A method for identifying internal crack damage in pavement based on damage classification

Abstract

The present invention relates to a method for identifying crack diseases inside a pavement based on disease classification. Based on the recognition results of pavement crack diseases by the YOLO deep learning network, the method of the present invention first filters the crack features of bridge structures, internal foreign objects, etc. by analyzing the time-frequency domain characteristics of electromagnetic signals, and then combines the experience of identifying artificial crack diseases to classify the crack diseases in the data set according to the degree of obviousness of the crack diseases, and establishes filtering and classification rules for crack diseases, thereby improving the recognition accuracy of crack diseases, so that the results of automatic recognition can be applied to actual projects, and can improve the efficiency of manual discrimination to a certain extent. The method of the present invention can achieve an accuracy rate of 96% for the recognition results of internal cracks in the pavement.

Description

Technical Field

[0001] The invention relates to the technical field of pavement internal crack monitoring based on radar signal processing, and in particular to a pavement internal crack disease identification method based on disease classification. Background Art

[0002] When identifying the characteristics of cracks, because their characteristics are mainly distributed between structural layers, the up and down fluctuations of the structural layers and foreign objects inside the road surface will produce characteristics similar to cracks, which will cause the automatic identification results to be misjudged. In addition, the current evaluation of the recognition effect of automatic identification is greatly affected by the accuracy of the established data set. However, due to the subjectivity of the identification results of cracks, the recognition results of the automatic identification software are not ideal and cannot be applied to actual engineering. Summary of the invention

[0003] In view of the problems existing in the prior art, the purpose of the present invention is to provide a method for identifying crack damage inside a pavement based on damage classification, the method specifically comprising:

[0004] Step (1): pre-process the detected radar data to make the characteristics of cracks inside the pavement more obvious;

[0005] Step (2): exporting the preprocessed radar slice map data and its corresponding electromagnetic signal data;

[0006] Step (3): Use the radar slice map data derived from step (2) to establish a crack disease data set, and label the data set, where crack diseases with obvious crack characteristics are labeled as C1, and crack diseases with relatively weak characteristics are labeled as C2;

[0007] Step (4): Bring the data set obtained in step (3) into the YOLO deep learning network model for training to obtain a network weight model;

[0008] Step (5): export radar data to be identified of the same size and type as in step (2), bring the obtained slice map data into the network weight model for training, and obtain the recognition result of pavement cracks;

[0009] Step (6): Analyze the electromagnetic signal data derived in step (5) to filter out crack-like features such as bridge structure and internal foreign matter;

[0010] Step (7): Further filter the filtered recognition results and classify the crack disease characteristics.

[0011] Beneficial effects of the present invention:

[0012] Based on the recognition results of pavement crack diseases by the YOLO deep learning network, the present invention first filters the crack features of bridge structures, internal foreign objects, etc. by processing and analyzing the time-frequency domain characteristics of electromagnetic signals, and then combines the recognition experience of artificial crack diseases to classify the crack diseases in the data set according to the obvious degree of the crack diseases, and establishes filtering and classification rules for crack diseases, thereby improving the recognition accuracy of crack diseases, so that the results of automatic recognition can be applied to actual projects. The method of the present invention can make the recognition result accuracy of internal cracks in the pavement reach 96%, and if suspected crack diseases are reviewed with manual assistance, the accuracy can reach 99%. Therefore, this method can improve the recognition accuracy of crack diseases while ensuring that the recognition results are not missed, and can improve the recognition efficiency of crack diseases to a certain extent.

