A method for determining pavement damage areas based on raw ground penetrating radar data

Abstract

The present invention relates to a method for determining road surface defect areas based on raw ground penetrating radar (GPR) data. The method comprises the following steps: 1. GPR data is collected on a sampled road section; 2. The GPR data is resampled to unify the sampling frequency of the GPR data; 3. Preprocessing is performed to highlight void areas and the data is exported as a matrix; 4. The content and size of the convolution kernel are set according to the shape of the Ricker wavelet of the GPR data, and the weight coefficient of the lateral data of the convolution kernel is set; 5. Convolution calculation is performed to obtain a judgment matrix; 6. A void judgment threshold is set; 7. The absolute value of the judgment matrix is ​​compared with the judgment threshold to obtain the location of the void area; 8. Voids in the road are identified using a preprocessing program and the judgment threshold. The method solves the problem that manual interpretation of GPR signals is time-consuming, labor-intensive, and highly subjective. The convolution method is used to enhance the difference between void area data and normal area data, thereby improving void identification accuracy. The method can be used to identify voids in road surfaces and improve the efficiency of road defect detection.

Description

Technical Field

[0001] The present invention belongs to the field of pavement maintenance and relates to a method for determining pavement damage areas by using convolution kernels to process raw ground penetrating radar data. Background Art

[0002] Ground Penetrating Radar (GPR) is a device designed for underground environmental detection. It primarily consists of an antenna and a main unit. The antenna includes a high-frequency transmitter and receiver. The transmitter and receiver can be triggered synchronously by a code wheel or a set trigger frequency. The transmitter turns on instantaneously, emitting a pulse wave, and then turns off. The receiver turns on and off after a set time window (generally 10-300 nanoseconds), awaiting the next trigger. During operation, the transmitter emits a high-frequency electromagnetic wave vertically downward from the ground to penetrate the underground dielectric layer during a single trigger. When the dielectric constants of the underground dielectric layers differ, the electromagnetic wave will be reflected and transmitted at the interface between the dielectric layers. Generally speaking, the greater the difference in dielectric constants between adjacent dielectrics at the interface, the stronger the reflected wave energy tends to be. Correspondingly, the lower the transmitted wave energy is, indicating that the deep signal will be attenuated. The receiver receives these reflected waves within a set time window and at a set sampling frequency. Because the reflected echoes from electromagnetic waves passing through dielectric layers at different depths reach the receiver at different times, the receiver's sample values ​​within a single trigger are arranged in a time series to form an A-scan (single-channel signal), yielding a radar signal corresponding to the subsurface structural layer beneath the trigger point. Subsequently, the A-scan sample values ​​are converted to grayscale values, and the multiple A-scan grayscale signals are sorted in parallel to form a B-scan (ground penetrating radar map). Based on the propagation speed of electromagnetic waves in the subsurface and the corresponding reception time intervals between adjacent sampling points, the depth of each dielectric boundary layer can be determined. Due to its continuous, rapid, non-destructive, and qualitative nature, GPR is widely used in geological exploration, road and tunnel lining defect detection, and building structure inspection.

[0003] Once the approximate dielectric constant distribution of the underlying structural layer is known, a radar echo image (B-scan) of the detection surface can be simulated. Based on the discrepancy between the actual detection image and the simulated result, anomalies in the underground structure can be determined. Due to the complex underground environment, B-scan images often contain a lot of noise. Even uneven materials in the same structural layer can cause noise due to differences in dielectric constant. Therefore, GPR data interpretation still relies on expert experience.

[0004] Traditional radar signal interpretation methods include machine learning algorithms based on time-frequency domain feature extraction and deep learning recognition methods based on radar B-scan radar images. For example, the support vector machine (SVM) algorithm, the most widely used in machine learning, is a binary classification model based on supervised learning. Its basic working principle is to find an optimal classification hyperplane that maximizes the marginal distance between two samples below this hyperplane. The larger the marginal distance, the farther apart the two samples are, and the better the classification structure. Since this algorithm cannot directly train models and identify radar signals, it requires manual extraction of time-frequency domain features, and recognition results are affected by the quality of the extracted features. Deep learning image recognition methods based on radar B-scan images rely on manual annotation of defect areas to obtain optimal weight parameter combinations during algorithm training. This method requires the collection of a large number of radar image sample sets containing defect features, and the defect areas in the samples must be manually annotated. Recognition results are affected by factors such as model parameters, sample size, and the quality of the annotations.

