A friction prediction method based on multi-scale decomposition of asphalt pavement texture

By performing multi-scale decomposition and denoising processing on asphalt pavement texture, a friction prediction model is constructed and integrated into the detection vehicle for scanning, the problems of poor repetition and difficulty in automatic prediction of road surface anti-slip performance analysis in the existing technology are solved, and the asphalt is realized Accurate identification and large-scale measurement of the anti-slip performance of the road surface.

CN118229661BActive Publication Date: 2024-11-19SOUTHWEST JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410435846.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-11
Publication Date
2024-11-19
Estimated Expiration
2044-04-11

AI Technical Summary

Technical Problem

The prior art has poor repeatability, low correlation of equipment indicators, inability to achieve continuous friction prediction, great influence on test vehicles and conditions, and cannot automatically and accurately predict asphalt road friction performance, resulting in difficulty in judging anti-slip performance.

Method used

A multi-scale decomposition method based on asphalt pavement texture is adopted. By obtaining two-dimensional texture data sets and friction data sets, denoising and multi-scale decomposition are carried out, friction prediction models are constructed, and integrated into the detection vehicle for scanning to obtain friction prediction results.

Benefits of technology

It realizes accurate identification and large-scale measurement of the anti-slip performance of asphalt pavement, improves the accuracy of pavement texture data, and can effectively predict the anti-slip performance of asphalt pavement.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118229661B_ABST
    Figure CN118229661B_ABST
Patent Text Reader

Abstract

The present invention discloses a friction prediction method based on multi-scale decomposition of asphalt pavement texture. The method comprises the following steps: obtaining a two-dimensional texture data set and a friction data set of an asphalt pavement; performing denoising and multi-scale decomposition processing on the two-dimensional texture data set of the asphalt pavement to obtain a multi-scale two-dimensional texture data set of the asphalt pavement; constructing a friction prediction model, and training the friction prediction model using the friction data set of the asphalt pavement and the multi-scale two-dimensional texture data set of the asphalt pavement to obtain a trained friction prediction model, and integrating the trained friction prediction model into a detection vehicle; scanning the asphalt pavement using the detection vehicle integrated with the trained friction prediction model to obtain the friction prediction result of the asphalt pavement. The present invention can automatically and accurately predict the friction of any asphalt pavement to determine whether the asphalt pavement meets the anti-skid requirements, thereby avoiding the time-consuming and road damage problems of manual measurement.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of pavement anti-skid performance analysis, and in particular to a friction prediction method based on multi-scale decomposition of asphalt pavement texture. Background Art

[0002] Anti-skid performance is one of the important properties of the road surface, which refers to the force generated when the tire rotates and slides along the road surface. Insufficient anti-skid performance will lead to a higher risk of collision, such as skidding and rear-end collision.

[0003] Traditional methods for direct measurement of road friction have problems with poor repeatability and low correlation between indicators of various devices. In addition, these methods can only achieve point-based friction prediction, and the results are discontinuous. Moreover, these methods are greatly affected by the test vehicle and experimental conditions, require professional technicians to test, cannot be used for large-scale road friction measurement, and will cause road damage. Although finite element simulation models can accurately estimate road friction, they require a lot of computing power and time.

[0004] Currently, the contact mechanical behavior of vehicle loads on elastic pavement is complex and difficult to analyze, so the role of pavement texture on anti-skid performance is also difficult to examine.

[0005] In summary, the existing methods have the problems of time-consuming and damaging the road surface due to manual measurement, and cannot automatically and accurately predict the friction performance of any asphalt pavement to determine whether the asphalt pavement section meets the anti-skid requirements. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a friction prediction method based on multi-scale decomposition of asphalt pavement texture, which can effectively extract the different effects of different asphalt pavement textures on anti-skid performance, thereby effectively predicting the anti-skid performance of asphalt pavement.

[0007] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0008] A friction prediction method based on multi-scale decomposition of asphalt pavement texture includes the following steps:

[0009] S1. Obtaining a two-dimensional texture dataset of an asphalt pavement and a friction dataset of an asphalt pavement;

[0010] S2, performing denoising and multi-scale decomposition processing on the asphalt pavement two-dimensional texture dataset in step S1 to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement;

[0011] S3, constructing a friction prediction model, and using the asphalt pavement friction dataset in step S1 and the asphalt pavement multi-scale two-dimensional texture dataset in step S2 to train the friction prediction model, obtaining the trained friction prediction model, and integrating the trained friction prediction model into the inspection vehicle;

[0012] S4. Scan the asphalt pavement section using the inspection vehicle integrated with the trained friction prediction model in step S3 to obtain the friction prediction result of the asphalt pavement.

