An Optimization Method for Identifying Internal Road Surface Defects Based on Statistical Features of Radar Signals
By combining high-speed, high-precision three-dimensional array ground-penetrating radar with YOLO deep learning and radar signal statistical feature optimization methods, the problems of low identification efficiency and poor accuracy in the detection of internal road surface defects have been solved, achieving efficient and accurate defect identification.
Patent Information
- Application Number
- CN202410292686.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-14
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-03-14
AI Technical Summary
Existing technologies for detecting internal road surface defects using 3D ground-penetrating radar suffer from problems such as large data volume, low recognition efficiency, and serious omissions and misjudgments in the recognition results, making it difficult to effectively replace manual recognition through deep learning.
Non-destructive testing is performed using a high-speed, high-precision three-dimensional array ground-penetrating radar. Training and test sets are established through preprocessing and manual screening of labeled data. The YOLO deep learning network model is used for training, and the recognition results are filtered and optimized by combining the statistical features of radar signals to improve recognition accuracy.
It significantly improves the accuracy of identifying internal road surface defects, with the accuracy rate for identifying major defects reaching over 98% and the accuracy rate for identifying relatively serious defects reaching over 95%, effectively solving the problems of identification efficiency and accuracy.
Smart Images

Figure CN118171144B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of pest detection, specifically relating to an optimized method for identifying internal road surface defects based on the statistical characteristics of radar signals. Background Technology
[0002] Three-dimensional ground-penetrating radar (GPR) is a non-destructive testing device that utilizes the principle of high-frequency electromagnetic wave reflection to detect the internal condition of a road surface. It uses high-frequency electromagnetic waves (with a main frequency ranging from tens to hundreds of megahertz to gigahertz) in the form of wide-bandwidth short pulses, transmitted from the ground through a transmitting antenna T. After reflection from the underground strata or the target object, the waves return to the ground and are received by a ground receiving antenna R. By analyzing the waveform data of the received wave field, an image of the underground target is obtained.
[0003] 3D ground-penetrating radar (GPR), as an important geophysical method for rapid, high-resolution, and non-destructive detection, holds significant importance and value in the research and engineering practice of pavement internal defect detection. GPR technology offers real-time performance and high precision, meeting the needs for rapid, accurate, and non-destructive detection of highway defects. Currently, when using 3D GPR for pavement internal condition detection, the identification of pavement defects mainly relies on human experience, resulting in problems such as large data volumes and low identification efficiency. Although deep learning-based automatic radar image recognition technology can improve recognition efficiency and reduce human workload, the target identification results still suffer from numerous missed and false positives, making it difficult to completely replace manual identification of radar detection results. Summary of the Invention
[0004] To address the problem of numerous missed and incorrect identifications of road surface defects in 3D radar images using deep learning, this invention proposes an optimized method for identifying internal road surface defects based on statistical features of radar signals, which can effectively improve the accuracy of deep learning in identifying internal road surface defects.
[0005] To achieve the above objectives, the present invention provides the following solution: An optimization method for identifying internal road surface defects based on statistical features of radar signals includes the following steps: Step 1: Use high-speed, high-precision three-dimensional array ground-penetrating radar to perform non-destructive testing on the asphalt pavement and obtain radar data of the pavement interior; Step 2: Preprocess the radar data inside the road surface to obtain the data characteristics of road surface defects; Step 3: Manually screen and label the data features of the internal road surface defects, and divide them into training set and test set in sequence; Step 4: Input the training set into the YOLO deep learning network model for training to obtain the network weight model; Step 5: Use the network weight model to perform target identification and classification on the pavement internal defect maps in the test set to obtain target identification results; Step 6: Calculate the statistical characteristics of the waveform data within the range of the target recognition results, and filter and optimize the target recognition results based on the statistical characteristics.
[0006] Preferably, the method for preprocessing the radar data inside the road surface includes: zero offset processing, zero-point adjustment, and data gain; The zero-bias processing is used to make the single-channel wave signal symmetrical from left to right; The zero-point adjustment is used to eliminate the influence of direct waves and align the radar echoes of each channel in the depth direction. The gain adjustment is used to enhance the echo signal of the internal defects of the road surface, highlighting the defects.