[0013] Ground penetrating radar detects cracks inside the pavement by analyzing the difference in the propagation speed of electromagnetic signals in different media. However, the electromagnetic signal attenuates quickly during propagation and is easily interfered by electromagnetic signals and noise in the surrounding environment. Preprocessing the collected radar data by means of zero bias removal, zero point adjustment and data gain can effectively remove interference signals and enhance the signal characteristics of cracks inside the pavement. Since the size of cracks is relatively small, they are generally identified by radar slice maps. However, radar maps are generally obtained based on electromagnetic signal data. The electromagnetic signal data derived from the disease can lay the foundation for filtering crack characteristics such as bridge structures and internal foreign bodies in the later stage. Since the crack disease in the same place often has characteristics at multiple depths in the radar slice map, and its characteristics show the characteristics of "weak-strong-weak", by grading the crack diseases in the data set, the characteristics of crack diseases can be better captured, the accuracy of crack disease identification can be improved, and the missed results can be reduced, while laying the foundation for the grading of recognition results. YOLO (You Only LookOnce) is a real-time target detection model that predicts the location and category of the target through a single forward propagation, and has the advantages of fast speed and high accuracy. Step 4 generates a disease feature extraction model based on the labeled crack disease data set, laying the foundation for the extraction of crack disease features in the measured data in step 5; step 5 lays the foundation for the automatic identification of crack diseases by identifying crack diseases in the measured data. The medium of crack diseases inside ordinary pavements is often air or water (with a dielectric constant of 1 or 81), while bridges or foreign objects are often made of metal (with a dielectric constant of ∞), which leads to a large difference between the electromagnetic signals collected by the ground penetrating radar and crack diseases, so that the crack features of bridge structures, internal foreign objects, etc. can be accurately identified by extracting electromagnetic signal features. The present invention can effectively improve the accuracy of crack disease identification by formulating a series of crack disease identification and filtering rules, and by grading crack diseases, it can ensure that crack diseases are considered controllable without missing crack diseases, and improve the efficiency of manual identification to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of different types of cracks in Example 2 of the present invention. DETAILED DESCRIPTION

[0015] The terms used in the present invention, unless otherwise specified, generally have the meanings commonly understood by those of ordinary skill in the art.

[0016] The present invention is further described in detail below in conjunction with specific examples and with reference to data. The following examples are only intended to illustrate the present invention and are not intended to limit the scope of the present invention in any way.

[0017] Example 1

[0018] A method for identifying crack damage inside a pavement based on damage classification, the method specifically comprising:

[0019] Step (1): pre-process the detected radar data to make the characteristics of cracks inside the pavement more obvious;

[0020] As a specific implementation scheme, the preprocessing method in step (1) includes: zero bias removal, zero point adjustment and data gain, and the gain coefficient of the waveform is adjusted; specifically, the parameter of zero bias removal is 90%; the zero point adjustment selects the first positive maximum positive amplitude of the direct wave of each channel; the data gain selects the compensation gain, the gain coefficient is controlled between 1.0-2.0, and the data chromatogram uniformly uses the grayscale chromatogram.

[0021] Step (2): exporting the preprocessed radar slice map data and its corresponding electromagnetic signal data;

[0022] As a specific implementation scheme, the size of the radar slice image exported in step (2) is selected to be 400×100, the number of electromagnetic signal channels is set to 400, and the overlap of each adjacent slice is 5%, that is, 20 channels of electromagnetic signal data.

[0023] Step (3): Use the radar slice map data derived from step (2) to establish a crack disease data set, and label the data set, where crack diseases with obvious crack characteristics are labeled as C1, and crack diseases with relatively weak characteristics are labeled as C2;

[0024] The specific marking method is as follows:

[0025] Normalize the electromagnetic signal amplitude at the corresponding depth of the radar slice map data derived in step (2). First, normalize the electromagnetic signal amplitude at the crack disease location and its surroundings (30 electromagnetic waves before and after), calculate the variance of each electromagnetic signal amplitude, and determine the difference threshold of C1 and C2 crack diseases to achieve the classification of the degree of obviousness of crack disease characteristics;

[0026] Among them, the normalization formula of the electromagnetic signal amplitude is:

[0027] In the formula, ——The normalized value of the amplitude of the electromagnetic wave of the i-th channel;

[0028] ——the amplitude of the electromagnetic wave of the i-th channel;

[0029] ——The error between the maximum and minimum amplitudes of electromagnetic waves around the crack disease;

[0030] Among them, the calculation formula for the variance of the electromagnetic signal amplitude at the crack disease location and the electromagnetic signal amplitude around it is:

[0031] In the formula, ——The average value of the normalized electromagnetic signal amplitude within the range; ——Amplitude variance of the normalized electromagnetic signal within the range;

[0032] Among them, C1 type crack disease Greater than or equal to 0.7, C2 type crack disease Between 0.3 and 0.7.

[0033] After step (3) is completed, data enhancement technology is used to make the data amounts of C1 and C2 in the data set the same or substantially the same. Specifically, the data enhancement technology refers to flipping, rotation, scale transformation, random extraction and / or color jittering.

