Ground penetrating radar time-frequency feature selection method and system based on crack morphology similarity

Abstract

The application discloses a ground penetrating radar time-frequency feature selection method and system based on crack morphology similarity, and belongs to the technical field of road engineering pavement maintenance. The application carries out time-frequency feature analysis on the extracted electromagnetic signal; calculates the change rate of the crack and the intact part, and maps the calculated change rate of each data in the crack marking frame to the crack identification marking frame; further calculates the maximum curvature of the change rate curve, and forms a marking point; analyzes the morphology similarity of the marking point and the crack, obtains a feature comprehensive score, and determines the feature importance. The application carries out feature screening based on the similarity between the crack morphology identified by the feature and the actual crack morphology, can directly show the sensitivity of different features to the crack, screens the most sensitive features to the crack, and can greatly reduce the number of features.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of pavement maintenance of road engineering, and particularly relates to a ground penetrating radar time-frequency feature selection method and system based on crack morphology similarity. BACKGROUND

[0002] The structure type commonly used in high-grade highways is semi-rigid base asphalt pavement. The semi-rigid base has certain slab body and stability. However, the semi-rigid base also has inherent shortcomings: under the action of environmental factors such as dry and wet and freeze-thaw cycles and repeated loads, cracks are extremely easy to occur. The cracks damage the integrity and continuity of the pavement structure and to some extent weaken the structural strength. The occurrence of cracks also reduces the waterproofness of the pavement system. Most of the water entering the pavement structure through the cracks is retained between the surface layer and the base layer due to the lack of proper drainage channels. Under the repeated action of driving loads, a large amount of dynamic water pressure is generated to scour the base layer, causing the cracks to widen and leading to the edge biting of the asphalt pavement on both sides of the cracks. The repeated action of this process eventually leads to the loss of support and connection between the base layer and the surface layer, resulting in the occurrence of network cracking and other damages in the asphalt surface layer, accelerating the damage of the asphalt pavement, and thus affecting the quality and service life of the highway. Therefore, during the service of the semi-rigid base asphalt pavement, it is necessary to regularly detect cracks to comprehensively understand the development process of the cracks on the pavement.

[0003] In the traditional technology, the crack detection methods mainly include drilling sampling, falling hammer deflection and ultrasonic detection. These types of detection methods have the problems of strong destructiveness, low efficiency, great influence of environment, etc., and are difficult to meet the needs of modern road maintenance for efficient and accurate detection. In contrast, ground penetrating radar (GPR) can realize continuous non-destructive detection at different depths due to its non-contact, high efficiency and high resolution, and has been widely applied in the fields of bridges, railways, tunnels, roads, etc. Then, in the identification of road crack diseases, feature selection is a key link that affects the prediction accuracy of GPR data. In addition, the GPR data of cracks generally has the characteristics of multi-dimension and high complexity, and the data scale is large, the nonlinear interaction between features is significant, the noise interference is serious, and the sample distribution is relatively discrete. Therefore, how to select reasonable features that match the cracks from the large-scale GPR data to improve the efficiency and accuracy of GPR detection is a problem that needs to be solved at present. SUMMARY

[0004] The application discloses a ground penetrating radar time-frequency feature selection method and system based on crack morphology similarity. The similarity between the crack morphology identified by the feature and the actual crack morphology is used for screening the features, which can intuitively indicate the sensitivity of different features to cracks, screen out the features most sensitive to cracks, and greatly reduce the number of features, thereby effectively solving at least one technical problem involved in the background technology.