[0005] Convolution kernels are widely used in image processing. They perform filtering and feature enhancement functions. By setting convolution kernels of different forms and sizes, different features in the image can be enhanced, highlighting the contrast between such features and the background. The image feature enhancement properties of convolution kernels provide a new approach to the interpretation of radar maps.

[0006] The invention patent with application number 201910100046.3 discloses a method for extracting time-frequency domain statistical features based on ground-penetrating radar A-scan signals. The patent proposes extracting 28 time-frequency domain features of radar A-scan signals for use in the modeling of subsequent machine learning classification algorithms to identify water damage diseases in pavement, and uses the PCA component analysis method to perform sensitive feature analysis, ultimately obtaining five principal components with a total contribution rate of 95%.

[0007] The invention patent with application number 202111673027.3 discloses a radar image artificial intelligence recognition method and device. This invention patent solves the problem that the existing automatic recognition method based on convolutional neural network has the shortcomings of variable morphological characteristics of internal diseases of the same pavement structure, and it is difficult to use a single feature to characterize the disease with universality and versatility. The patent uses the recognition results of multiple judgment methods to vote and finally obtain the recognition results of the disease, that is, the classification of the disease area.

[0008] The aforementioned machine learning methods based on extracting time-frequency domain feature parameters from radar A-scan signals and deep learning methods based on radar B-scan images require the collection of large amounts of raw radar A-scan signals or radar B-scan images. While these methods can identify the type of damage within radar images, they cannot determine the size and dimensions of the damaged area, rendering these results of limited guidance for pavement maintenance. The model quality of machine learning methods based on time-frequency domain features is affected by the eigenvectors, sample size, and sample quality. The recognition results of deep learning methods based on radar B-scan images are also affected by sample size and model parameters. Currently, a standardized method for identifying damage in GPR is lacking.

[0009] The above two methods ignore the information contained in the single-channel radar A-scan signal and the correlation between adjacent A-scan signals, resulting in misjudgment of damage. In order to solve the problem of accurate identification of hollow damage areas, this patent combines the characteristics of the A-scan signal containing damage scale information and the advantage of the convolution network in deep transfer learning that can effectively obtain damage characteristics. The convolution kernel is used to process the raw data of the ground penetrating radar, highlighting the characteristics of the damage area, and designing the content and size of the convolution kernel. The lateral weight coefficient of the convolution kernel is set, and the hollow judgment threshold is selected to achieve accurate judgment of the lateral scale of the damage area. The scale information of the damage area can be used to analyze the safety of the pavement structure, which will provide a scientific basis for pavement maintenance, thereby achieving the purpose of extending the service life of the pavement. Summary of the Invention

[0010] Aiming at the problem that the interpretation of ground penetrating radar (GPR) signals in the prior art still relies on manual experience and judgment, the present invention provides a method for automatically identifying void defects using a convolution kernel and ground penetrating radar signals to solve the problem of automatic interpretation of GPR signals.

[0011] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0012] A method for determining pavement damage areas based on raw ground-penetrating radar data is proposed. This method primarily utilizes the Z-score normalization method and the convolution kernel feature enhancement method to automatically identify gaps in ground-penetrating radar signals. The method proceeds as follows:

[0013] Step 1: Radar data acquisition: Select road sections with voids as sampling sections, and use ground penetrating radar to collect GPR data on the sampling sections. During the acquisition process, mark the road sections with voids in the GPR data acquisition software to obtain the original radar signal data S[s, n], where s is the number of sampling points for each A-scan in the GPR data. Different GPR devices, antennas of different frequencies, or different sampling time windows may have different values ​​for s. n is the number of A-scan channels contained in the GPR data, which is used for subsequent processing and identification.

[0014] Step 2: GPR signal resampling processing, resampling the rows of the data set matrix S[s,n] into the matrix A[m,n], where m is the value of the unified number of rows after the data matrix is ​​resampled.

[0015] Step 3: GPR signal preprocessing, processing the GPR data A[m,n] after resampling in step 2, including static correction, Z-score standardization and background removal.

[0016] Step 4: Preprocess data sample storage. The GPR data obtained after preprocessing in step 3 is saved as a sample set X[m,n]. The sample set X is an m×n matrix, where m is the number of sampling points per A-scan after processing, and n is the number of A-scan channels contained in the GPR data.