[0034] Step (4): Bring the data set obtained in step (3) into the YOLO deep learning network model for training to obtain the network weight model.

[0035] Step (5): Export the radar data to be identified of the same size and type as step (2), bring the obtained slice map data into the network weight model for training, and obtain the recognition result of the pavement cracks.

[0036] Step (6): Analyze the electromagnetic signal data derived in step (5) to filter out crack-like features such as bridge structure and internal foreign matter;

[0037] Specifically, in step (6), crack-like features such as bridge structure and internal foreign matter are filtered by first performing Fourier transformation on the collected electromagnetic signal data, then extracting the time-frequency domain feature values ​​of the collected electromagnetic signal in sequence, and filtering the crack-like features such as bridge structure and internal foreign matter by determining the threshold of the time-frequency domain feature values;

[0038] Among them, since the single-channel electromagnetic signal is a discrete signal with limited sampling data points, discrete Fourier transform is required, and the formula is: ;

[0039] Where: ——The amplitude corresponding to the k-frequency component;

[0040] x n ——is a discrete time signal;

[0041] N is the number of sampling data points, which is 256 in the present invention; e is a constant; i is an imaginary unit;

[0042] Among them, the time-frequency domain characteristic signals used include:

[0043] Energy amplitude variance: the second-order center distance of the energy amplitude of a single radar channel

[0044] Right now ;

[0045] Energy amplitude waveform attenuation degree: waveform factor of the energy amplitude of a single radar channel

[0046] Right now ;

[0047] Among them, xi represents the i-th sampling point of a certain wave, and represents the amplitude mean of a single wave;

[0048] make , when the P value is greater than 5, it is a bridge or a foreign object.

[0049] Step (7): Further filter the filtered recognition results and classify the crack disease characteristics. The specific method is as follows:

[0050] S1: Determination of the continuity of the recognition results in the depth direction, that is, if the overlap area of ​​the recognition frames at the same position at adjacent depths is more than 50%, it is considered to be the same crack disease, otherwise it is two or more crack diseases;

[0051] S2 classifies and filters based on the number, confidence, and crack type of crack identification results at the same location;

[0052] The crack classification and filtering rules mainly include:

[0053] In the crack disease identification results at the same position, when the number of crack identification results is less than 3, the identification results are directly filtered; when the number of crack identification results is greater than or equal to 3, the judgment of S22 is performed;

[0054] In the crack disease identification results at the same position of S22, if there is a C1 type crack disease, the following identification is performed:

[0055] 1) When the number of identification results of C1 crack damage is greater than or equal to 5, it is determined to be a clear crack; the so-called clear crack means that the crack clearly exists, and its identification result does not need to be manually reviewed later (the same below);

[0056] 2) When the number of identification results of C1 cracks is less than 5, if the confidence of the identified C1 crack damage is greater than or equal to 0.7, it is determined to be a clear crack;

[0057] 3) When the number of identified C1 cracks is less than 5, if the identified C1 crack disease has a confidence level less than 0.7, the following judgment is made:

[0058] a If the confidence level of C1 crack disease is greater than or equal to 0.5, and there is no C2 crack disease, it is judged as suspected crack. The suspected crack means that there is a high possibility of crack existence, and the identification result needs to be manually reviewed and confirmed later (the same below);

[0059] b If the confidence level of C1 crack disease is greater than or equal to 0.5, and there are 4 or more C2 crack diseases, it is confirmed as a clear crack disease;

[0060] c If the confidence level of C1 crack disease is greater than or equal to 0.5, and there are less than 4 C2 crack diseases, if the confidence level of C2 crack diseases is greater than 0.8, it is a clear crack disease. If the confidence level of the identified C2 crack disease is less than or equal to 0.8, it is judged as a suspected crack;

[0061] d. If the confidence level of the C1 crack disease is less than 0.5, and if the number of C2 crack diseases is greater than 2, it is judged as a suspected crack disease. If there are only 2 or less C2 crack diseases, the judgment result is filtered;

[0062] S23 When only C2 crack disease exists in the crack disease identification results at the same location, then

[0063] 1) When the number of C2 crack disease identification results is greater than or equal to 5, it is judged as suspected cracks;

[0064] 2) When the number of C2 crack disease discrimination results is less than 5, but there are more than 2 results with a confidence level greater than 0.8 in the discrimination results, it is judged as a suspected crack. If there are less than 2 results with a confidence level greater than 0.8 in the discrimination results, the discrimination results are filtered.