[0005] To achieve the above object, the technical scheme of the present application is:

[0006] A ground penetrating radar time-frequency feature selection method based on crack morphology similarity, comprising the following steps:

[0007] Step S1, obtaining the ground penetrating radar detection result of the road surface, taking the road surface area corresponding to the radar slice graph of the crack annotation box as the target area, and obtaining the radar electromagnetic signal of the target area, wherein the radar electromagnetic signal of the target area includes the radar electromagnetic signal of the crack position and the intact position;

[0008] Step S2, performing time-frequency feature analysis on the radar electromagnetic signal extracted in step S1, and extracting a plurality of time domain features and frequency domain features respectively;

[0009] Step S3, taking any one time-frequency feature as a target feature, obtaining the feature values of the target area under different channels and calculating the feature difference value and the feature change rate, and taking the ratio of the maximum value and the minimum value of the three analysis indexes under all channels as the scaling coefficient; wherein the feature difference value represents the feature difference value of the crack position and the intact position, and the feature change rate represents the ratio of the feature difference value and the feature value;

[0010] Step S4, selecting the index with the maximum scaling coefficient among the three analysis indexes as the optimal analysis index of the target feature, mapping the optimal analysis index value of the target feature to the radar slice graph, and marking the position corresponding to the maximum value of the optimal analysis index on the radar slice graph;

[0011] Step S5, calculating the horizontal distance between the marked point and the crack position and the variance of the horizontal distance, taking the weighted sum of the horizontal distance and the variance of the horizontal distance as the evaluation index, calculating the score of the target feature, and traversing all time domain features and frequency domain features to obtain the scores of all time-frequency features;

[0012] Step S6, sorting all time-frequency feature scores from high to low, selecting a number of time-frequency features with high scores as the preliminary screening features, and then statistically analyzing the preliminary screening features of a plurality of target areas, counting the number of each preliminary screening feature appearing under all target areas, and sorting the preliminary screening features from high to low according to the number of appearances, and selecting a number of preliminary screening features with high ranking as the final analysis features.

[0013] As a preferred improvement, the crack annotation box is obtained by the YOLOv8 model performing crack recognition on the radar slice graph.

[0014] As a preferred improvement, the process of obtaining the radar electromagnetic signal of the target area specifically comprises the following steps:

[0015] Step S11, obtaining the name of the radar slice graph where the crack annotation box is located;

[0016] Step S12, according to the naming rules of the radar slice map, the name determined by step S11 extracts the trace number range of the radar slice map;

[0017] Step S13, according to the trace number range determined in step S12, the radar electromagnetic signal of the target area is extracted from the ground penetrating radar detection result.

[0018] As a preferred improvement, the time domain features include mean, absolute mean, maximum, square difference, cubic difference, quartic difference, root mean square, standard deviation, peak-to-peak value, square of mean square value, Peak-to-RMS, skewness, kurtosis, Max-to-mean, RMS-to-abs, Max-to-abs, energy, power, waveform factor, peak factor, impact factor, zero-crossing rate, average amplitude difference, weighted average, dynamic range, instantaneous mean difference, short-time power change rate, amplitude non-stationarity, autocorrelation, extreme value number; the frequency domain features include spectral mean, spectral standard deviation, spectral skewness, spectral kurtosis, weighted frequency average, weighted mean square difference, weighted variance, fourth central moment, peak factor, skewness ratio, third kurtosis, fourth kurtosis, low frequency energy ratio, medium frequency energy ratio, high frequency energy ratio, spectral centroid, spectral entropy, spectral flatness, frequency variance, bandwidth, harmonic distortion rate, main frequency, spectral skewness, signal-to-noise ratio.

[0019] As a preferred improvement, the evaluation index is represented as:

[0020]

[0021] In the formula: is the comprehensive score, is the horizontal distance normalized value; is the variance normalized value of the horizontal distance; , respectively represent , the weight of.

[0022] As a preferred improvement, the preliminary screening features are the top 10 time-frequency features of each target area score; the final analysis features are the top 10 preliminary screening features in the order of the number of occurrences of all target areas.

[0023] A ground penetrating radar time-frequency feature selection system based on crack morphology similarity, comprising:

[0024] A signal acquisition module is configured to obtain a ground penetrating radar detection result of a road surface, take a road surface area corresponding to a radar slice map of a crack annotation box as a target area, and obtain a radar electromagnetic signal of the target area, wherein the radar electromagnetic signal of the target area includes radar electromagnetic signals of crack positions and intact positions.