[0017] Step 5: Convolution kernel matrix C is set. According to the number of sampling points occupied by a single Ricker wavelet of a single channel A-scan data in the original radar signal data A[m,n] and the number of A-scan channels determined by the convolution kernel at one time, the size of the convolution kernel matrix C is set. The expression of Ricker wavelet is In the formula, f0 is the main frequency, t is the time, and each row of the convolution kernel matrix C is multiplied by the distribution weight coefficient.

[0018] Step 6: GPR data convolution processing, use the convolution kernel C set in step 5 to convolve the sample X[m,n] in step 4, and obtain the convolved sample set data F[k,l]. The dimension of the convolved sample set data is calculated according to the size of the convolution kernel C and the step size during the convolution calculation. The sample data set F[k,l] is also called the judgment matrix.

[0019] Step 7: Set the critical threshold for empty areas. Divide the sample set F[k, l] in step 6 by the threshold. Take the absolute value of the data in the sample set F[k, l] and compare it with the threshold. Judge and output the recognition results of the empty areas and normal areas in the sample set F[k, l]. Save this threshold and use it for the next judgment of the empty areas of the GPR data. The loop program waits for the next data input.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The present invention adopts standardization to process GPR data, simplifies the GPR signal post-processing method, and is convenient for road surface survey engineering technicians to use.

[0022] The present invention applies a convolution kernel to GPR raw data, highlighting the characteristics of the disease. The convolution kernel used is the number of sampling points occupied by a single Ricker wavelet of single-channel radar signal A-scan data in the vertical direction, and is multiplied by a distribution weight coefficient in the horizontal direction. The horizontal width of the convolution kernel is set according to the gap width determined at the same time as needed, thereby realizing automatic processing of raw radar data without manual intervention, simplifying the radar data processing flow, and improving radar data processing efficiency.

[0023] The present invention considers the information of the single-channel radar A-scan signal itself and the mutual information of adjacent A-scans to solve the abnormal situation of isolated defect points in existing machine learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of air-out identification and positioning.

[0025] Figure 2 It is a schematic diagram of the longitudinal data of the convolution kernel.

[0026] Figure 3 It is a schematic diagram of the lateral distribution weight coefficient of the convolution kernel.

[0027] Figure 4 This is a complete diagram of the convolution kernel.

[0028] Figure 5 This is the original GPR map display.

[0029] Figure 6 This is a display diagram of some disease data of the original GPR data.

[0030] Figure 7 This is the GPR map display after preprocessing.

[0031] Figure 8 This is a display diagram of some GPR disease data after preprocessing.

[0032] Figure 9 This is a display diagram of the indoor hollow model.

[0033] Figure 10 This is a comparison of the recognition results of the ANN algorithm in the indoor model.

[0034] Figure 11 This is the recognition result of the indoor model by the patented method.

[0035] Figure 12 It is the recognition result obtained by ANN machine learning method based on time-frequency feature extraction of A-scan.

[0036] The meanings of the numbers in the figure are: 1 - manually set threshold for determining gaps; 2 - two gaps are identified as one gap; 3 - normal areas are mistakenly identified as gaps. DETAILED DESCRIPTION

[0037] The present invention will be described in further detail below with reference to the accompanying drawings.

[0038] like Figure 1 As shown, the complete steps of the method of the present invention are as follows:

[0039] Step 1: Radar data collection: Select the indoor model with void damage ( Figure 9 ) Use GPR for acquisition. Specifically, a GPR ground-coupled antenna is used to perform non-destructive testing on the ground. The antenna frequency is ≥500 MHz and the sampling interval is ≤50 mm. The original radar signal data S[s, n] is obtained. s is the number of sampling points for each A-scan in the GPR data. Different GPR devices, different frequency antennas, or different sampling time windows may have different values ​​of s. n is the number of A-scan channels contained in the GPR data, which is used for subsequent processing and identification.

[0040] Step 2: GPR signal resampling processing. When different GPR devices are used in step 1, due to the differences in antenna frequency, sampling frequency and sampling time window, the A-scan in the collected signal has different numbers of sampling points in the same sampling time window. Therefore, the original GPR data collected in step 1 is resampled so that all GPR data have a unified sampling frequency and contain the same number of sampling points in the same sampling time window. The number of rows of the data set matrix S[s,n] is resampled to the matrix A[m,n], where m is the value of the unified number of rows in the data matrix after resampling.

[0041] Step 3: Raw data preprocessing: GPR data is processed, including static correction, Z-score normalization and background removal.