[0013] Furthermore, step S1 includes the following steps:

[0014] S11, determining the test road section measuring points, and using a detection vehicle equipped with a high-speed laser scanner to scan the asphalt pavement texture at the test road section measuring points to obtain a two-dimensional texture data set of the asphalt pavement;

[0015] Each individual data of the asphalt pavement two-dimensional texture data set is 512×512 two-dimensional elevation point data, and the horizontal spacing between the two-dimensional elevation points is 0.05 mm;

[0016] S12. Use a dynamic friction tester to measure the friction performance of the test road section measuring points in step S11 to obtain an asphalt pavement friction data set.

[0017] Further, step S2 includes the following steps:

[0018] S21, performing denoising processing on the asphalt pavement two-dimensional texture dataset in step S1 to obtain a denoised asphalt pavement two-dimensional texture dataset;

[0019] S22. Use discrete Fourier transform to perform multi-scale decomposition processing on the asphalt pavement two-dimensional texture dataset after denoising in step S21, and obtain texture information with wavelengths of 0.1-0.5 mm, 0.5-4 mm, 4 mm-16 mm, 16 mm-50 mm and greater than 50 mm, so as to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement.

[0020] Further, step S21 includes the following steps:

[0021] S211, calculating the mean value of the two-dimensional texture data in the asphalt pavement two-dimensional texture data set, replacing the two-dimensional elevation point values ​​exceeding 2000 or less than -2000 with the mean value of the two-dimensional texture data, and obtaining an optimized asphalt pavement two-dimensional texture data set;

[0022] S212, using a two-dimensional Fourier transform to decompose the optimized asphalt pavement two-dimensional texture data set in step S211 into asphalt pavement macro texture data and asphalt pavement micro texture data;

[0023] S213, performing bilateral filtering and denoising on the asphalt pavement macro texture data in step S212 to obtain optimized asphalt pavement macro texture data, which is expressed as:

[0024]

[0025] Where: H(x, y) is the pixel value in the optimized macro texture data, and the optimized macro texture data is the set of all H(x, y), W p is the normalized weight, (x, y) is the coordinate of the current pixel, (i, j) is the coordinate of other pixels in the neighborhood, Ω is the pixel neighborhood, G s is the spatial weight function, P·|| is the norm 2, G r is the grayscale weight function, I(x,y) is the grayscale value of the current pixel, and I(i,j) is the grayscale value of other pixels in the neighborhood;

[0026] S214, performing MAD method outlier denoising on the asphalt pavement micro texture data in step S212 to obtain first optimized asphalt pavement micro texture data, expressed as:

[0027]

[0028] in: is the pixel value of the first optimized micro texture data, and the first optimized micro texture data is all A collection of, M is the median of the micro texture data, X i is the pixel value in the micro texture data, k is the threshold multiple, MAD is the MAD method outlier of the micro texture data, otherwise is other cases;

[0029] S215, performing wavelet transform denoising on the first optimized asphalt pavement micro-texture data in step S214 to obtain second optimized asphalt pavement micro-texture data, which is expressed as:

[0030]

[0031] in: is an image of the second optimized micro texture data, the second optimized micro texture data is all A set of, a is the scale parameter, b is the location parameter, is the wavelet coefficient after denoising, ψ a,b (x,y) is the wavelet basis function;

[0032] S216: reconstruct the optimized asphalt pavement macro texture data in step S213 and the second optimized asphalt pavement micro texture data in step S215 to obtain a denoised asphalt pavement two-dimensional texture data set.

