 are processed into a grid. Based on the HSV color space segmentation technology, different color detection points are configured for each classification category to obtain a visualized classification result map. Step S5: Introduce hierarchical clustering method to perform spatial clustering on the detection points classified as "grouting" to form multiple candidate clusters. Use RANSAC algorithm to fit a grouting line segment for each candidate cluster. Step S6: Use the grouting continuity rate as a quantitative evaluation index to assess the grouting continuity, wherein the grouting continuity rate represents the ratio of the grouting segment to the total length of the crack; In step S2, redundant features are removed using the correlation coefficient between features, specifically including: The Pearson correlation coefficient is used to measure the degree of linear correlation between two features; Set a correlation coefficient threshold. If the correlation coefficient between two features is greater than the threshold, it is determined that there is a strong linear correlation between the two features. The feature with lower importance is removed, and only the feature with higher importance is retained. Otherwise, both features are retained. In step S2, an exhaustive calculation method is used to determine the optimal combination among multiple features, specifically including: Perform a full combination enumeration on the feature set retained after removing redundant features, traversing all possible feature subsets; For each feature combination, the K-fold cross-validation method is used to evaluate its classification performance, and the average value of the evaluation index of each fold is taken as the performance evaluation result of the feature combination. Select the feature combination with the highest evaluation index score as the key feature set.

2. The method according to claim 1, characterized in that, In step S2, the features obtained from the time-frequency analysis include time-domain features and frequency-domain features. The time-domain features include: mean, absolute mean, maximum value, difference of squares, cubic difference, fourth difference, root mean square, standard deviation, peak-to-peak value, squared mean square, peak-to-RMS, skewness, kurtosis, max-to-mean, RMS-to-abs, max-to-abs, energy, power, waveform factor, peak factor, impulse factor, zero-crossing rate, average amplitude difference, and weighted average. Average value, dynamic range, instantaneous mean difference, short-time power change rate, amplitude non-stationarity, autocorrelation, number of extrema; frequency domain characteristics include: spectral mean, spectral standard deviation, spectral skewness, spectral kurtosis, weighted frequency average, weighted root mean square deviation, weighted variance, fourth-order center distance, peak factor, skewness ratio, third-order kurtosis, fourth-order kurtosis, low-frequency energy ratio, mid-frequency energy ratio, high-frequency energy ratio, spectral centroid, spectral entropy, spectral flatness, frequency variance, bandwidth, harmonic distortion rate, dominant frequency, spectral skewness, signal-to-noise ratio.

3. The method according to claim 1, characterized in that, In step S2, "calculating the importance score of each feature based on the random forest model and using the cumulative importance index to initially screen all features" specifically includes the following process: The importance of a statistical feature is measured by the amount of impurity reduction it produces during the splitting of all decision tree nodes, resulting in an importance score. Sort all features in descending order of importance score; The importance scores of each feature are accumulated sequentially according to their sorting order, before defining the preceding... The cumulative importance index is the proportion of the sum of the importance scores of each feature to the sum of the importance scores of all features. A cumulative importance threshold is set, and accumulation stops when the cumulative importance index reaches the threshold, and the previous values ​​are taken. One feature is used as the initial screening feature, and other features are eliminated.

4. The method according to claim 1, characterized in that, In step S3, the classification model is the XGBoost model.

5. The method according to claim 1, characterized in that, In step S4, the classification results from step S3 are processed into a grid, specifically including: The annotation boxes in the classification results are located based on HSV color space segmentation technology; The vertical direction of the labeled box is divided equally according to the number of channels of the ground penetrating radar, and the horizontal direction is dynamically divided equally according to the channel number to form a set of grid cells. Each grid cell corresponds to the classification result of a detection point. The detection points are color-coded according to the classification category: green is marked for those predicted as "intact", red is marked for those predicted as "cracks", and blue is marked for those predicted as "grouting".

6. The method according to claim 1, characterized in that, Step S5 specifically includes: Extract the set of all marker points representing "grouting" from the visualization classification result image; The set of marked points is spatially clustered using a hierarchical clustering algorithm based on the Ward link criterion, with Euclidean distance as the metric. Clustering is terminated when the minimum inter-cluster distance exceeds a preset threshold, thus forming candidate clusters. The candidate clusters are filtered for validity, and clusters with fewer than a preset threshold number of points or a vertical span less than a preset proportion of the total height of the annotation box are removed. For each candidate cluster, apply the RANSAC algorithm to fit a straight line to obtain the direction vector of the cluster, and retain the fitted straight line segments whose angle with the direction vector is within a preset range. Project all points within the retained cluster onto their respective direction vectors, take the minimum and maximum values ​​of the projection interval as the two endpoints of the grouting segment, and output all valid grouting segments.

7. A system for performing the pavement crack grouting continuity evaluation method based on ground-penetrating radar signals according to any one of claims 1-6, characterized in that, include: The dataset construction module builds a ground-penetrating radar dataset with three categories of labels: "grouting", "cracks", and "intact". The feature selection module performs time-frequency analysis on samples in the ground penetrating radar dataset to obtain various features. It calculates the importance of each feature based on the random forest model and sorts them. It uses the correlation coefficient between features to remove redundant features and uses an exhaustive calculation method to determine the optimal combination of multiple features as the key feature set. The classification module constructs a classification model, trains the classification model using the key feature set, and after training, collects ground-penetrating radar data for the grouting section of the road surface crack to be evaluated to form a prediction dataset, extracts the key feature set of the prediction dataset and sends it into the trained classification model, and outputs the classification result. The visualization module performs gridding on the classification results and configures different color detection points for each category based on HSV color space segmentation technology to obtain a visualized classification result map. The clustering module introduces a hierarchical clustering method to spatially cluster the detection points classified as "grouting" to form multiple candidate clusters. The RANSAC algorithm is then used to fit a grouting line segment to each candidate cluster. The evaluation module uses the grouting continuity rate as a quantitative evaluation index to assess the continuity of grouting, wherein the grouting continuity rate represents the ratio of the grouting segment to the total length of the crack.

Images