[0042] Static correction: In the A[m, n] signal, the ground position is mapped to the first row of data in A[m, n]. If the ground is at the jth position in each data channel, k∈[1, m], then rows 1 through 10 of matrix A are deleted to obtain a new matrix A, i.e., A[1:j,:]=0. Subsequent sampling points corresponding to defects can be more accurately mapped to their actual locations. Because the ground direct wave peak amplitude is the largest in the raw GPR data, this embodiment preferably removes all data before sampling point j corresponding to the maximum absolute value of each A-scan sampling value, forming a new matrix Z[w, n], where w=mj.

[0043] Z-score normalization: Normalizing the data in each column of the matrix Z can eliminate the effect of the static operating point deviating from zero during normal operation of the ground penetrating radar electronic components. For GPR data from different manufacturers and antennas, the processed data range is scaled to (-10, 10), which helps to unify subsequent identification methods. Z-score normalization ultimately obtains the processed matrix B[w, n], as shown in Equation (1);

[0044] (1)

[0046] Where, X i is the A-scan data of channel i before normalization, X i = Z[:, i], μ is the mean of the sample matrix Z, σ is the standard deviation of the sample matrix Z, Y i is the data of lane i after normalization.

[0047] Background removal: Subtract the average value of the data in each row of matrix B from the data in that row to improve the characteristics of the defective area in the radar data, remove the background interference of the radar signal, suppress the direct wave and other clutter, and highlight the target signal. The formula is shown in formula (2):

[0048] X[i,:]=B[i,:]-mean(B[i,:]) (2)

[0050] Where X is the matrix after background removal, and its dimension is [w, n].

[0051] Step 4: The GPR data obtained after the resampled GPR data preprocessing in step 3 is used as the sample set X[w,n]. The sample set X is a w×n matrix, w is the number of sampling points of each A-scan after processing, and n is the number of A-scan channels contained in the GPR data.

[0052] Step 5: Set the convolution kernel C, where C is a u×v matrix, and u is the number of sampling points occupied by a single Ricker wavelet in the current A-scan data. Each column of A-scan data is sampled and filled with equal spacing between the main peak and side lobes of the Ricker wavelet (the Ricker wavelet expression is f0 is the antenna transmission frequency, t is the time) Where u is the number of sampling points for setting the Ricker wavelet, and v is the number of A-scan channels determined by the convolution kernel at one time.

[0053] Set the vertical data of the convolution kernel, and set the column vectors of the convolution kernel to the main peak and side lobe area of ​​the Ricker wavelet (t∈[-i, i], R(t)<0.05R(0))t∈[-i,i], for R(t)<0.05R(0) is the equally spaced sampling data. If a Ricker wavelet has 3 sampling points and contains 3 A-scans, that is, u=3, v=3, then the data in the convolution kernel is

[0054] Set the weight of each data in the convolution kernel row direction, and multiply each row of the convolution kernel matrix C by the distribution weight coefficient. In this embodiment, the Gaussian distribution weight coefficient is taken as an example. The equation of the Gaussian distribution weight coefficient is σ is the standard deviation, and the data in the convolution kernel C is In this embodiment, C is a 45×20 matrix. Figure 2 This is a schematic diagram of the longitudinal data of the convolution kernel in this example. Figure 3 This is a diagram of the horizontal distribution weight coefficient data of the convolution kernel in this example. Figure 4 This is a complete diagram of the convolution kernel in this example.

[0055] Step 6: Convolve the resampled data with the convolution kernel with a step size of p. Output the convolution result of each cycle to the discriminant matrix F[k, l] to obtain a matrix F of [k, l], where k×l=((wu) / p+1)×((nv) / p+1). In this embodiment, the iterative convolution calculation step size is preferably l.

[0056] Step 6: Calibrate on the standard data and determine the threshold. Manually select the empty area and set the empty critical threshold based on the convolution result. In this example, the threshold is set to 10. Figure 10To manually select the empty area and set the threshold map, the maximum absolute value of each column of the discriminant matrix F[k, l] is compared with the threshold. If it is less than or equal to the threshold, the A-scan corresponding to that column is set to normal and the A-scan data in that column is given a judgment label of 0 (normal channel). If it is greater than the threshold, the A-scan corresponding to that column is set to empty and the A-scan data in that column is given a judgment label of 1 (empty channel). The row index corresponding to the maximum value of the A-scan data with the label of 1 (empty channel) is returned and converted into a depth position.