[0033] Further, step S22 includes the following steps:

[0034] S221, converting the denoised asphalt pavement two-dimensional texture data set in step S21 into the frequency domain by discrete Fourier transform to obtain the frequency component of the two-dimensional texture of the asphalt pavement, expressed as:

[0035]

[0036] Where: X1[k,l] is the frequency component of the two-dimensional texture of the asphalt road surface, [k,l] is the horizontal and vertical coordinates of the spectrum domain, M is the number of rows, N is the number of columns, m1[x,y] is the pixel value of the two-dimensional texture in the denoised two-dimensional texture data set, [x,y] is the horizontal and vertical coordinates;

[0037] S222, sequentially retaining the texture information of the target wavelengths of 0.1-0.5 mm, 0.5-4 mm, 4 mm-16 mm, 16 mm-50 mm and greater than 50 mm in the frequency components of the two-dimensional texture of the asphalt pavement in step S221, and filtering out the clutter outside the target wavelength to obtain the frequency components of the two-dimensional texture of the asphalt pavement after the clutter is filtered out;

[0038] S223, converting the frequency components of the asphalt pavement two-dimensional texture after the clutter is filtered out in step S222 into the asphalt pavement two-dimensional texture through inverse Fourier transform to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement, which is expressed as:

[0039]

[0040] Among them: m2[x,y] is the pixel value of the multi-scale two-dimensional texture of the asphalt pavement, the multi-scale two-dimensional texture dataset of the asphalt pavement is the set of all m2[x,y], and X2[k,l] is the frequency component of the two-dimensional texture of the asphalt pavement after filtering out clutter.

[0041] Further, step S3 includes the following steps:

[0042] S31, matching the asphalt pavement friction dataset in step S1 with the asphalt pavement multi-scale two-dimensional texture dataset in step S2, and dividing the matched datasets into a training set, a test set, and a validation set in a ratio of 8:1:1;

[0043] S32, constructing a friction prediction model including a residual network for extracting information of different scales, an aggregation network for fusing information of different scales, and an output module for obtaining a friction prediction result of an asphalt pavement;

[0044] S33. Use the training set, test set and validation set in step S31 to train, test and validate the friction prediction model in step S32, obtain the trained friction prediction model, and integrate the trained friction prediction model into the inspection vehicle.

[0045] Further, in step S32, constructing a residual network for extracting information of different scales includes the following steps:

[0046] A1. The multi-scale two-dimensional texture data of asphalt pavement with a size of 512×512×1 is used as the first input data of the friction prediction model. It first passes through a convolution kernel with a size of 7×7, a number of channels of 64, and a stride of 2, and then passes through normalization and ReLU activation function, and then passes through the maximum pooling with a stride of 2 to obtain the first feature data with a size of 128×128×64;

[0047] A2, taking the first feature data of size 128×128×64 in step A1 as the second input data, and obtaining the second feature data of size 128×128×64 through three first residual modules; the first residual module first passes through a convolution kernel of size 3×3, number of channels 64, and stride 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 64, and stride 1 and standardization, and then adds it to the second input data, and then passes through the Relu activation function;

[0048] A3, taking the second feature data of size 128×128×64 in step A2 as the third input data, first passing through a second residual module, and then passing through three third residual modules to obtain the third feature data of size 64×64×128; the second residual module first passes through a convolution kernel of size 3×3, number of channels 128, and stride 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 128, stride 1 and standardization, and then adds it to the third input data, and then passes through the Relu activation function; the third residual module first passes through a convolution kernel of size 3×3, number of channels 128, and stride 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 128, stride 1 and standardization, and then adds it to the third input data, and then passes through the Relu activation function;

[0049] A4, taking the third feature data of size 64×64×128 in step A3 as the fourth input data, first passing through a fourth residual module, and then passing through five fifth residual modules to obtain fourth feature data of size 32×32×256; the fourth residual module first passes through a convolution kernel of size 3×3, number of channels 256, and stride 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 256, stride 1 and standardization, and then adds it to the fourth input data, and then passes through the Relu activation function; the fifth residual module first passes through a convolution kernel of size 3×3, number of channels 256, and stride 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 256, stride 1 and standardization, and then adds it to the fourth input data, and then passes through the Relu activation function;

[0050] A5. Use the fourth feature data of size 32×32×256 in step A4 as the fifth input data, first pass through a sixth residual module, and then pass through two seventh residual modules to obtain the fifth feature data of size 16×16×512; the sixth residual module first passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1 and standardization, and then adds it to the fifth input data, and then passes through the Relu activation function; the seventh residual module first passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1 and standardization, and then adds it to the fifth input data, and then passes through the Relu activation function.

