Three-dimensional ground penetrating radar pavement damage multi-dimensional diagnosis method based on deep learning

Abstract

The invention discloses a deep learning-based three-dimensional ground penetrating radar pavement damage multi-dimensional diagnosis method, which comprises the following steps: collecting and sorting three-dimensional ground penetrating radar road detection data, the data comprising N B-SCAN maps, M C-SCAN maps, road positions and point coordinates, and preprocessing the B-SCAN maps and the C-SCAN maps; based on a YOLO v12 model, obtaining a first identification result for the preprocessed B-SCAN atlas; obtaining a second identification result based on the preprocessed C-SCAN map; the first recognition result and the second recognition result are further judged through the optimized multi-dimensional cross recognition rule to determine the final recognition result, in the diagnosis method, the recognition model is optimized, and the recall rate and the accuracy rate of the mode in graph recognition are improved to the maximum extent; and the false abnormal signals are further screened, so that the model is improved, the recognition precision is improved, the multi-dimensional cross recognition rule is optimized, and the omission ratio and the accuracy of abnormal defects are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to a three-dimensional ground penetrating radar pavement damage multi-dimensional diagnosis method based on deep learning. BACKGROUND

[0002] The cavities in the soil under the road caused by underground pipeline leakage and other factors may cause road collapse and other problems, affecting the normal passage of pedestrians and vehicles, and threatening the safety of the city. Ground penetrating radar is the most widely used road non-destructive testing technology at present. It uses high-frequency electromagnetic data to identify the distribution of soil media and other structures to accurately assess the soil condition with minimal damage. The radar transmits signals to the underground through multiple transmitters, and the signals reflected from the damaged area are different from those reflected from the normal soil, so as to identify the underground defects of the road. The three-dimensional ground penetrating radar collects B-SCAN maps by scanning the region of interest, and forms C-SCAN maps by imaging the data points at the same depth in multiple B-SCAN maps.

[0003] At present, the data collected by the ground penetrating radar is converted into reflected waveform images by special software such as GPR Examiner and MALA Object Imageper, and visual analysis is performed by experienced engineers. However, the inspector needs a lot of experience to manually interpret the GPR image, because these data usually have considerable environmental complexity and measurement noise. Visual interpretation of three-dimensional ground penetrating radar data is time-consuming and laborious. SUMMARY

[0004] In order to solve the above-mentioned problems existing in the prior art, the present application provides a three-dimensional ground penetrating radar pavement damage multi-dimensional diagnosis method based on deep learning. The recognition model is optimized in the diagnosis method, which maximizes the recall rate and accuracy of the model in the spectrum recognition. Further, the pseudo-abnormal signal is screened and removed, the recognition accuracy of the model is improved, and the multi-dimensional cross-recognition rule is optimized to reduce the missed detection rate and accuracy of abnormal defects.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a three-dimensional ground penetrating radar pavement damage multi-dimensional diagnosis method based on deep learning, comprising

[0006] Collect and organize three-dimensional ground penetrating radar road detection data, including N B-SCAN maps, M C-SCAN maps, road position and point coordinates, wherein the B-SCAN maps and C-SCAN maps are preprocessed;

[0007] Obtain a first recognition result based on the YOLO v12 model for the preprocessed B-SCAN map;

[0008] Obtain a second recognition result based on the preprocessed C-SCAN map;

[0009] The first recognition result and the second recognition result are further determined by the optimized multi-dimensional cross-identification rule to obtain a final recognition result.

[0010] Further, the process of obtaining the first recognition result comprises

[0011] The YOLO v12 model is constructed and trained;

[0012] The preprocessed B-SCAN graph is identified by using the YOLO v12 model to obtain a preliminary recognition result, and the preliminary recognition result comprises normal signals, abnormal signals or well signals;

[0013] The pseudo-abnormal signals are screened out from the abnormal signals by using the signal pattern recognition method, and the pseudo-abnormal signals are corrected to normal signals;

[0014] The positions and quantities of the abnormal signals or well signals are determined by using the spatial positioning and stereoscopic recognition method of the B-SCAN graph, and the correctness of the abnormal signals or well signals is verified by using the positions and quantities to obtain a final first recognition result.

