This application is the national stage entry of International Application No. PCT/CN2020/120601, filed on Oct. 13, 2020, which is based upon and claims priority to Chinese Patent Application No. 202010140645.0 filed on Mar. 3, 2020, the entire contents of which are incorporated herein by reference.
The present invention belongs to the field of road maintenance, and relates to detection of moisture damage of asphalt pavements, in particular to a method for constructing a recognition model of a moisture damage of an asphalt pavement, and a method and system for recognizing the moisture damage of the asphalt pavement.
Ground penetrating radar (Ground Penetrating Radar, GPR) is an instrument for detecting a condition under the surface of earth and imaging by radar impulse waves, and its principle is detecting electromagnetic contrasts in a medium by emitting and receiving high frequency electromagnetic (EM) waves via an antenna. GPR uses a high frequency wireless radio wave which is usually polarized, the EM wave is emitted under the surface of the earth, and when the EM wave strikes an object buried under the surface of the earth or reaches a boundary with variable dielectric constants, a reflected wave received by the antenna will record a signal difference of a reflection echo. As GPR has the characteristics of high precision, high efficiency and non-destructive detection, GPR has already been applied to road maintenance, for example, GPR has been widely applied to cavitation of a tunnel substrate, pavement cavitation, recognition of dowel steel or a reinforcing bar of a building or a bridge deck slab and underground pipelines. Due to the difficulty in GPR data explanation, current GPR analysis depends on human analysis, which is both in time-wasting and labor-wasting. In particular, as a result of long highway mileage, it is hard to meet an intelligent detection demand for pavement maintenance with current artificial experience recognition method. Therefore, it is an urgent need for intelligent analysis and recognition method of the GPR signal.
Taking a common pavement defect as an example, the moisture damage is one of major reasons which cause early damage of the asphalt pavement. The moisture damage will damage strength and durability of an asphalt aggregate, which causes obvious defects of the asphalt pavement such as looseness, longitudinal and transverse cracks, upheavals, slurry pumping or whitening, pot holes and peeling of an asphalt layer, such that the service life of the pavement is lowered and the traffic safety is affected. Asphalt pavements in China have surpassed 4000 thousand km, such that it is of great significance to guarantee the traffic safety by quick detection of the moisture damage.
With development of computer technology and artificial intelligence (for example, deep learning), target detection based on GPR image has been applied to ground penetrating radar, for example, reinforcing steel bar detection with a hyperbola characteristic. Recognition based on GPR image needs an intense measurement to obtain a high quality image, which cannot meet the defect detection demand with traffic speed vehicle-mounted GPR system. Aiming at the problem, the patent application (a method for recognizing a moisture damage based on a time-frequency statistic characteristic of a ground penetrating signal 201910100046.3) is provided by the writer. A method for detecting the moisture damage based on a GPR signal is realized by statistical analysis through parameters of a time domain and a frequency domain, and meanwhile, the moisture damage and the normal pavement are different in time domain and frequency domain.
Wavelet analysis has an ability of representing a local characteristic of the signal in time and frequency domains and has the characteristic that by self-adaption, a low frequency signal changes slowly and a high frequency signal changes quickly, which can suppress noises effectively. Continuous wavelet transform (CWT) can represent a frequency component and a time interval corresponding to moisture damage and can explain a transmission characteristic of the GPR signal energy in the asphalt pavement. By means of its time-frequency characteristic diagram, the method has rich characteristics and further can realize automatic detection of the moisture damage more accurately.
Aiming at deficiencies and defects in the prior art, the present invention aims to provide a method for constructing a recognition model of a moisture damage of an asphalt pavement, and a method and system for recognizing the moisture damage of the asphalt pavement. The technical problem that it is hard to detect the moisture damage of the asphalt pavement automatically in the prior art is solved.
In order to solve the technical problem, the present invention adopts a technical scheme as follows:
The present invention further has the technical characteristics that
Further, in the field data acquisition process in the step 1, a sampling frequency is 10-20 times of a main frequency of an antenna.