[0051] Furthermore, in step S32, constructing an aggregation network for fusing information of different scales includes the following steps:

[0052] B1. Concatenate the six fifth feature data with a size of 16×16×512 output by the residual network to obtain the sixth feature data with a size of 16×16×3072;

[0053] B2, the sixth feature data of size 16×16×3072 in step B1 is first standardized, and then passed through a convolution kernel of size 3×3, number of channels 512, stride 1 and Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the seventh feature data of size 8×8×512;

[0054] B3, the seventh feature data with a size of 8×8×512 in step B2 is first passed through a convolution kernel with a size of 3×3, a channel number of 256, and a stride of 1 and a Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the eighth feature data with a size of 4×4×256;

[0055] B4, the eighth feature data with a size of 4×4×256 in step B3 is first passed through a convolution kernel with a size of 3×3, a number of channels of 128, and a stride of 1 and a Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the ninth feature data with a size of 2×2×128;

[0056] B5, the ninth feature data of size 2×2×128 in step B4 is first passed through a convolution kernel of size 3×3, number of channels 64, stride 1 and Relu activation function, and then through maximum pooling with stride 2 to obtain the tenth feature data of size 1×1×64;

[0057] B6. The tenth feature data with a size of 1×1×64 in step B5 is passed through a convolution kernel with a size of 1×1, a number of channels of 32, and a step size of 1 and a Relu activation function to obtain an eleventh feature data with a size of 1×1×32.

[0058] Furthermore, in step B1, the 6 fifth feature data with a size of 16×16×512 are texture data of five different scales with wavelengths of 0.1-0.5mm, 0.5-4mm, 4mm-16mm, 16mm-50mm and greater than 50mm in the asphalt pavement multi-scale two-dimensional texture dataset and the texture data in the asphalt pavement two-dimensional texture dataset, which are output by the residual network respectively.

[0059] Further, in step S32, constructing an output module for obtaining the friction prediction result of the asphalt pavement includes the following steps:

[0060] C1. Flatten the eleventh feature data of size 1×1×32 output by the aggregation network to obtain 32 twelfth feature data;

[0061] C2. The 32 twelfth feature data in step C1 are first passed through a fully connected layer with 128 nodes, then through a Relu activation function, and then through a fully connected layer with 1 node to obtain the friction prediction result of the asphalt pavement.

[0062] The beneficial effects of the present invention are:

[0063] (1) The present invention uses a detection vehicle equipped with a high-speed laser scanner to obtain a two-dimensional texture data set of asphalt pavement at a test section measuring point, and uses an innovative denoising method to denoise the collected two-dimensional texture data set of asphalt pavement, thereby improving the accuracy of the two-dimensional texture data of the asphalt pavement;

[0064] (2) The detection vehicle of the present invention can accurately identify the friction of any asphalt pavement by using the integrated friction prediction model to determine its anti-skid performance, and can be used for large-scale road friction measurement;

[0065] (3) The present invention uses Fourier transform to extract road surface textures of different scales and analyzes the different effects of road surface textures of different scales on the friction of the asphalt road surface, so that the friction prediction results of the asphalt road surface are more comprehensive and accurate, thereby being able to more effectively predict the anti-skid performance of the asphalt road surface. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A schematic diagram of the process of friction prediction method based on multi-scale decomposition of asphalt pavement texture;

[0067] Figure 2 Schematic diagram of multi-scale decomposition of two-dimensional texture data of asphalt pavement;

[0068] Figure 3 This is a schematic diagram of the friction prediction model structure;

[0069] Figure 4 Schematic diagram of the loss function change during the friction prediction model training process;

[0070] Figure 5 Schematic diagram of the distribution of prediction results of the friction prediction model. DETAILED DESCRIPTION

[0071] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0072] like Figure 1 As shown, a friction prediction method based on multi-scale decomposition of asphalt pavement texture includes steps S1-S4, which are as follows:

[0073] S1. Obtain asphalt pavement two-dimensional texture dataset and asphalt pavement friction dataset.

[0074] In an optional embodiment of the present invention, the present invention uses a detection vehicle equipped with a high-speed laser scanner to obtain a two-dimensional texture data set of an asphalt pavement, and uses a dynamic friction tester to obtain a friction data set of an asphalt pavement.

[0075] Step S1 includes the following steps:

[0076] S11, determining the test road section measuring points, and using a detection vehicle equipped with a high-speed laser scanner to scan the asphalt pavement texture at the test road section measuring points to obtain a two-dimensional texture data set of the asphalt pavement;

[0077] Each individual data of the asphalt pavement two-dimensional texture data set is 512×512 two-dimensional elevation point data, and the horizontal spacing between the two-dimensional elevation points is 0.05 mm.