[0015] Further, the signal pattern recognition method comprises

[0016] The highlighted abnormal area in the B-SCAN graph is identified, the highlighted abnormal area is extracted to a new blank canvas, the stripe pattern of the highlighted area is analyzed, and when the top radiation coefficient is a positive number, the pseudo-abnormal signal is determined and corrected to a normal signal. Further, the process of obtaining the second recognition result comprises

[0017] According to the preprocessed C-SCAN graph, the area with white color in the graph is an abnormal soil density area, and is marked and drawn to a new canvas;

[0018] The spatial positioning and stereoscopic recognition method of the C-SCAN graph is used to confirm the positions of suspected abnormalities or suspected wells, and specifically, M C-SCAN graphs are divided into BxH regions, the positions of abnormal regions are recorded, and the depth sizes are counted;

[0019] In the M C-SCAN graphs of a single point, if more than three C-SCAN graphs but not more than M graphs appear continuously, it is determined as a suspected abnormality; if more than 12 graphs appear continuously, it is determined as a suspected well;

[0020] If none of the above conditions occurs, it is determined that there is no soil abnormal signal.

[0021] Further, the multi-dimensional cross-stereoscopic recognition rule comprises

[0022] According to the second identification result, if the second identification result of the point is no anomaly, the final identification result is no anomaly;

[0023] If the second identification result of the point is suspected anomaly or suspected water well, further according to the first identification result, if the first identification result of the point is not determined as anomaly or water well, and no anomaly or water well is detected in the N B-SCAN maps, the final identification result is normal;

[0024] If the first identification result of the point is not determined as anomaly or water well, but anomaly or water well is detected in the N B-SCAN maps, and the signal interval exceeds one map, the final identification result is anomaly or water well;

[0025] If the first identification result of the point is determined as anomaly or water well, the final identification result is anomaly or water well. Further, the spatial positioning and stereoscopic identification method of the B-SCAN map comprises

[0026] Divide the N B-SCAN maps into K*V regions, record the positions of the anomaly signals and water well regions, and determine the positions and quantities of the anomalies or water wells;

[0027] In the N B-SCAN maps of a single point, if anomaly signals or water well signals appear on more than three B-SCAN maps, and the signal interval is further determined as anomaly or water well signal; if the signal interval is not more than one map, it is determined as normal signal.

[0028] Further, the process of constructing and training the YOLO v12 model comprises

[0029] Collect a large number of B-SCAN maps, according to the actual exploration situation, use anchor box to label anomaly signals and water well regions, and generate a VOC format data set,

[0030] Construct the YOLO v12 model by analyzing the imaging characteristics of the anomaly signals and water wells in the B-SCAN maps, wherein the YOLOv12 model comprises a backbone network, a neck network and a detection head;

[0031] Train the YOLO v12 model using the VOC format data set;

[0032] Input the B-SCAN map into the trained YOLO v12 model to obtain the identification result.

[0033] Further, the backbone network comprises five convolution modules, the convolution modules are used to extract image features, C3k2 is arranged between the second and third convolution modules and between the third and fourth convolution modules, and the C3k2 is used to assist in extracting features, and A2C2f is arranged in the fourth and fifth convolution blocks, the A2C2f is a unique feature, and comprises two ordinary convolution modules and an ABlock module of a local attention mechanism.

[0034] Further, the neck network comprises two groups of "up-sampling modules, A2C2f and splicing layers" to form a middle column, one group of "convolution modules, splicing layers and A2C2f" and one group of convolution modules, splicing layers and C2k2" are cascaded to form a right column, the middle column receives feature parameters in the second C2k3 and the first A2C2f in the backbone network through the first splicing layer and the second splicing layer respectively, and is used for fusing shallow layer features; the first splicing layer of the right column receives shallow layer feature information of the first A2C2f in the middle column, and the second splicing layer receives deep feature information of the last A2C2f in the backbone network.

[0035] Further, the preprocessing of the B-SCAN atlas is to denoise a large amount of background signals in the original B-SCAN atlas, and the soil density abnormal area of the C-SCAN atlas is enhanced in color contrast.