[0065] Example 2

[0066] Based on the method in Example 1, the present application performs pavement internal crack disease identification based on disease classification on a typical highway in Jiangsu Province, and the specific steps are as follows:

[0067] Step (1): preprocessing the detected radar data, the processing methods include: zero bias removal, zero point adjustment and data gain, and appropriately adjusting the gain coefficient of the waveform to make the characteristics of the crack disease inside the pavement more obvious;

[0068] Specifically: the radar detection data obtained is a 14-channel three-dimensional array ground penetrating radar, and the antenna type is a 1.2 GHz ground-coupled radar antenna.

[0069] The specific parameters of the radar preprocessing method are:

[0070] When performing radar data preprocessing, the parameter for zero bias removal is set to 90%; the zero point adjustment selects the first positive maximum positive amplitude of the direct wave of each channel; the data gain selects the compensation gain, the gain coefficient is controlled between 1.0-2.0, and the data chromatogram uniformly uses the grayscale chromatogram;

[0071] The radar data zero point adjustment should be done by making the positions of the structural layers of each channel as close as possible without obvious breaks;

[0072] Step (2): Export the preprocessed radar slice map data (one slice map is exported for each depth point) and export its corresponding electromagnetic signal data;

[0073] Specifically, the size of the exported radar slice image is selected as 400×100, the number of electromagnetic signal channels is set to 400, and the overlap of each adjacent slice is 5%, that is, 20 channels of electromagnetic signal data;

[0074] Step (3): Use the radar slice map data derived from step (2) to establish a crack disease dataset and label the dataset, where crack diseases with obvious crack characteristics are labeled as C1, and crack diseases with relatively weak characteristics are labeled as C2 (e.g. Figure 1 shown);

[0075] The specific marking method is as follows:

[0076] Normalize the electromagnetic signal amplitude at the corresponding depth of the radar slice map data derived in step (2). First, normalize the electromagnetic signal amplitude at the crack disease location and its surroundings, calculate the variance of each electromagnetic signal amplitude, and determine the difference threshold of C1 and C2 crack diseases to achieve the classification of the degree of obviousness of crack disease characteristics;

[0077] Among them, the normalization formula of the electromagnetic signal amplitude is:

[0078] In the formula, ——The normalized value of the amplitude of the electromagnetic wave of the i-th channel;

[0079] ——the amplitude of the electromagnetic wave of the i-th channel;

[0080] ——The error between the maximum and minimum amplitudes of electromagnetic waves around the crack disease;

[0081] Among them, the calculation formula for the variance of the electromagnetic signal amplitude at the crack disease location and the electromagnetic signal amplitude around it is:

[0082] In the formula, ——The average value of the normalized electromagnetic signal amplitude within the range; ——Amplitude variance of the normalized electromagnetic signal within the range;

[0083] Among them, C1 type crack disease Greater than or equal to 0.7, C2 type crack disease Between 0.3 and 0.7.

[0084] After the data set is labeled, data enhancement technology is used to make the data volume of C1 and C2 in the data set basically the same. There are many methods of data enhancement, such as flipping, rotating, scaling, random extraction, color jittering, etc. to augment the data set, thereby improving the robustness of the recognition results.

[0085] According to the collected ground penetrating radar detection data, the present invention selects 100 electromagnetic wave data corresponding to obvious crack diseases (C1), crack diseases with weaker characteristics (C2), bridges, and foreign objects. The proportion of the P value of the electromagnetic signal within the above range is shown in Table 1 below.

[0086] Table 1 The proportion of P values ​​of different types of cracks in the corresponding range

[0087]

[0088] From the above table, we can see that the use of P value to filter the features of bridges and foreign objects has a very accurate effect, which can effectively eliminate the influence of crack features such as bridges and foreign objects on the recognition results and improve the accuracy of crack disease recognition; because the amplitude of the electromagnetic signal has a higher degree of refinement, the use of Differentiating between C1 and C2 crack diseases can more accurately compare the characteristic differences between C1 and C2 crack diseases.

[0089] In this step, C1 refers to crack damage with obvious crack characteristics, which is not necessarily a real crack and needs to be identified according to the method in step (7).