[0025] a feature analysis module configured to perform time-frequency feature analysis on the extracted radar electromagnetic signals, and extract a plurality of time-domain features and frequency-domain features respectively;

[0026] an analysis index construction module configured to take any one of the time-frequency features as a target feature, obtain feature values of the target region under different channels and calculate a feature difference value and a feature change rate, and take a ratio of a maximum value and a minimum value of the three analysis indexes under all channels as a scaling coefficient; wherein the feature difference value represents a feature difference value between a crack position and an intact position, and the feature change rate represents a ratio of the feature difference value and the feature value;

[0027] a marking module configured to select an index with the largest scaling coefficient among the three analysis indexes as an optimal analysis index of the target feature, map an optimal analysis index value of the target feature to a radar slice graph, and mark a position corresponding to a maximum value of the optimal analysis index on the radar slice graph;

[0028] a scoring module configured to calculate a horizontal distance between a marked point and a crack position and a variance of the horizontal distance, take a weighted sum of the horizontal distance and the variance of the horizontal distance as an evaluation index, calculate a score of the target feature, and traverse all time-domain features and frequency-domain features to obtain scores of all time-frequency features;

[0029] a screening module configured to sort scores of all time-frequency features from high to low, select a plurality of time-frequency features with high scores as preliminary screening features, and then perform statistical analysis on the preliminary screening features of the plurality of target regions, count a number of occurrences of each preliminary screening feature under all target regions, sort the preliminary screening features from high to low according to the number of occurrences, and select a plurality of preliminary screening features with high rankings as final analysis features.

[0030] The present application has the following advantages:

[0031] The present application performs screening of features based on similarity between a crack morphology identified by features and an actual crack morphology, can intuitively indicate sensitivity of different features to cracks, and can screen features most sensitive to cracks, so that the number of features can be greatly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings, wherein:

[0033] Figure 1 a schematic diagram showing a radar slice;

[0034] Figure 2 a curve diagram representing eigenvalues under different channels;

[0035] Figure 3 a curve diagram representing eigenvalue differences under different channels;

[0036] Figure 4 a curve diagram representing eigenvalue change rates under different channels;

[0037] Figure 5 a mapping diagram representing eigenvalue change rates;

[0038] Figure 6 a curve diagram representing marker points;

[0039] Figure 7 a morphological similarity analysis result diagram;

[0040] Figure 8 a feature importance ranking result diagram. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0042] Please refer to Figures 1-8 The present embodiment provides a ground penetrating radar time-frequency feature selection method based on crack morphology similarity, comprising the following steps:

[0043] Step S1, obtain the ground penetrating radar detection result of the road surface, take the road surface area corresponding to the radar slice graph of the crack annotation box as the target area, and obtain the radar electromagnetic signal of the target area, wherein the radar electromagnetic signal of the target area includes the radar electromagnetic signal of the crack position and the intact position.

[0044] In the present embodiment, the ground penetrating radar is used in cooperation with the YOLOv8 model, the YOLOv8 model is used for crack identification of the radar slice graph, the crack annotation box is obtained, the position in the crack annotation box is the crack position, and the position outside the crack annotation box is the intact position. This kind of technology belongs to the prior art in the art, for example, the scheme disclosed in patent CN120808181A.

[0045] The process of obtaining the radar electromagnetic signal of the target area specifically comprises the following steps:

[0046] Step S11, obtain the name of the radar slice graph where the crack annotation box is located;

[0047] Step S12, according to the naming rules of the radar slice map, the channel number range of the radar slice map is extracted through the name determined in step S11;

[0048] Step S13, according to the channel number range determined in step S12, the radar electromagnetic signal of the target area is extracted from the ground penetrating radar detection result.