[0057] Step 8: Draw the recognition results. Use the original GPR map as the drawing background, and draw the judgment results and converted depth output in step 6 in the form of lines on the GPR map to obtain the final recognition results.

[0058] Step 9: The program returns to step 1 and waits for the input of GPR data.

[0059] Figure 5 is the original GPR spectrum A[m,n] collected, Figure 6 The partial sampling values ​​corresponding to the spectrum matrix, antenna frequency 800MHz, sampling frequency 12495.497MHz, sampling spacing 10mm / channel, obtained after the raw data preprocessing process in step 3 Figure 7 The GPR map X[m,n] and Figure 8 Comparison of some sampling values ​​corresponding to the spectrum matrix Figure 6 and 8 The numerical value in can be obtained. After standardization, the data measurement of the spectrum matrix is ​​10 3 Reduced to 10 0 This helps unify radar data collected from different antennas and powers. After processing, the sample set X[386, 1160] is obtained. A single radar wavelength in this B-scan spectrum occupies approximately 20 sampling points. The program is set to simultaneously analyze 20 A-scan data channels, so the convolution kernel C[20, 20] is set. Convolution kernel C is applied to sample set X with a step size of P = 1, resulting in a discriminant matrix F[367, 1141]. The maximum absolute value and index of each column in the discriminant matrix F are taken. If the maximum value is greater than a threshold of 10, it is determined that a void defect exists, and the A-scan data channel is assigned a label of 1 (void defect).

[0060] Figure 11The output results of the discrimination program in the ground penetrating radar map of the indoor model are plotted in the figure. In the figure, the blue box represents the void and the green line represents the corresponding depth. The highlighted part in the figure is the known void area, and the judgment method of the present invention judges this as 1. The gray background area is known to be a normal area, and the judgment method of the present invention judges this as 0. The wireframe area and the highlighted area basically coincide with each other, and the depth information of the diseased area is given, which shows that the feasibility and accuracy of the recognition method of the present invention are guaranteed.

[0061] At the same time, in order to verify the recognition effect of the method of the present invention, the ANN algorithm in machine learning is used for comparison. Through the collection of data sets, preprocessing of radar data, extraction of time-frequency characteristics of radar data and ANN algorithm recognition model training, an ANN void recognition model is established. The data of the indoor void model collected by the GPR equipment is input into the ANN recognition model, and the following is obtained: Figure 12 The ANN recognition results shown in Figure 11 The method of the present invention can effectively identify the voids in the indoor model and give the size of each void in the model. Figure 12 In the ANN machine learning method based on the time-frequency feature extraction of A-scan, although the void area is identified, it fails to effectively identify each void separately. Figure 12 In the case of number 2, the result of identifying two gaps as the same gap is also as follows Figure 12 In the case of number 3, the normal area is mistakenly identified as a void, and the identification size of each void is not given. Figure 11 and Figure 12 The comparison shows that the feasibility and accuracy of the identification method of the present invention are guaranteed.

[0062] The above content is only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for determining road surface damage areas based on raw ground penetrating radar data, characterized in that: The following steps are involved: Step 1: Radar data collection: Select a road section with voids as a sampling section, use ground penetrating radar to collect GPR data on the sampling section, and obtain a data set S. During the collection process, mark the road section with voids in the GPR data collection software; Step 2: GPR signal resampling processing, resampling the rows of the data set S into the data set matrix A; Step 3, GPR data A preprocessing: GPR data A is processed, including static correction, Z-score normalization and background removal; Step 4: Export the GPR data pre-processed in step 3 into a matrix X. The matrix elements are radar waveform data. The matrix dimension is m×n, where m is the number of sampling points of each A-scan waveform after GPR data processing, and n is the number of A-scan samples contained in the data collected by GPR. Step 5: Set the convolution kernel C: The convolution kernel C is a u×v matrix, where u is the number of sampling points occupied by a single Ricker wavelet in an A-scan data in the GPR data matrix in step 4, and v is the number of A-scan channels determined by the convolution kernel at one time. The column vectors of the convolution kernel C are set to the equally spaced sampling data of the main peak and sidelobe areas of the Ricker wavelet. The Ricker wavelet expression is: f0 is the antenna transmission frequency, t is the time, t∈[-q,q], If the number of sampling points of a Ricker wavelet is 3, and the convolution kernel convolves 3 A-scans each time, that is, u=3, v=3, then the data in the convolution kernel is Multiply each row of the convolution kernel by the distribution weight coefficient G(x), and the data in the convolution kernel C is Step 6, convolution calculation: perform convolution operation on the GPR data matrix X obtained in step 4 and the convolution kernel C designed in step 5, with a step size of p, and output the convolution result of each cycle to the discriminant matrix F; Step 7: Set the critical threshold for missing values: Analyze and compare the global data distribution of the discriminant matrix F obtained in step 6, and set the critical threshold for missing values; Step 8, GPR air void data discrimination: Compare the absolute maximum value of each column of the discrimination matrix F obtained in step 6 with the threshold set in step 7. If it is less than or equal to the threshold, the A-scan corresponding to the column is set to normal, and the A-scan of the corresponding column is given a judgment result of 0. If it is greater than the threshold, the A-scan corresponding to the column is set to air void, and the A-scan of the corresponding column is given a judgment result of 1. The index corresponding to the maximum value of the A-scan of the column in the discrimination matrix F is returned and converted into depth information. When the value in the discrimination matrix corresponding to the index is positive, the A-scan data of the column is marked as air void, and when it is negative, it is marked as a water-containing area.