[0090] Step (4): Bring the data set obtained in step (3) into the YOLO deep learning network model for training to obtain a network weight model;

[0091] Specifically, the parameters of the training model are:

[0092] The convolution kernel size is 3, the stride is 2, the epoch is 100, the batch size is 16, the initial learning rate is 0.01, the cyclic learning rate is 0.2, the optimizer uses SGD, its momenrum is 0.937, the weight decay is 0.0005, the IOU training threshold is 0.2, the loss function is the cross entropy loss function, and the class is 2.

[0093] Step (5) exporting radar data to be identified of the same size and type as step (2), bringing the obtained slice map data into the network weight model for training, and obtaining the recognition result of the pavement cracks;

[0094] Step (6): Analyze the electromagnetic signal data derived in step (5) to filter out crack-like features such as bridge structure and internal foreign matter;

[0095] Specifically, when filtering the crack-like features of the bridge structure, internal foreign matter, etc. in step (6), the collected electromagnetic signal data is first subjected to Fourier transformation, and then the time-frequency domain characteristic values ​​of the collected electromagnetic signal are extracted in sequence, and the filtering of the crack-like features of the bridge structure, internal foreign matter, etc. is achieved by determining the threshold value of the time-frequency domain characteristic value;

[0096] Among them, since the single-channel electromagnetic signal is a discrete signal with limited sampling data points, discrete Fourier transform is required, and the formula is:

[0097] Where: ——The amplitude corresponding to the k-frequency component;

[0098] x n ——is a discrete time signal;

[0099] N is the number of sampling data points, which is 256 in the present invention; e is a constant; i is an imaginary unit;

[0100] Among them, the time-frequency domain characteristic signals used include:

[0101] Energy amplitude variance: the second-order center distance of the energy amplitude of a single radar channel

[0102] Right now ;

[0103] Energy amplitude waveform attenuation degree: waveform factor of the energy amplitude of a single radar channel

[0104] Right now ;

[0105] Among them, xi represents the i-th sampling point of a certain wave, and represents the amplitude mean of a single wave;

[0106] make , when the P value is greater than 5, it is a bridge or a foreign object.

[0107] Step (7): further filtering the filtered recognition results and classifying the crack disease characteristics, thereby improving the accuracy of crack disease recognition;

[0108] The specific method is as follows:

[0109] S1: Determination of the continuity of the recognition results in the depth direction, that is, if the overlap area of ​​the recognition frames at the same position at adjacent depths is more than 50%, it is considered to be the same crack disease, otherwise it is two or more crack diseases;

[0110] S2 classifies and filters based on the number, confidence, and crack type of crack identification results at the same location;

[0111] The crack classification and filtering rules mainly include:

[0112] In the crack disease identification results at the same position, when the number of crack identification results is less than 3, the identification results are directly filtered; when the number of crack identification results is greater than or equal to 3, the judgment of S22 is performed;

[0113] In the crack disease identification results at the same position of S22, if there is a C1 type crack disease, the following identification is performed:

[0114] 1) When the number of identification results of C1 crack damage is greater than or equal to 5, it is determined to be a clear crack; the so-called clear crack means that the crack clearly exists, and its identification result does not need to be manually reviewed later (the same below);

[0115] 2) When the number of identification results of C1 cracks is less than 5, if the confidence of the identified C1 crack damage is greater than or equal to 0.7, it is determined to be a clear crack;

[0116] 3) When the number of identified C1 cracks is less than 5, if the identified C1 crack disease has a confidence level less than 0.7, the following judgment is made:

[0117] a If the confidence level of C1 crack disease is greater than or equal to 0.5, and there is no C2 crack disease, it is judged as suspected crack. The suspected crack means that there is a high possibility of crack existence, and the identification result needs to be manually reviewed and confirmed later (the same below);

[0118] b If the confidence level of C1 crack disease is greater than or equal to 0.5, and there are 4 or more C2 crack diseases, it is confirmed as a clear crack disease;

[0119] c If the confidence level of C1 crack disease is greater than or equal to 0.5, and there are less than 4 C2 crack diseases, if the confidence level of C2 crack diseases is greater than 0.8, it is a clear crack disease. If the confidence level of the identified C2 crack disease is less than or equal to 0.8, it is judged as a suspected crack;

[0120] d. If the confidence level of the C1 crack disease is less than 0.5, and if the number of C2 crack diseases is greater than 2, it is judged as a suspected crack disease. If there are only 2 or less C2 crack diseases, the judgment result is filtered;

[0121] S23 When only C2 crack disease exists in the crack disease identification results at the same location, then

[0122] 1) When the number of C2 crack disease identification results is greater than or equal to 5, it is judged as suspected cracks;

[0123] 2) When the number of C2 crack disease discrimination results is less than 5, but there are more than 2 results with a confidence level greater than 0.8 in the discrimination results, it is judged as a suspected crack. If there are less than 2 results with a confidence level greater than 0.8 in the discrimination results, the discrimination results are filtered.