[0007] Preferably, when manually screening and marking the data characteristics of the pavement internal defects, the pavement internal defects are mainly divided into four types: poor interlayer bonding of structural layers, loose and broken structure, structural faults, and foreign objects.
[0008] Preferably, the statistical characteristics include: energy amplitude and phase, in-phase axis characteristics, frequency domain spectrum characteristics, and electromagnetic wave attenuation characteristics; Wherein, the energy amplitude and phase are the average value and sign of the energy amplitude of a single radar wave, expressed as: ,in, xi Indicates the first wave of a certain channel i One sampling point; The in-phase axis characteristic is the depth difference between the maximum amplitudes of two adjacent radar waves, expressed as: ,in d 1 represents the maximum amplitude depth of the previous radar wave. d 2 represents the maximum amplitude depth of the subsequent radar wave; The frequency domain spectral characteristics are the average amplitude values of radar waves in the frequency domain, expressed as: Where s(k) is the radar spectrum, k =1, 2, ..., M ,in M The number of frequency components; The electromagnetic wave attenuation characteristic is the average energy attenuation coefficient of a single radar wave, expressed as follows: ,in, xi Indicates the first wave of a certain channel i One sampling point.
[0009] Preferably, the method for filtering and optimizing the target identification results based on the statistical features includes: Based on the target recognition results, the signal characteristics of different types of pavement defects are analyzed, and the importance of each statistical feature for the identification of pavement defects, i.e., the weight coefficient, is determined. By calculating and analyzing the product of the statistical features and weight coefficients of the target identification results, the filtering thresholds for various types of pavement internal defects are determined. Based on the filtering threshold, the target recognition results are filtered and optimized.
[0010] Preferably, the specific process of filtering and optimizing the target recognition results based on the filtering threshold includes: By calculating and analyzing the product of the maximum value of each statistical feature and the weight coefficient within a single disease range of the target identification results, the filtering threshold for various diseases is determined. The credibility of the disease identification results based on the network weight model in the test set is calculated, and diseases with credibility below the set filtering threshold are filtered out to optimize the target identification results.
[0011] The preferred method for calculating disease reliability is as follows: In the formula, y To assess the credibility of this disease, k 1, k 2, k 3, k 4 represents the weighting coefficients for energy amplitude and phase, in-phase axis characteristics, frequency domain spectrum characteristics, and electromagnetic wave attenuation characteristics. x 1, x 2, x 3, x 4 represents the calculated energy amplitude and phase, in-phase axis characteristics, frequency domain spectrum characteristics, and electromagnetic wave attenuation characteristics within each diseased area.
[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention preprocesses complete 3D radar detection data using methods such as zero-bias adjustment, zero-point adjustment, and data gain, sequentially exporting waveform and image data. Training and test datasets are then established through manual selection and annotation. The training dataset is input into a YOLO deep learning network model to train and adjust the network parameters. After training, the test dataset is input into the deep learning network, which then uses the network to identify internal road surface defects from ground-penetrating radar echo images. Based on this, statistical characteristics of the waveform data within the identification result range are calculated, and the identification results are filtered and optimized. This method effectively improves the accuracy of deep learning in identifying internal road surface defects. This invention can achieve an accuracy rate of over 98% for identifying major internal road surface defects and over 95% for identifying relatively severe defects. Attached Figure Description
[0013] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0014] Figure 1 This is a flowchart of the method for optimizing the deep learning identification results of pavement internal defects based on radar signal statistical features in this embodiment; Figure 2 These are data images obtained by three-dimensional ground-penetrating radar in this embodiment, where (a) is the original radar data image and (b) is the radar data image obtained after preprocessing. Figure 3 This is a schematic diagram of the exported image data size suitable for the YOLO deep learning neural network in this embodiment; Figure 4 This is a characteristic diagram of poor interlayer bonding defects in the internal structure of the road surface in this embodiment; Figure 5 This is a characteristic diagram of the loose and broken interlayer structure of the road surface in this embodiment; Figure 6 This is a feature diagram of the fault defects in the internal structural layer of the pavement structure in this embodiment; Figure 7 This is a feature diagram of foreign objects inside the road surface structure in this embodiment. Detailed Implementation
[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0017] Example 1 like Figure 1 As shown, this invention proposes a method for optimizing the deep learning identification results of pavement internal defects based on radar signal statistical features. The method specifically includes: Step 1: Use high-speed, high-precision three-dimensional array ground-penetrating radar to perform non-destructive testing on asphalt pavement; Step 2: Preprocess the collected radar data inside the road surface, including zero bias processing, zero point adjustment and data gain, so as to eliminate the influence of clutter and highlight the data characteristics of road surface defects. Step 3: Export the radar waveform and image data processed in Step 2, and manually filter and label the exported data, and divide it into training set and test set in sequence; Step 4: Input the training set image data obtained in Step 3 into the YOLO deep learning network model for training to obtain the network weight model; Step 5: Input the test image data obtained in Step 4 into the weights of the deep learning network model that has finished training, so as to identify and classify the internal road surface defects detected by ground penetrating radar. Step 6: Calculate the statistical characteristics of the waveform data within the target recognition result range in Step 5, thereby filtering the recognition results and improving their accuracy.