Further, pre-processing procedure of GPR data includes removing direct current (DC) offset algorithm (to subtract the DC drift), static correction algorithm (to cut the air layer and correct the layer of asphalt pavement), a background removal algorithm, a band-pass filtering algorithm and a sliding average algorithm.
Further, in the step 5, a dimension size of the input data is 28*28 and the classification label is a number, wherein the normal pavement is 0, the bridge joint is 1 and the moisture damage is 2.
Further, the present invention provides a method for constructing a recognition model of a moisture damage of an asphalt pavement, including the following steps:
Further, the present invention further provides a system for recognizing a moisture damage of an asphalt pavement, including a data acquisition and pre-processing module and the recognition model according to claim 1,
Compared with the prior art, the present invention has the benefits that
Further description of specific embodiments of the present invention in detail will be made below in combination with drawings.
Specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the specific embodiments below and equivalent transformations made based on the technical scheme of the application shall fall within the scope of protection of the present invention.
The number of samples (A-Scan sample, single radar trace, or waveform) selected in the embodiment is 22453, wherein the number of the moisture damage samples is 8215, the number of the normal pavement samples is 8215 and the number of the bridge joint samples is 6023.
The embodiment provides a method for recognizing a moisture damage of an asphalt pavement as shown in the
S1, a GPR field survey data set is pre-processed to obtain an initial data set with a moisture damage, a bridge joint and a normal pavement, wherein in the field data collection process, a moisture damaged region of the pavement is marked in GPR acquisition software, and in the embodiment, a region with slurry pumping (stripping) or whitening in the pavement is marked;
In the field data acquisition process, a sampling frequency is 10-20 times of a central frequency of an antenna.
Pre-processing procedure of GPR data includes removing direct current (DC) offset algorithm (to subtract the DC drift), static correction algorithm (to cut the air layer and correct the layer of asphalt pavement), a background removal algorithm, a band-pass filtering algorithm and a sliding average algorithm.
in the step 1, the contrast of the GPR image is ranged from 1.2 to 1.6, preferably 1.4 in the embodiment.
A process of acquiring the GPR data set of the moisture damage: when GPR antenna arrived at a observed whitening area, a marker of corresponding GPR traces will be recorded by data acquisition software, and main characteristics of the moisture damage are determined by plenty of investigation of living examples:
The bottom image in the
S2, continuous wavelet transform is performed on the initial data set by using continuous wavelet transform, and an amplitude of wavelet transform taken construct a first time-frequency image data set;
A continuous wavelet transform formula is as follows:
When |a| is smaller than 1, the mother wavelet j* (a function with a continuous property in the time domain and the frequency domain) is compressed and has a small supporting degree in a timer shaft, and corresponds to a high frequency because the mother wavelet is narrowed and changes quickly. On the contrary, when |a| is greater than 1, the mother wavelet is widened and changes slowly, and corresponds to a low frequency. The scale factor α represents stretching related to frequency.
In the
After continuous wavelet transform (CWT), the amplitude-frequency size of the array of the sample is 51*237 and the first time-frequency data set is constructed based on original samples.
S3, a image in the first time-frequency image set is filtered to obtain a second time-frequency image so as to construct a second first time-frequency image data set;
it can be known from contrastive analysis, the energy of the three targets (Moisture damage-Moisture, Normal pavement-Normal and bridge joint-Joint) are concentrated at 0.4-4 GHz in frequency domain and 0-3 ns in the time domains a.
The time-frequency data is filtered by using Morse wavelet, and filtering parameters are data within a range of frequency domain [0.4-4 GHz] and time domain [0-3 ns] (corresponding to a sample at a A-scan sampling point [1,100]). After filtering processing, the amplitude-frequency size of the array of each sample after continuous wavelet transform (CWT) is reduced from 51*237 to 15*100, which is defined as the second time-frequency image set.