[0049] The name of the radar slice map adopts a standardized naming method, which provides position information for the subsequent accurate extraction of electromagnetic signals. The naming rules are: detection road section_file number_left boundary channel number of the radar slice map_right boundary channel number of the radar slice map_slice position (depth). As shown in Figure 1 Figure 1 A radar slice map (with crack annotation box) in a certain road crack detection is shown, and the name of the radar slice map is “1_01_15681_15980_78”, wherein “1” represents the detection road section; “01” represents the file number; “15681” represents the left boundary channel number of the radar slice map; “15980” represents the right boundary channel number of the radar slice map; “78” represents the slice position (depth).

[0050] The electromagnetic signal of the crack position is accurately extracted through the channel number range (15681_15980) in the slice map name and the fixed number of channels (300 channels of data).

[0051] Step S2, time-frequency feature analysis is carried out on the radar electromagnetic signal extracted in step S1, and a plurality of time domain features and frequency domain features are extracted.

[0052] Due to the serious noise interference of crack disease and the relatively discrete sample distribution, in order to select as many features reflecting crack disease as possible, in this embodiment, 54 time-frequency features are selected, of which 30 are time domain features as shown in Table 1, and 24 are frequency domain features as shown in Table 2.

[0053] Table 1 Time domain feature table

[0054]

[0055] Table 2 Frequency domain feature table

[0056]

[0057] Step S3, taking any one time-frequency feature as a target feature, the feature values of the target area under different channels are obtained and the feature difference value and the feature change rate are calculated, and the ratio of the maximum value and the minimum value of the three analysis indexes of the feature value, the feature difference value and the feature change rate under all channels is taken as the scaling coefficient; wherein the feature difference value represents the feature difference value of the crack position and the intact position, and the feature change rate represents the ratio of the feature difference value and the feature value.​

[0058] The feature values ​​at any location in the target area are directly acquired from the spectrum of time-frequency features. For the calculation of feature difference and feature change rate at intact locations, the feature difference is obtained by subtracting it from the feature value at the crack location and taking the absolute value. Then, the feature difference is divided by the feature value to obtain the feature change rate at the intact location. Similarly, for the calculation of feature difference and feature change rate at crack locations, the feature difference is obtained by subtracting it from the feature value at the intact location and taking the absolute value. Then, the feature difference is divided by the feature value to obtain the feature change rate at the crack location.

[0059] The feature value of the target feature is used to reflect the change of the signal at the crack location itself; the feature difference is used to reflect the difference between the crack location and the intact location; the feature change rate is used to reflect the degree of difference between the crack location and the intact location, thus reflecting the signal fluctuation at the crack location from three dimensions.

[0060] During ground-penetrating radar (GPR) detection, signals from intact locations typically remain stable within a certain range without drastic abrupt changes; however, signals from crack locations exhibit significant fluctuations. The three analytical indicators provided in this invention can reflect the fluctuations at crack locations from different perspectives, but the degree of fluctuation varies among these indicators. To determine which indicator more intuitively reflects the fluctuations at crack locations, this invention defines a scaling factor:

[0061] A large scaling factor indicates that the analysis index fluctuates violently and can more intuitively reflect the fluctuation of the crack location signal, while a small scaling factor indicates that the analysis index fluctuates gently and cannot intuitively reflect the fluctuation of the crack location signal.

[0062] Taking the mean P0 as the target feature as an example, the characteristic values, characteristic differences, and characteristic rates of change of the crack location are as follows: Figures 2-4 As shown in Table 3, the calculation results of the scaling factor are as follows:

[0063] Table 3 Scaling Factor Calculation Table

[0064]

[0065] from Figures 2-4 It can be seen that the fluctuation trends of eigenvalues, eigend differences, and eigenvalue change rates are basically consistent. As can be seen from Table 3, the eigenvalue scaling factor of the mean P0 is the smallest, while the eigenvalue change rate scaling factor is the largest. This indicates that the eigenvalue change rate can better reflect the fluctuation of the crack location signal and can be better used to distinguish between crack locations and intact locations.