[0124] The present invention selected 3 expressways in Jiangsu Province to be overhauled, and used the method in Example 2 to identify crack diseases inside the road surface. A total of 47 clear crack diseases and 137 suspected crack diseases were identified on the 3 typical expressways, and the coincidence rate between the identification results and the manual discrimination results was more than 90%. On this basis, the accuracy of the identification results was analyzed by manual review. Among the 47 clear crack diseases, only 2 crack diseases were not manually recognized, and the identification accuracy was about 96%; among the 137 suspected crack diseases, 6 crack diseases were not manually recognized, and the identification accuracy was more than 95%.

[0125] Table 2 Human approval of the recognition results of each highway

[0126]

[0127] In summary, the method of the present invention can achieve an accuracy rate of 96% in the recognition of cracks inside the pavement. If human assistance is used to review suspected crack diseases, the accuracy rate can reach 99%. Therefore, the method can improve the recognition accuracy of crack diseases while ensuring that no identification results are missed, and can also improve the recognition efficiency of crack diseases to a certain extent.

[0128] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for identifying crack damage inside a pavement based on damage classification, characterized in that: The method specifically comprises: Step (1): pre-process the detected radar data to make the characteristics of cracks inside the pavement more obvious; Step (2): exporting the preprocessed radar slice map data and its corresponding electromagnetic signal data; Step (3): Use the radar slice map data derived from step (2) to establish a crack disease data set, and label the data set, where crack diseases with obvious crack characteristics are labeled as C1, and crack diseases with relatively weak characteristics are labeled as C2; Step (4): Bring the data set obtained in step (3) into the YOLO deep learning network model for training to obtain a network weight model; Step (5): export the radar data to be identified of the same size and type as in step (2), bring the obtained slice map data into the network weight model for training, and obtain the recognition result of the pavement cracks; Step (6): Analyze the electromagnetic signal data derived in step (5) to filter out crack features of bridge structure and internal foreign matter; Step (7): further filtering the filtered recognition results and classifying the crack disease characteristics; The specific marking method in step (3) is as follows: Normalize the electromagnetic signal amplitude at the corresponding depth of the radar slice map data derived in step (2). First, normalize the electromagnetic signal amplitude at the crack disease location and its surroundings, calculate the variance of each electromagnetic signal amplitude, and determine the difference threshold of C1 and C2 crack diseases to achieve the classification of the degree of obviousness of crack disease characteristics; Among them, the normalization formula of the electromagnetic signal amplitude is: ； In the formula, ——The normalized value of the amplitude of the electromagnetic wave in the i-th channel; ——the amplitude of the electromagnetic wave of the i-th channel; ——The error between the maximum and minimum amplitudes of electromagnetic waves around the crack disease; Among them, the calculation formula for the variance of the electromagnetic signal amplitude at the crack disease location and the electromagnetic signal amplitude around it is: ； In the formula, ——The average value of the normalized electromagnetic signal amplitude within the range; ——Amplitude variance of the normalized electromagnetic signal within the range; Among them, C1 type crack disease Greater than or equal to 0.7, C2 type crack disease Between 0.3 and 0.7; The specific method of step (7) is as follows: S1: Determination of the continuity of the recognition results in the depth direction, that is, if the overlap area of ​​the recognition frames at the same position at adjacent depths is more than 50%, it is considered to be the same crack disease, otherwise it is two or more crack diseases; S2 classifies and filters based on the number, confidence, and crack type of crack identification results at the same location; The crack classification and filtering rules mainly include: In the crack disease identification results at the same position, when the number of crack identification results is less than 3, the identification results are directly filtered; when the number of crack identification results is greater than or equal to 3, the judgment of step S22 is performed; In the crack disease identification results at the same position of S22, if there is a C1 type crack disease, the following identification is performed: 1) When the number of identification results of C1 crack damage is greater than or equal to 5, it is determined to be a clear crack; the clear crack means that the crack clearly exists, and the identification result does not need to be manually reviewed later; 2) When the number of identification results of C1 cracks is less than 5, if the confidence of the identified C1 crack damage is greater than or equal to 0.7, it is determined to be a clear crack; 3) When the number of identified C1 cracks is less than 5, if the identified C1 crack disease has a confidence level less than 0.7, the following judgment is made: a If the confidence level of C1 crack disease is greater than or equal to 0.5, and there is no C2 crack disease, it is judged as suspected crack. The suspected crack means that the crack may exist, and the identification result needs to be manually reviewed and confirmed later; b If the confidence level of C1 crack disease is greater than or equal to 0.5, and there are 4 or more C2 crack diseases, it is confirmed as a clear crack disease; c If the confidence level of C1 crack disease is greater than or equal to 0.5, and there are less than 4 C2 crack diseases, if the confidence level of C2 crack diseases is greater than 0.8, it is a clear crack disease. If the confidence level of the identified C2 crack disease is less than or equal to 0.8, it is judged as a suspected crack; d. If the confidence level of the C1 crack disease is less than 0.5, and if the number of C2 crack diseases is greater than 2, it is judged as a suspected crack disease. If there are only 2 or less C2 crack diseases, the judgment result is filtered; S23 When only C2 crack disease exists in the crack disease identification results at the same location, then 1) When the number of C2 crack disease identification results is greater than or equal to 5, it is judged as suspected cracks; 2) When the number of C2 crack disease discrimination results is less than 5, but there are more than 2 results with a confidence level greater than 0.8 in the discrimination results, it is judged as a suspected crack. If there are less than 2 results with a confidence level greater than 0.8 in the discrimination results, the discrimination results are filtered.