S4, normalization processing is performed on the image in the second time-frequency image set to obtain a third time-frequency image so as to construct a third time-frequency image set, and a moisture damage classification label is annotated for the image of the third time-frequency image set; and
Normalization processing specifically includes:
The maximum value in the matrix A is defined, and if elements in the matrix exceed the maximum value, the value is assigned as the maximum value maxCWT and otherwise, the original value is reserved, and a calculating method of some point A(i,j) in the matrix specifically includes:
The defined matrix A is normalized and is mapped to the range of [0, 255], and the calculating method is as follows:
A(i,j)=k×A(i,j)/MaxCWT, where i∈[1,m],j∈[1,n] (4)
Normalization of the amplitude after continuous transform can be achieved according to the steps.
It is found by researches that image with different resolutions are different in accuracy in the recognition model, and the picture resolution affect a model recognition result directly. Therefore, as a preferred scheme, in the embodiment, the resolution of the third time-frequency image is adjusted: the third time-frequency image data set is zoomed to 28*28 directly to obtain a fourth time-frequency image data set;
As shown in the
The GPRMCNN deep learning model adopting the fourth time-frequency image set is divided into a training set, a test set and a verification set with the distribution proportion being 60%, 20% and 20%. A specific model training method includes training the designed mixed deep learning model by using a TL(Transfer Learning), wherein parameters are set as follows: ‘InitialLearnRate’:0.005, ‘MaxEpochs’:15, ValidationFrequency’:30.
The classification label is a number, wherein the normal pavement is 0, the bridge joint is 1 and the moisture damage is 2.
The deep learning model uses a classification precision index to measure performance of the model.
The classification result is drawn in an overlapped manner on the GPR image of pavement, the result is as shown in
The comparative example provides the method for detecting the moisture damage of the asphalt pavement. Other steps of the method are same as those in the embodiment 1 and the difference is merely that filtering and normalization processing are not performed in the step 2. The initial data spectra are as shown in
The precision of model test is as shown in
The comparative example provides the method for detecting the moisture damage of the asphalt pavement. The method is as same as the embodiment I in step 1. The method adopting a conventional time-frequency characteristic extraction method constructs an ANN (artificial neural network, belonging to a machine learning method), the specific steps being as same as those of the method for recognizing the moisture damage based on the time-frequency statistics characteristic of the ground penetrating signal (the Chinese patent application number is 201910100046.3).
The precision of model test is as same as ANN as shown in
The comparative example provides the method for detecting the moisture damage of the asphalt pavement. Other steps of the method are same as those in the embodiment 1 and the difference is merely that when the time-frequency image is zoomed, the time-frequency image is not stored as a 28*28 matrix but a 224*224 RGB picture; then the obtained time-frequency image is input to the Resenet18 directly, and the training parameters are set as ‘Momentum’, 0.9 and MiniBatchSize’, 64, respectively.
The precision of the test result is as same as Resenet18 as shown in
Contrastive analysis is performed on the embodiment 1, the comparative example 2 and the comparative example 3.
Number | Date | Country | Kind |
---|---|---|---|
202010140645.0 | Mar 2020 | CN | national |
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/CN2020/120601 | 10/13/2020 | WO |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO2021/174857 | 9/10/2021 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
5835053 | Davis | Nov 1998 | A |
20190301109 | Roberts | Oct 2019 | A1 |
20220276374 | Zhang | Sep 2022 | A1 |
Number | Date | Country |
---|---|---|
107238821 | Oct 2017 | CN |
109782274 | May 2019 | CN |
110458028 | Nov 2019 | CN |
110552281 | Dec 2019 | CN |
110736985 | Jan 2020 | CN |
111476088 | Jul 2020 | CN |
2985309 | Jul 2016 | EP |
Number | Date | Country | |
---|---|---|---|
20220276182 A1 | Sep 2022 | US |