[0078] S12. Use a dynamic friction tester to measure the friction performance of the test road section measuring points in step S11 to obtain an asphalt pavement friction data set.

[0079] S2. Perform denoising and multi-scale decomposition processing on the asphalt pavement two-dimensional texture dataset in step S1 to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement.

[0080] In an optional embodiment of the present invention, the present invention performs denoising on the asphalt pavement two-dimensional texture dataset, and uses discrete Fourier transform to perform multi-scale decomposition on the denoised two-dimensional texture dataset to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement.

[0081] Step S2 includes the following steps:

[0082] S21 , performing denoising processing on the asphalt pavement two-dimensional texture dataset in step S1 to obtain a denoised asphalt pavement two-dimensional texture dataset.

[0083] Step S21 includes the following steps:

[0084] S211. Calculate the mean of the two-dimensional texture data in the asphalt pavement two-dimensional texture data set, replace the two-dimensional elevation point values ​​greater than 2000 or less than -2000 with the mean of the two-dimensional texture data, and obtain an optimized asphalt pavement two-dimensional texture data set.

[0085] S212: Decompose the optimized asphalt pavement two-dimensional texture data set in step S211 into asphalt pavement macro texture data and asphalt pavement micro texture data using two-dimensional Fourier transform.

[0086] S213, performing bilateral filtering and denoising on the asphalt pavement macro texture data in step S212 to obtain optimized asphalt pavement macro texture data, which is expressed as:

[0087]

[0088] Where: H(x, y) is the pixel value in the optimized macro texture data, and the optimized macro texture data is the set of all H(x, y), W p is the normalized weight, which is used to ensure that the sum of all weights is 1, (x, y) is the coordinate of the current pixel, (i, j) is the coordinate of other pixels in the neighborhood, Ω is the pixel neighborhood, and G s is the spatial weight function, which calculates the weight according to the spatial distance between pixels, P·|| is the norm 2, G r is the grayscale weight function, which calculates the weight according to the difference between pixel values. I(x,y) is the grayscale value of the current pixel, and I(i,j) is the grayscale value of other pixels in the neighborhood.

[0089] S214, performing MAD method outlier denoising on the asphalt pavement micro texture data in step S212 to obtain first optimized asphalt pavement micro texture data, expressed as:

[0090] MAD=median(|X i -M|),

[0091]

[0092] in: is the pixel value of the first optimized micro texture data, and the first optimized micro texture data is all A collection of, M is the median of the micro texture data, X i is the pixel value in the micro texture data, k is the threshold multiple, which is 3 or 2.5, MAD is the MAD method outlier of the micro texture data, median() is the median in the calculation brackets, and otherwise is other cases.

[0093] S215, performing wavelet transform denoising on the first optimized asphalt pavement micro-texture data in step S214 to obtain second optimized asphalt pavement micro-texture data, which is expressed as:

[0094]

[0095]

[0096]

[0097] in: is an image of the second optimized micro texture data, the second optimized micro texture data is all A set of, a is the scale parameter, b is the location parameter, is the wavelet coefficient after denoising, ψ a,b(x, y) is the wavelet basis function, W(a, b) is the wavelet coefficient, x is the horizontal coordinate of the current pixel, y is the vertical coordinate of the current pixel, f(x, y) is the original pixel value of the image, sgn is the sign function used to extract the sign of the wavelet coefficient, max is the maximum value sign, and T is the wavelet coefficient threshold.

[0098] S216: reconstruct the optimized asphalt pavement macro texture data in step S213 and the second optimized asphalt pavement micro texture data in step S215 to obtain a denoised asphalt pavement two-dimensional texture data set.

[0099] S22. Use discrete Fourier transform to perform multi-scale decomposition processing on the asphalt pavement two-dimensional texture dataset after denoising in step S21, and obtain texture information with wavelengths of 0.1-0.5 mm, 0.5-4 mm, 4 mm-16 mm, 16 mm-50 mm and greater than 50 mm, so as to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement.

[0100] like Figure 2 As shown, the present invention provides a multi-scale decomposition schematic diagram of two-dimensional texture data of asphalt pavement.