[0066] Step S4, the maximum scaling coefficient of the three analysis indicators is selected as the optimal analysis indicator of the target feature, and the optimal analysis indicator value of the target feature is mapped to the radar slice graph. The position corresponding to the maximum value of the optimal analysis indicator is marked on the radar slice graph.

[0067] At the crack position, the electromagnetic signal of the ground penetrating radar will change obviously. The signal fluctuation is most violent at the position of the maximum value of the optimal analysis indicator, indicating that the position is the crack position, which is marked.

[0068] Taking the mean P0 as an example, by analyzing the change law of the mean P0 feature change rate, the influence of the difference of the road surface material on the change of the feature value can be avoided as much as possible. The feature change rates calculated by the radar electromagnetic signals of the 14 channels are mapped to the radar slice graph respectively, and the results are shown in Figure 5 Figure 5 from top to bottom are 1-14 channels. The marked results are shown in Figure 6

[0069] Step S5, the horizontal distance between the marked point and the crack position and the variance of the horizontal distance are calculated. The weighted sum of the horizontal distance and the variance of the horizontal distance is taken as an evaluation index to calculate the score of the target feature. All time domain features and frequency domain features are traversed to obtain the scores of all time-frequency features.

[0070] The marked point is the crack point identified from the target feature. The marked point curve is formed by the marked points in all channels. By calculating the horizontal distance between the marked point curve and the actual crack position (the position in the crack marking box) and the variance, the similarity between the marked point curve and the crack morphology can be analyzed. The change law of each feature at the crack can be clearly analyzed, and the sensitivity of different features to the crack can be intuitively and clearly shown. This lays a foundation for subsequent analysis of the determination of the feature.

[0071] The horizontal distance is used to measure the distance between the marked point curve and the crack curve in the shape space. The smaller the distance, the closer the marked point sequence to the crack morphology. The variance of the horizontal distance is used to describe whether the distribution of the marked point is stable, and to evaluate whether the marked point is close to the crack.

[0072] Under the 14 channels, the horizontal distance between the marked point and the crack position and the variance of the horizontal distance are recorded in Table 4, and the similarity analysis results are shown in Figure 7

[0073] Table 4 Morphological similarity result table

[0074]

[0075]

[0076] ​​​

[0077] The evaluation index is expressed as:

[0078]

[0079] In the formula, is a comprehensive score, is a horizontal distance normalized value; is a horizontal distance variance normalized value; , respectively represent , the weight of P.

[0080] This embodiment extracts the feature score results of a target region, as shown in Table 5:

[0081] Table 5 Feature score result table

[0082]

[0083]

[0084] In step S6, the scores of all time-frequency features are sorted from high to low, and a number of time-frequency features with high scores are selected as preliminary screening features. Then, statistical analysis is performed on the preliminary screening features of multiple target regions, the number of occurrences of each preliminary screening feature in all target regions is counted, and the preliminary screening features are sorted from high to low according to the number of occurrences, and a number of preliminary screening features with high ranking are selected as final analysis features.

[0085] In this embodiment, the preliminary screening features are the top 10 time-frequency features with high scores in each target region. As can be seen from Table 5, the preliminary screening features of this target region are P2, P32, P41, P50, P8, P28, P39, P45, P34, and P49. The score results of 100 target regions are obtained in the same way as the corresponding target region in Table 5, and the preliminary screening features of each target region are obtained in turn.