[0018] In this embodiment, step 1: A three-dimensional ground-penetrating radar is used to perform non-destructive testing on the asphalt pavement. For example... Figure 2 As shown.
[0019] In step 1, when testing the asphalt pavement, it is necessary to ensure that the equipment is working properly and the parameters are set reasonably so that the test data is free from missing channels, omissions, or incomplete direct waves.
[0020] The three-dimensional ground-penetrating radar is a 14-channel three-dimensional array ground-penetrating radar with a 1.2GHz ground-coupled radar antenna.
[0021] In this embodiment, step 2: preprocessing the internal radar data of the road surface generated in step 1, including zero bias processing, zero point adjustment and data gain, so as to eliminate the influence of clutter and highlight the data characteristics of internal road surface defects.
[0022] The specific method for radar data preprocessing is as follows: Zero-bias processing is used to make single-channel wave signals symmetrical from left to right; Zero-point adjustment is used to eliminate the influence of direct waves and align the radar echoes of each channel in the depth direction; Gain adjustment is used to enhance the echo signal of internal road surface defects, highlighting the defects.
[0023] In this embodiment, step 3: Export the radar waveform and image data processed in step 2, and manually filter and label the exported data, dividing it into training and test sets in sequence. For example... Figure 3 As shown.
[0024] When exporting the processed radar data, it is necessary to ensure the original state of the disease features in the map, while satisfying the size model recognition of the YOLO deep learning network model, and to ensure that there is an overlap between two adjacent map data, with the overlap accounting for more than 10% of the map area.
[0025] Specifically, the data export method mainly involves image data export. Based on the image size requirements for YOLO deep learning, a 400×400 pixel image size is selected. Each image exports 200 channels, meaning one channel of data occupies one image pixel. There is overlap between two images, with the overlap exceeding 10% of the entire image, i.e., ≥20 overlapping channels.
[0026] The data filtering and labeling described above primarily categorize pavement internal defects into four main types: poor interlayer bonding, internal structural loosening and fragmentation, structural faults, and foreign objects. The graphic characteristics of each type of pavement internal defect are as follows: Poor interlayer bonding of structural layers: Obvious highlights appear at the interfaces of structural layers, with discontinuous phase axes and multiple reflections, such as... Figure 4 .
[0027] Loose and fragmented structure within the structural layers: Significant energy enhancement is observed within the structural layers, with disordered phase axes and the superposition of multiple parabolic features, such as... Figure 5 .
[0028] Structural fault: The pavement structural layers show obvious faulting and a certain degree of highlighting, such as... Figure 6 .
[0029] Foreign matter: Significant energy enhancement is observed within the structural layers, accompanied by pronounced hyperbolic stacking, such as... Figure 7 .
[0030] In this embodiment, step 4: The training set image data obtained in step 3 is fed into the YOLO deep learning network model for training to obtain the network weight model.
[0031] The YOLO deep learning network model, which is currently the mainstream model, is adopted. The manually labeled training set is fed into the network model, and the ratio of the number of training sets for various diseases is 1:1:1:1, so as to train the network model.