[0036] Advantages:

[0037] 1. The Area Attention attention mechanism used in the prior art model maintains a large receptive field in a simple way, greatly reduces the calculation complexity, and improves the model running efficiency, but in actual ground penetrating radar detection, based on the analysis of defects, it is found that the recognition of defects does not need to pay too much attention to the far end area, therefore, the model is improved in the application, the Window Attention attention mechanism is adopted, the model pays more attention to the difference between the defect itself and the surrounding, without a large receptive field, and the recall rate and accuracy of the model in recognizing the abnormality of the B-SCAN atlas are maximized.

[0038] 2. Based on the existing model in identifying the abnormality of the B-SCAN atlas and the water well signal, there is a false abnormal signal recognition into an abnormal signal, which leads to false recognition, therefore, the B-SCAN abnormal signal pattern recognition algorithm is established in the application, and the false abnormal signal in the abnormal signal is further discriminated, so as to improve the precision.

[0039] 3, The spatial positioning method is also developed, which further determines the position and number of anomalies or water wells in the B-SCAN map and the C-SCAN map, and then combines the optimized multi-dimensional cross stereo recognition rule to determine the final recognition result, wherein the multi-dimensional cross stereo recognition rule is optimized based on the existing recognition rule, which avoids the existing recognition rule from causing a large number of false recognitions and requiring a large amount of review work in the case that a large number of B-SCANs are not abnormal and C-SCANs judge that there are suspected anomalies. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is the specific implementation roadmap of the present application.

[0041] Figure 2 is a three-dimensional ground penetrating radar B-SCAN map and C-SCAN map disassembly method.

[0042] Figure 3 is the image after preprocessing of the B-SCAN map.

[0043] Figure 4 is the enhanced image of the C-SCAN map.

[0044] Figure 5 is the structure diagram of the YOLO v12 target detection network.

[0045] Figure 6 is the imaging feature of the abnormal area.

[0046] Figure 7 is an example of the abnormal area and the Window Attention mechanism.

[0047] Figure 8 is the extraction process of the highlighted abnormal area in the B-SCAN map.

[0048] Figure 9 is an example of the B-SCAN map area division method.

[0049] Figure 10 is the logic diagram of the B-SCAN map interpretation.

[0050] Figure 11 is the soil density abnormal area extraction process in the C-SCAN map in the implementation case.

[0051] Figure 12 is an example of the C-SCAN map area division method in the implementation case.

[0052] Figure 13 is the logic diagram of the C-SCAN map interpretation.

[0053] Figure 14To optimize the logical schematic diagram of multi-dimensional cross stereo recognition rules. DETAILED DESCRIPTION

[0054] To further illustrate the technical means and effects taken by the present application to achieve the predetermined inventive purpose, the specific embodiments, structures, features and effects according to the present application are described in detail below in combination with the drawings and preferred embodiments.

[0055] Referring to FIG. Figure 1 The present application provides a three-dimensional ground penetrating radar pavement damage multi-dimensional diagnosis method based on deep learning, which comprises the following steps:

[0056] Collect and organize three-dimensional ground penetrating radar road detection data, which includes N B-SCAN maps, M C-SCAN maps, road positions and point coordinates, wherein the B-SCAN maps and C-SCAN maps are preprocessed;

[0057] Based on the YOLO v12 model, the first recognition result is obtained from the preprocessed B-SCAN map;

[0058] Based on the preprocessed C-SCAN map, the second recognition result is obtained;

[0059] The first recognition result and the second recognition result are further determined by multi-dimensional cross recognition rules to determine the final recognition result.