[0086] The number of occurrences of each preliminary screening feature in all target regions is counted, and through statistical analysis, the top 10 time-frequency features with high ranking are P21, P25, P17, P31, P26, P34, P45, P16, P33, and P3, as shown in Table 6. Through two screenings, the number of features is reduced from 54 to 10, and the feature quantity is reduced by 81%. The final analysis features can well present the characteristics of cracks and can be used for subsequent machine learning, which can greatly reduce the dimension of data and improve the processing speed of the model while ensuring accurate crack recognition. Figure 8

[0087] ​The embodiment also provides a ground penetrating radar time-frequency feature selection system based on crack morphology similarity, comprising:

[0088] A signal collection module is configured to acquire a ground penetrating radar detection result of a road surface, take a road surface region corresponding to a radar slice graph of a crack annotation frame as a target region, and acquire radar electromagnetic signals of the target region, wherein the radar electromagnetic signals of the target region include radar electromagnetic signals of crack positions and intact positions;

[0089] A feature analysis module is configured to perform time-frequency feature analysis on the extracted radar electromagnetic signals, and extract multiple time domain features and frequency domain features respectively.

[0090] An analysis index construction module is configured to take any one time-frequency feature as a target feature, acquire feature values of the target region under different channels and calculate a feature difference value and a feature change rate, and take a ratio of maximum values and minimum values of three analysis indexes, i.e., the feature values, the feature difference value and the feature change rate, under all channels as a scaling coefficient; wherein the feature difference value represents a feature difference value of the crack positions and the intact positions, and the feature change rate represents a ratio of the feature difference value and the feature value.

[0091] A marking module is configured to select an index with the largest scaling coefficient among the three analysis indexes as an optimal analysis index of the target feature, map an optimal analysis index value of the target feature to a radar slice graph, and mark a position corresponding to a maximum value of the optimal analysis index on the radar slice graph.

[0092] A scoring module is configured to calculate a horizontal distance between a marking point and a crack position and a variance of the horizontal distance, take a weighted sum of the horizontal distance and the variance of the horizontal distance as an evaluation index, calculate a score of the target feature, traverse all time domain features and frequency domain features, and obtain scores of all time-frequency features.

[0093] A screening module is configured to sort scores of all time-frequency features from high to low, select a plurality of time-frequency features with high scores as preliminary screening features, and then perform statistical analysis on the preliminary screening features of multiple target regions, count a number of appearances of each preliminary screening feature under all target regions, sort the preliminary screening features from high to low according to the number of appearances, and select a plurality of preliminary screening features with high sorting as final analysis features.

[0094] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are only illustrative but not limiting, and those skilled in the art can make many forms without departing from the purpose of the present application under the inspiration of the present application, which are all within the protection of the present application.

Claims

1. A method for selecting time-frequency features of ground-penetrating radar based on crack morphology similarity, characterized in that, Includes the following steps: Step S1: Obtain the ground-penetrating radar detection results of the road surface. Take the road surface area corresponding to the radar slice map where the crack marking box is located as the target area and obtain the radar electromagnetic signal of the target area. The radar electromagnetic signal of the target area includes the radar electromagnetic signals of the crack location and the intact location. Step S2: Perform time-frequency feature analysis on the radar electromagnetic signal extracted in step S1, and extract multiple time-domain features and frequency-domain features respectively; Step S3: Using any time-frequency feature as the target feature, obtain the feature values ​​of the target area under different channels and calculate the feature difference and feature change rate. Use the ratio of the maximum to minimum value of the three analysis indicators—feature value, feature difference, and feature change rate—under all channels as the scaling factor. Among them, the feature difference represents the feature difference between the crack location and the intact location, and the feature change rate represents the ratio of the feature difference to the feature value. Step S4: Select the indicator with the largest scaling factor among the three analysis indicators as the optimal analysis indicator of the target feature, map the optimal analysis indicator value of the target feature to the radar slice image, and mark the position corresponding to the maximum value of the optimal analysis indicator on the radar slice image. Step S5: Calculate the horizontal distance between the marker point and the crack location, as well as the variance of the horizontal distance. Use the weighted sum of the horizontal distance and the variance of the horizontal distance as the evaluation index to calculate the score of the target feature. Iterate through all time-domain features and frequency-domain features to obtain the scores of all time-frequency features. Step S6: Sort all time-frequency features from high to low scores, select the top-scoring time-frequency features as initial screening features, and then perform statistical analysis on the initial screening features of multiple target regions. Count the number of times each initial screening feature appears in all target regions, sort them from high to low according to the number of occurrences, and select the top-ranked initial screening features as the final analysis features.