[0032] The training parameters are defined as follows: kernel size 3, stride 2, epoch 100, batch size 16, initial learning rate 0.01, recurrent learning rate 0.2, SGD optimizer with a momenrum of 0.937, weight decay of 0.0005, IOU training threshold of 0.2, cross-entropy loss function, and class 4.
[0033] In this embodiment, step 5: input the test set image data obtained in step 4 into the weights of the deep learning network model that has finished training, so as to identify and classify the internal road surface defects detected by ground penetrating radar.
[0034] The target identification is divided into four categories: poor interlayer bonding of structural layers, loose and broken structure, structural fault, and foreign objects. The first category is poor interlayer bonding of structural layers, the second category is loose and broken structure, the third category is structural fault, and the fourth category is foreign objects.
[0035] In this embodiment, step 6: perform maximum and minimum normalization processing on the waveform data within the target recognition result range in step 5, and calculate the statistical characteristics of the processed waveform data, thereby filtering the recognition results and improving the accuracy of the recognition results.
[0036] The statistical characteristics of the ground-penetrating radar waveform data include: Energy Amplitude and Phase: The average value and sign of the energy amplitude of a single radar wave. Right now , where x i This represents the i-th sampling point of a certain channel wave; In-phase axis characteristics: Depth difference in the maximum amplitude of two adjacent radar waves Right now Where d1 is the maximum amplitude depth of the previous radar wave, and d2 is the maximum amplitude depth of the next radar wave; Frequency domain spectral characteristics: Average amplitude of radar wave in the frequency domain Right now , where s(k) is the radar spectrum (frequency-amplitude), k=1,2,...,M, where M is the number of frequency components; Electromagnetic wave attenuation characteristics: mean energy attenuation coefficient of a single radar wave Right now , where x i This represents the i-th sampling point of a certain channel wave.
[0037] By analyzing the labeling results in the training set, we studied the signal characteristics of different types of road surface defects, thereby determining the importance of four parameters for defect identification and their weighting coefficients.
[0038] Then, by calculating and analyzing the product of the four statistical features of the artificially labeled results in the training set with the weight coefficients, the filtering thresholds for various diseases are determined.
[0039] The waveform data of the target recognition results obtained from the YOLO deep learning network model in step 5 are subjected to maximum and minimum normalization, and the energy amplitude, energy concentration, energy amplitude variance, and energy fluctuation of the waveform data are calculated. The maximum value of each parameter within each disease range is multiplied by the weight coefficient of each parameter and analyzed to determine the filtering threshold of each type of disease. The credibility of the disease recognition results based on the YOLO deep learning network model in the training set is calculated, and diseases with credibility below the set threshold are filtered out, thereby improving the accuracy of the recognition results and optimizing the recognition results.
[0040] The method for calculating the reliability of disease is as follows: In the formula, y represents the confidence level of the disease, k1, k2, k3, and k4 are the weighting coefficients of energy amplitude and phase, in-phase axis characteristics, frequency domain spectrum characteristics, and electromagnetic wave attenuation characteristics, respectively, and x1, x2, x3, and x4 are the calculated values of energy amplitude and phase, in-phase axis characteristics, frequency domain spectrum characteristics, and electromagnetic wave attenuation characteristics within each disease area.
[0041] By comparing the disease's confidence level y with the threshold obtained from the analysis, identification results below the threshold are filtered out, thereby effectively reducing misjudgments and omissions in the identification results and improving the accuracy of the identification results. The accuracy rate for identifying major diseases reaches over 98%, and the accuracy rate for identifying relatively severe diseases can reach over 95%.
[0042] The weighting coefficients of the four parameters obtained so far are as follows: energy amplitude and phase are 3, in-phase axis characteristics are 3, frequency domain spectrum characteristics are 2, and electromagnetic wave attenuation characteristics are 2. The filtering thresholds for various defects are: poor interlayer bonding defects are 7, loose and broken defects within the structural layer are 5, structural faults are 3, and foreign objects are 2.