[0060] Specifically, referring to FIG. Figure 2 In the three-dimensional ground penetrating radar signal derivation software, a set of three-dimensional data is output every 30 meters, and the three-dimensional data includes multiple B-SCAN maps (depth vertical direction information), multiple C-SCAN maps (horizontal lateral information), road positions and point coordinates. Specifically, the number of B-SCAN maps is N, wherein N is the number of measuring lines of the three-dimensional ground penetrating radar device; the number of C-SCAN maps is M, wherein M is the result of dividing the effective detection depth (unit: m) of the radar by 0.1 m, that is, a C-SCAN map is generated every 0.1 m, and all data is stored in the corresponding database with the road name and number, and the actual coordinate position information of each area is stored in a TXT format file. Further example: in the embodiment, the three-dimensional GPR data is collected by using the vehicle-mounted three-dimensional GPR device Stream UP, and the data acquisition and visualization tools used are uMap and IQMaps. According to the software and interface protocol, we can extract the original B-SCAN (18) and C-SCAN (25) maps of each area from the three-dimensional GPR data.

[0061] The collected original B-SCAN map and C-SCAN map are preprocessed; the original B-SCAN map contains a large amount of background signal, which makes it difficult to identify abnormal signals. In view of the signal noise characteristics, a Gaussian function band-pass filter is used to preprocess the B-SCAN map, and the specific formula is as follows:

[0062] F filtered (u,v)=F(u,v)·H(u,v)

[0063]

[0064] Wherein, F filtered (u,v) is the expression of the image after filtering in the frequency domain; F(u,v) is the expression of the image before filtering in the frequency domain; H(u,v) is the frequency response function of the Gaussian band-pass filter; f c is the center frequency between the lower cutoff frequency and the upper cutoff frequency; Δf is the bandwidth; n is the order of the filter, which determines the steepness of the filter. The B-SCAN map before and after processing is shown in Figure 3 .

[0065] For the original C-SCAN map, the soil density anomaly area is enhanced, and the color contrast between the abnormal area and other normal areas is enhanced. Referring to Figure 4 .

[0066] Referring to Figure 5 , then based on the preprocessed B-SCAN map, the first recognition result is obtained through the YOLO v12 model.

[0067] Specifically, the YOLO v12 model is constructed, and the YOLO v12 model comprises a backbone network, a neck network and a detection head; the backbone network comprises five convolution modules, and the convolution modules are used for extracting image features; a C3k2 is arranged between the second convolution module and the third convolution module and between the third convolution module and the fourth convolution module, and the C3k2 is used for assisting in extracting features; an A2C2f is arranged in the fourth convolution block and the fifth convolution block, and the A2C2f is a unique feature and comprises two ordinary convolution modules and an ABlock module of a local attention mechanism, which is a Window Attention attention mechanism; the neck network comprises two groups of “up-sampling modules, A2C2f and splicing layers” in cascade to form a middle column, one group of “convolution modules, splicing layers and A2C2f” and one group of convolution modules, splicing layers and C2k2” in cascade to form a right column; the first splicing layer and the second splicing layer of the middle column receive feature parameters in the second C2k3 and the first A2C2f in the backbone network respectively, and are used for fusing shallow features; the first splicing layer of the right column receives shallow feature information of the first A2C2f of the middle column, and the second splicing layer receives deep feature information of the last A2C2f of the backbone network. The residual efficient layer aggregation network (R-ELAN) is used in the backbone network and the neck network, the feature aggregation mode is redesigned based on the scaling technology; the A2C2f in the backbone network and the neck network is added with a weight of 0.1, so that the original feature aggregation capability can be retained while the calculation cost is reduced. The detection head is composed of a detection layer, is responsible for the final target detection task, integrates feature information from the neck network, and outputs the category and position information of the detected target.

[0068] The YOLO v12 model is trained. Three-dimensional ground penetrating radar data collected from more than 100 km of urban roads are used for model training and algorithm verification. The specific division is shown in Table 1.

[0069] Table 1

[0070]

[0071] Referring to as Figure 6 , Figure 7As shown, all the anomalies and water wells on the B-SCAN map are labeled using the LabelMe labeling tool. The labeled data is compiled into a VOC format dataset for training and verification of the B-SCAN image signal detection model. The dataset is divided into a training set and a verification set in a ratio of 8:2, resulting in 1224 data training samples (i.e. 944 defects and 280 water wells) and 306 verification samples (i.e. 236 defects and 70 water wells). The YOLO v12 target detection network is used to detect anomalies and water well signals in the B-SCAN map, and the imaging characteristics of the anomaly signal are analyzed. The irregularly reflected area in the rectangular region is analyzed. The training and verification of the model are carried out on a server equipped with dual AMD EPYC 9534 processors and NVIDIA GeForce RTX A600048G graphics cards. The training parameters are set as follows: 1000 rounds (early stopping strategy when the verification loss curve is stable), batch size of 8, learning rate of 0.001, input size of 512, and 4 threads for data loading to improve training speed. At the same time, we compared the original structure and other more popular attention mechanism methods, and the results are shown in Table 2.