2. The ground-penetrating radar time-frequency feature selection method based on crack morphology similarity according to claim 1, characterized in that, The crack annotation boxes are obtained by using the YOLOv8 model to identify cracks in the radar slice image.

3. The ground-penetrating radar time-frequency feature selection method based on crack morphology similarity according to claim 1, characterized in that, The process of acquiring radar electromagnetic signals in the target area includes the following steps: Step S11: Obtain the name of the radar slice containing the crack annotation box; Step S12: According to the naming rules of radar slices, extract the channel number range of the radar slice using the name determined in step S11. Step S13: Based on the channel number range determined in step S12, extract the radar electromagnetic signal of the target area from the ground penetrating radar detection results.

4. The ground-penetrating radar time-frequency feature selection method based on crack morphology similarity according to claim 1, characterized in that, 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, square of 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, weighted average, dynamic range, instantaneous mean difference, short-time power change rate, amplitude non-stationarity, autocorrelation, and number of extrema. The frequency-domain features include spectral mean, spectral standard deviation, spectral skewness, spectral kurtosis, weighted frequency average, weighted mean square, 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, and signal-to-noise ratio.

5. The ground-penetrating radar time-frequency feature selection method based on crack morphology similarity according to claim 1, characterized in that, The evaluation indicators are expressed as follows: In the formula: For comprehensive scoring, This is the normalized value of the horizontal distance; This represents the normalized variance of the horizontal distance. , They represent , The weight.

6. The ground-penetrating radar time-frequency feature selection method based on crack morphology similarity according to claim 1, characterized in that, The initial screening features are the top 10 time-frequency features that score highest for each target region; the final analysis features are the top 10 initial screening features that appear most frequently across all target regions.

7. A ground-penetrating radar time-frequency feature selection system based on crack morphology similarity, characterized in that, include: The signal acquisition module is used to acquire the ground-penetrating radar detection results of the road surface. Taking the road surface area corresponding to the radar slice map where the crack marking box is located as the target area, the module acquires the radar electromagnetic signal of the target area. The radar electromagnetic signal of the target area includes the radar electromagnetic signals of the crack location and the intact location. The feature analysis module is used to perform time-frequency feature analysis on the extracted radar electromagnetic signals, extracting multiple time-domain features and frequency-domain features respectively; The analysis index construction module takes any time-frequency feature as the target feature, obtains the feature values ​​of the target area under different channels, and calculates the feature difference and feature change rate. The ratio of the maximum value to the minimum value of the three analysis indices (feature value, feature difference, and feature change rate) under all channels is used as the scaling factor. Among them, the feature difference represents the feature difference between the crack location and the intact location, and the feature change rate represents the ratio of the feature difference to the feature value. The marking module is used to select the indicator with the largest scaling factor among the three analysis indicators as the optimal analysis indicator of the target feature, map the optimal analysis indicator value of the target feature to the radar slice image, and mark the position corresponding to the maximum value of the optimal analysis indicator on the radar slice image. The scoring module calculates the horizontal distance between the marker point and the crack location, as well as the variance of the horizontal distance. It uses the weighted sum of the horizontal distance and the variance of the horizontal distance as the evaluation index to calculate the score of the target feature. It iterates through all time-domain features and frequency-domain features to obtain the scores of all time-frequency features. The filtering module sorts all time-frequency features from highest to lowest score, selects the top-scoring time-frequency features as initial screening features, and then performs statistical analysis on the initial screening features of multiple target regions. It counts the number of times each initial screening feature appears in all target regions and sorts them from highest to lowest according to the number of occurrences. The top-ranked initial screening features are then selected as the final analysis features.

Images