[0043] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for identifying and optimizing internal road surface diseases based on radar signal statistical characteristics, characterized in that, Includes the following steps: Step 1: Use high-speed, high-precision three-dimensional array ground-penetrating radar to perform non-destructive testing on the asphalt pavement and obtain radar data of the pavement interior; Step 2: Preprocess the radar data inside the road surface to obtain the data characteristics of road surface defects; Step 3: Manually screen and label the data features of the internal road surface defects, and divide them into training set and test set in sequence; Step 4: Input the training set into the YOLO deep learning network model for training to obtain the network weight model; Step 5: Use the network weight model to perform target identification and classification on the pavement internal defect maps in the test set to obtain target identification results; Step 6: Calculate the statistical characteristics of the waveform data within the range of the target recognition results, and filter and optimize the target recognition results based on the statistical characteristics; The statistical characteristics include: energy amplitude and phase, in-phase axis characteristics, frequency domain spectrum characteristics, and electromagnetic wave attenuation characteristics. Wherein, the energy amplitude and phase are the average value and sign of the energy amplitude of a single radar wave, expressed as: ,in, xi Indicates the first wave of a certain channel i One sampling point; The in-phase axis characteristic is the depth difference between the maximum amplitudes of two adjacent radar waves, expressed as: ,in d 1 represents the maximum amplitude depth of the previous radar wave. d 2 represents the maximum amplitude depth of the subsequent radar wave; The frequency domain spectral characteristics are the average amplitude values of radar waves in the frequency domain, expressed as: Where s(k) is the radar spectrum, k =1, 2, ..., M ,in M The number of frequency components; The electromagnetic wave attenuation characteristic is the average energy attenuation coefficient of a single radar wave, expressed as follows: ,in, xi Indicates the first wave of a certain channel i One sampling point.
2. The method for identifying and optimizing internal road surface defects based on radar signal statistical features according to claim 1, characterized in that, The methods for preprocessing the radar data inside the road surface include: zero bias processing, zero point adjustment, and data gain. The zero-bias processing is used to make the single-channel wave signal symmetrical from left to right; The zero-point adjustment is used to eliminate the influence of direct waves and align the radar echoes of each channel in the depth direction. The gain adjustment is used to enhance the echo signal of the internal defects of the road surface, highlighting the defects.
3. The method for identifying and optimizing internal road surface defects based on radar signal statistical features according to claim 1, characterized in that, When manually screening and labeling the data characteristics of the internal pavement defects, the internal pavement defects are mainly divided into four types: poor interlayer bonding of structural layers, loose and broken structure, structural faults, and foreign objects.
4. The method for identifying and optimizing internal road surface defects based on radar signal statistical features according to claim 1, characterized in that, Methods for filtering and optimizing target recognition results based on the aforementioned statistical characteristics include: Based on the target recognition results, the signal characteristics of different types of pavement defects are analyzed, and the importance of each statistical feature for the identification of pavement defects, i.e., the weight coefficient, is determined. By calculating and analyzing the product of the statistical features and weight coefficients of the target identification results, the filtering thresholds for various types of pavement internal defects are determined. Based on the filtering threshold, the target recognition results are filtered and optimized.
5. The method for identifying and optimizing internal road surface defects based on radar signal statistical features according to claim 4, characterized in that, The specific process of filtering and optimizing the target recognition results based on the filtering threshold includes: By calculating and analyzing the product of the maximum value of each statistical feature and the weight coefficient within a single disease range of the target identification results, the filtering threshold for various diseases is determined. The credibility of the disease identification results based on the network weight model in the test set is calculated, and diseases with credibility below the set filtering threshold are filtered out to optimize the target identification results.
6. The method for identifying and optimizing internal road surface defects based on radar signal statistical features according to claim 5, characterized in that, The method for calculating the reliability of disease is as follows: ; In the formula, y To assess the credibility of this disease, k 1, k 2, k 3, k 4 represents the weighting coefficients for energy amplitude and phase, in-phase axis characteristics, frequency domain spectrum characteristics, and electromagnetic wave attenuation characteristics. x 1, x 2, x 3, x 4 represents the calculated energy amplitude and phase, in-phase axis characteristics, frequency domain spectrum characteristics, and electromagnetic wave attenuation characteristics within each diseased area.
Citation Information
Patent Citations
Radar-based road internal disease three-dimensional display and characteristic signal determination method
CN115902877A
Ground penetrating radar road multiple hidden disease real-time identification method based on YOLO
CN117269954A