2. The method for identifying internal crack damage in pavement based on damage classification according to claim 1 is characterized in that: The filtering method of step (6) is: firstly perform Fourier transform on the collected electromagnetic signal data, then extract the time-frequency domain eigenvalues ​​of the collected electromagnetic signal in turn, and filter the crack characteristics of the bridge structure and internal foreign matter by determining the threshold of the time-frequency domain eigenvalue.

3. The method for identifying pavement internal crack damage based on damage classification according to claim 2 is characterized in that: The specific filtering steps of step (6) are as follows: Since the single-channel electromagnetic signal is a discrete signal with limited sampling data points, discrete Fourier transform is required, and the formula is: ； Where: ——The amplitude corresponding to the k-frequency component; x n ——is a discrete time signal; N——the number of sampling data points; Among them, the time-frequency domain characteristic signals used include: Energy amplitude variance: the second-order center distance of the energy amplitude of a single radar channel Right now ; Energy amplitude waveform attenuation degree: waveform factor of the energy amplitude of a single radar channel Right now ; Among them, xi represents the i-th sampling point of a certain wave, Indicates the mean amplitude of a single-channel wave; make , when the P value is greater than 5, it is a bridge or a foreign object.

4. The method for identifying pavement internal crack damage based on damage classification according to claim 1 is characterized in that: After the dataset is labeled in step (3), data augmentation technology is used to make the data volume of C1 and C2 in the dataset the same or substantially the same.

5. The method for identifying pavement internal crack damage based on damage classification according to claim 4 is characterized in that: Data augmentation techniques refer to flipping, rotation, scaling, random cropping, and / or color jittering.

6. The method for identifying pavement internal crack damage based on damage classification according to claim 1 is characterized in that: The preprocessing method in step (1) includes: zero bias removal, zero point adjustment and data gain, and the gain coefficient of the waveform is adjusted; specifically, the parameter of zero bias removal is 90%; the zero point adjustment selects the first positive maximum positive amplitude of the direct wave of each channel; the data gain selects the compensation gain, the gain coefficient is controlled between 1.0-2.0, and the data chromatogram uniformly uses the grayscale chromatogram.

7. The method for identifying pavement internal crack damage based on damage classification according to claim 1 is characterized in that: The size of the radar slice image exported in step (2) is selected as 400×100, the number of electromagnetic signal channels is set to 400, and the overlap of each adjacent slice is 5%, that is, 20 channels of electromagnetic signal data.

Images