[0072] Table 2

[0073]

[0074] The results show that the optimized model has the highest recall rate and precision for anomaly recognition, and has a higher prediction speed.

[0075] The YOLO v12 model is used to identify the preprocessed B-SCAN map to obtain preliminary identification results, which include normal signals, anomaly signals or water well signals. Since there are pseudo-anomaly signals in the anomaly signals, the signal pattern recognition method is introduced to filter out the pseudo-anomaly signals from the anomaly signals and correct the pseudo-anomaly signals to normal signals; as shown in Figure 8 The specific steps include: extracting the highlighted abnormal area in the B-SCAN map, and extracting it to a new blank canvas to form a stripe pattern. Based on the imaging principle of ground penetrating radar B-SCAN map, for example, taking void as an example, the dielectric constant of air is much smaller than that of road material, the reflection coefficient R at the upper interface of the void is negative, the reflection coefficient R at the lower interface is positive, and the reflection wave at the two interfaces is one negative and one positive. This case will show a flat positive upper boundary in the map, with the color changing from bright to dark and then to bright, and the difference is obvious. Because of the difference of ground penetrating radar equipment, some black color is negative reflection coefficient, and some white color is negative reflection coefficient. In this application, the top of the stripe pattern is black, which is determined as an anomaly signal; if the top is white, it is determined as a pseudo-anomaly signal and is corrected to a normal signal.

[0076] Referring to Figure 9 ,Figure 10 As shown, the position and number of abnormal signals or well signals are determined by the spatial positioning method of the B-SCAN atlas, and the correctness of the abnormal signals or well signals is verified by the position and number to obtain the final first identification result. Specifically, the N B-SCAN atlases are divided into KxV regions, the positions of the abnormal signals and well regions are recorded, and the positions and numbers of the abnormalities or wells are determined; in the N B-SCAN atlases of a single point, if abnormal signals or well signals appear on more than three B-SCAN atlases, and the signal interval does not exceed one atlas, it is further determined as an abnormal or well signal; if it only appears on one B-SCAN atlas or the signal interval exceeds one atlas, it is determined as a normal signal.

[0077] Further, 18 B-SCAN atlases are divided into 12x4 regions, the positions of abnormal signals and well regions are recorded, and a single signal is allowed to cross multiple regions, such as Figure 9 As shown. Then in the 18 B-SCAN atlases under a single point, the positions and numbers of abnormalities or wells are clearly distinguished. For one of the abnormal or well signals, if it appears continuously in three or more atlases, and the signal interval does not exceed one atlas, it is further determined as an abnormal signal, if it only appears in one B-SCAN atlas, or the signal interval exceeds one atlas, it is further determined as a normal signal, and the final first identification result is obtained after verification.

[0078] Referring to as Figures 11-13 As shown, the process of obtaining the second identification result includes: according to the pre-processed C-SCAN atlas, the region with white color in the atlas is the soil density abnormal region, and is marked and drawn on a new canvas, and the spatial positioning method of the C-SCAN atlas is used to determine the positions of suspected abnormalities or suspected wells. Specifically, the M C-SCAN atlases are divided into BxH regions, the positions of the abnormal regions are recorded and the depth size is counted; in the M C-SCAN atlases of a single point, if more than three C-SCAN atlases but not more than M atlases appear continuously, it is determined as a suspected abnormality; if more than 12 atlases appear continuously, it is determined as a suspected well; if none of the above occurs, it is determined as no soil abnormal signal.