2. The method for determining road surface damage areas based on ground penetrating radar raw data according to claim 1, characterized in that: In step 1, a ground-coupled antenna radar is used to perform non-destructive testing of the road surface. The antenna frequency of the ground-penetrating radar is ≥500 MHz, and the sampling interval is ≤50 mm.

3. The method for determining road surface damage areas based on ground penetrating radar raw data according to claim 1, characterized in that: In step 2, the GPR signals collected by different devices and different antenna frequencies are resampled, and the data matrix obtained after resampling has a uniform number of rows.

4. The method for determining road surface damage areas based on ground penetrating radar raw data according to claim 1, characterized in that: In step 3, the radar data is subjected to static correction and Z-score normalization: Static correction maps the ground position to the first line of B-scan data, so that the sampling points corresponding to subsequent diseases can be more accurately mapped to the actual location; Z-score normalization normalizes the data of each column of B-scan to eliminate the influence of the static working point deviation from zero point when the ground penetrating radar electronic components are working normally. In addition, for GPR data of different manufacturers and antennas, the processed data range is scaled to (-10, 10), which helps to unify the subsequent identification methods. The formula of Z-score normalization is shown in formula (1): Where i is the i-th sampling point of A-scan, X i is the sampling value of each single channel A-scan before standardization, μ is the sample mean, σ is the sample standard deviation, Y i is the standardized data; Background removal: Using formula (2), the average value of each row of B-scan data is subtracted from the data in that row. This is used to improve the characteristics of the defective area in the radar data, remove the background interference of the radar signal, suppress the direct wave and other clutter, and highlight the target signal: X[i,:]=B[i,:]-mean(B[i,:]) (2) Where B is the GPR data matrix before background removal, X is the GPR data matrix after background removal, and mean is the function for finding the average value of the matrix.

5. The method for determining road surface damage areas based on ground penetrating radar raw data according to claim 1, characterized in that: In step 6, the dimension of the discriminant matrix F is Where m is the number of sampling points of each A-scan waveform after GPR data processing, n is the number of A-scan samples contained in the data collected by GPR, u is the number of sampling points occupied by a single Ricker wavelet in an A-scan data in the GPR data matrix in step 4, v is the number of A-scans that need to be determined in one convolution operation of the convolution kernel, and p is the step size of the convolution calculation.

6. The method for determining road surface damage areas based on ground penetrating radar raw data according to claim 1, characterized in that: In step seven, based on the absolute value data matrix of the data in the judgment matrix F obtained by convolution calculation in step six, the empty area is manually selected, and the empty critical threshold is set according to the convolution result.

7. The method for determining road surface damage areas based on ground penetrating radar raw data according to claim 1, characterized in that: In step eight, the A-scan data in each column of the judgment matrix is ​​compared with the threshold value, and the results 0 and 1 obtained are plotted according to the coordinates of the original sampling points of the radar map, with the radar map as the background. 1 represents the area with degassing disease, and 0 represents the normal area. Finally, it is represented as several wireframes on the radar map. The more the wireframe area overlaps with the degassing disease area, the higher the feasibility and accuracy of the threshold setting.

8. The method for determining road surface damage areas based on ground penetrating radar raw data according to claim 4, characterized in that: In step 5, the distribution weight coefficient for multiplying the rows of data of the convolution kernel C is a Gaussian weight coefficient, for example, but may also be other distribution coefficients that can change the distribution of the rows of data of the convolution kernel.

Images