[0079] Further example, 25 C-SCAN maps are divided into 11x8 regions, the location of each soil density anomaly is recorded, and the depth size of the location is counted, as shown in the figure. Then in the 25 C-SCAN maps of a single point, if the soil anomaly in a certain region appears continuously in 3 or more maps (i.e. 0.3 meters in depth), but does not exceed 12 maps (i.e. 1.2 meters in depth), it is determined to be a suspected anomaly; if the region appears continuously in 12 or more maps (i.e. 1.2 meters in depth), it is determined to be a suspected water well. Through this method, it can be determined whether there is an anomaly or a water well in the region; otherwise, it is determined that there is no soil anomaly signal.

[0080] Referring to Figure 14 After obtaining the first recognition result and the second recognition result, the final recognition result is determined by the optimized multi-dimensional cross-identification rule, and the optimized multi-dimensional cross-identification rule includes:

[0081] According to the second recognition result, if the second recognition result of the point is no anomaly, the final recognition result is no anomaly;

[0082] If the second recognition result of the point is a suspected anomaly or a suspected water well, the first recognition result of the point is further determined, if the first recognition result of the point is not determined as an anomaly or a water well, and no anomaly or water well is detected in N B-SCAN maps, the final recognition result is normal;

[0083] If the first recognition result of the point is not determined as an anomaly or a water well, but an anomaly or a water well has been detected in N B-SCAN maps, and the signal interval exceeds one map, the final recognition result is an anomaly or a water well;

[0084] If the first recognition result of the point is determined as an anomaly or a water well, the final recognition result is an anomaly or a water well.

[0085] The application also provides a detection result table 3 based on only the optimized YOLO v12 for B-SCAN map recognition, the addition of signal pattern recognition, and the addition of the optimized multi-dimensional cross-identification rule recognition.

[0086] Table 3

[0087]

[0088] The results show that compared with only using a single B-SCAN map interpretation method, after adding pattern recognition, the recall rate and accuracy of 150 point anomalies and water wells are increased by 6% to 13%; when using spatial positioning and stereoscopic recognition, multi-dimensional cross-interpretation algorithm, the recall rate is increased by about 10%, and the accuracy is increased by 30%. Therefore, the application shows that the application of deep learning in road underground disease diagnosis has made substantial progress, which helps to reduce the risk of road collapse and underground infrastructure damage.

Claims

1. A method for multi-dimensional diagnosis of road surface damage based on three-dimensional ground penetrating radar, characterized in that: Comprising Collect and organize three-dimensional ground penetrating radar road detection data, the data including N B-SCAN maps, M C-SCAN maps, road positions and point coordinates, wherein the B-SCAN maps and C-SCAN maps are preprocessed; Obtaining a first identification result based on the YOLO v12 model for the preprocessed B-SCAN map; Obtaining a second identification result based on the preprocessed C-SCAN map; Further determining the final identification result through the optimized multi-dimensional cross-identification rule for the first identification result and the second identification result.

2. The deep learning-based multi-dimensional diagnosis method for pavement damage of three-dimensional ground penetrating radar according to claim 1, characterized in that: The process of obtaining the first identification result includes Constructing and training the YOLO v12 model; Using the YOLO v12 model to identify the preprocessed B-SCAN map to obtain a preliminary identification result, the preliminary identification result including normal signals, abnormal signals or water well signals; Screening false abnormal signals from the abnormal signals through a signal pattern recognition method and correcting the false abnormal signals to normal signals; Determining the position and quantity of the abnormal signals or water well signals through a spatial positioning and stereoscopic recognition method of the B-SCAN map, and verifying the correctness of the abnormal signals or water well signals through the position and quantity to obtain the final first identification result.

3. The deep learning-based multi-dimensional diagnosis method for pavement damage of three-dimensional ground penetrating radar according to claim 2, characterized in that: The signal pattern recognition method includes Identifying the highlighted abnormal area in the B-SCAN map, extracting the highlighted abnormal area to a new blank canvas, analyzing the stripe pattern of the highlighted area, and determining the false abnormal signal and correcting it to the normal signal when the top radiation coefficient is positive.

4. The deep learning-based multi-dimensional diagnosis method for pavement damage of three-dimensional ground penetrating radar according to claim 3, characterized in that: The process of obtaining the second identification result includes According to the preprocessed C-SCAN map, the area with white color in the map is the soil density abnormal area, and is labeled and drawn to a new canvas; Using the spatial positioning and stereoscopic recognition method of the C-SCAN map to confirm the suspected abnormal or suspected water well position, specifically dividing the M C-SCAN maps into BxH regions, recording the position of the abnormal region and counting its depth size; In the M C-SCAN maps of a single point, if more than three C-SCAN maps but not more than M maps appear continuously, it is determined as a suspected abnormality; if more than 12 continuous appearances, it is determined as a suspected water well; If none of the above occurs, it is determined that there is no soil abnormal signal.

5. The deep learning-based multi-dimensional diagnosis method for pavement damage of three-dimensional ground penetrating radar according to claim 4, characterized in that: The optimized multi-dimensional cross-stereoscopic recognition rule includes According to the second identification result, if the point has no abnormality, the final identification result is normal; If the point has a suspected abnormality or a suspected water well, further determine the point according to the first identification result, if the point has no abnormality or water well, and no abnormality or water well is detected in the N B-SCAN maps, the final identification result is normal; If the point has no abnormality or water well, but abnormality or water well has been detected in the N B-SCAN maps, and the signal interval exceeds one map, the final identification result is abnormal or water well; If the point is determined to be abnormal or a water well, the final identification result is abnormal or a water well. 6.The deep learning-based multi-dimensional diagnosis method for pavement damage using a three-dimensional ground penetrating radar according to claim 2 or 5, characterized in that: The spatial positioning and stereoscopic recognition method of the B-SCAN map includes N B-SCAN maps are divided into K*V regions, the positions of abnormal signals and well regions are recorded, and the positions and quantities of the abnormal or well are determined; In N B-SCAN maps of a single point, if abnormal signals or well signals appear on more than three B-SCAN maps, and the signals are further determined as abnormal or well signals; if the signals appear on only one B-SCAN map or the signal interval exceeds one map, it is determined as a normal signal. 7.The deep learning-based multi-dimensional diagnosis method of pavement damage by ground penetrating radar according to claim 6, characterized in that: The process of constructing and training the YOLO v12 model includes A large number of B-SCAN maps are collected, and according to the actual exploration situation, the anchor frame is used to label abnormal signals and well regions, and a VOC format data set is generated, A YOLO v12 model is constructed by analyzing the imaging characteristics of abnormal signals and wells in B-SCAN maps, and the YOLO v12 model includes a backbone network, a neck network and a detection head. The VOC format data set is used to train the YOLO v12 model. The B-SCAN map is input into the trained YOLO v12 model to obtain the recognition result. 8.The deep learning-based three-dimensional ground penetrating radar multi-dimensional diagnosis method for pavement damage according to claim 7, characterized in that: The backbone network includes five convolution modules, the convolution modules are used to extract image features, C3k2 is arranged between the second and third convolution modules and between the third and fourth convolution modules, C3k2 is used to assist in extracting features, A2C2f is arranged in the fourth and fifth convolution blocks, the A2C2f is a unique feature, including two ordinary convolution modules and an ABlock module of a local attention mechanism. 9.The deep learning-based multi-dimensional diagnosis method of pavement damage by ground penetrating radar according to claim 8, characterized in that: The neck network includes two groups of "up-sampling module, A2C2f and splicing layer" cascaded to form a middle column, one group of "convolution module, splicing layer and A2C2f" and one group of convolution module, splicing layer and C2k2" are cascaded to form a right column, the first splicing layer and the second splicing layer of the middle column receive the feature parameters in the second C2k3 and the first A2C2f in the backbone network respectively, and are used to fuse shallow features; the first splicing layer of the right column receives the shallow feature information of the first A2C2f of the middle column, and the second splicing layer receives the deep feature information of the last A2C2f of the backbone network.

10. The deep learning-based multi-dimensional diagnosis method for pavement damage of three-dimensional ground penetrating radar according to claim 9, characterized in that: The preprocessing of the B-SCAN map is to denoise a large amount of background signals in the original B-SCAN map, and the soil density abnormal area of the C-SCAN map is enhanced in color contrast.

Images