[0101] Step S22 includes the following steps:

[0102] S221, converting the denoised asphalt pavement two-dimensional texture data set in step S21 into the frequency domain by discrete Fourier transform to obtain the frequency component of the two-dimensional texture of the asphalt pavement, expressed as:

[0103]

[0104] Among them: X1[k,l] is the frequency component of the two-dimensional texture of the asphalt road surface, [k,l] is the horizontal and vertical coordinates in the spectrum domain, M is the number of rows, N is the number of columns, m1[x,y] is the pixel value of the two-dimensional texture in the denoised two-dimensional texture dataset, and [x,y] is the horizontal and vertical coordinates.

[0105] S222. The texture information of the target wavelengths of 0.1-0.5 mm, 0.5-4 mm, 4 mm-16 mm, 16 mm-50 mm and greater than 50 mm in the frequency components of the two-dimensional texture of the asphalt pavement in step S221 is retained in turn, and clutter outside the target wavelength is filtered out to obtain the frequency components of the two-dimensional texture of the asphalt pavement after the clutter is filtered out.

[0106] S223, converting the frequency components of the asphalt pavement two-dimensional texture after the clutter is filtered out in step S222 into the asphalt pavement two-dimensional texture through inverse Fourier transform to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement, which is expressed as:

[0107]

[0108] Among them: m2[x,y] is the pixel value of the multi-scale two-dimensional texture of the asphalt pavement, the multi-scale two-dimensional texture dataset of the asphalt pavement is the set of all m2[x,y], and X2[k,l] is the frequency component of the two-dimensional texture of the asphalt pavement after filtering out clutter.

[0109] S3, constructing a friction prediction model, and using the asphalt pavement friction dataset in step S1 and the asphalt pavement multi-scale two-dimensional texture dataset in step S2 to train the friction prediction model, obtaining the trained friction prediction model, and integrating the trained friction prediction model into the inspection vehicle.

[0110] In an optional embodiment of the present invention, the present invention constructs a friction prediction model, such as Figure 3 Then, the present invention uses the asphalt pavement friction data set and the asphalt pavement multi-scale two-dimensional texture data set to train the friction prediction model, obtain the trained friction prediction model, and integrate the trained friction prediction model into the detection vehicle.

[0111] Step S3 includes the following steps:

[0112] S31, matching the asphalt pavement friction dataset in step S1 with the asphalt pavement multi-scale two-dimensional texture dataset in step S2, and dividing the matched datasets into a training set, a test set, and a validation set in a ratio of 8:1:1.

[0113] S32, constructing a friction prediction model including a residual network for extracting information of different scales, an aggregation network for fusing information of different scales, and an output module for obtaining friction prediction results of an asphalt pavement.

[0114] The present invention constructs a residual network for extracting information of different scales, comprising the following steps:

[0115] A1. The multi-scale two-dimensional texture data of the asphalt pavement with a size of 512×512×1 is used as the first input data of the friction prediction model. It first passes through a convolution kernel with a size of 7×7, 64 channels, and a step size of 2, and then passes through normalization and ReLU activation function, and then passes through the maximum pooling with a step size of 2 to obtain the first feature data with a size of 128×128×64.

[0116] A2. Use the first feature data of size 128×128×64 in step A1 as the second input data, and obtain the second feature data of size 128×128×64 through three first residual modules; the first residual module first passes through a convolution kernel of size 3×3, number of channels 64, and step size 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 64, and step size 1 and standardization, and then adds it to the second input data, and then passes through the Relu activation function.

[0117] A3. Use the second feature data of size 128×128×64 in step A2 as the third input data, first pass through a second residual module, and then pass through three third residual modules to obtain the third feature data of size 64×64×128; the second residual module first passes through a convolution kernel of size 3×3, number of channels 128, and step size 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 128, step size 1 and standardization, and then adds it to the third input data, and then passes through the Relu activation function; the third residual module first passes through a convolution kernel of size 3×3, number of channels 128, and step size 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 128, step size 1 and standardization, and then adds it to the third input data, and then passes through the Relu activation function.

[0118] A4. Use the third feature data of size 64×64×128 in step A3 as the fourth input data, first pass through a fourth residual module, and then pass through five fifth residual modules to obtain fourth feature data of size 32×32×256; the fourth residual module first passes through a convolution kernel of size 3×3, number of channels 256, and step size 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 256, step size 1 and standardization, and then adds it to the fourth input data, and then passes through the Relu activation function; the fifth residual module first passes through a convolution kernel of size 3×3, number of channels 256, and step size 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 256, step size 1 and standardization, and then adds it to the fourth input data, and then passes through the Relu activation function.

[0119] A5. Use the fourth feature data of size 32×32×256 in step A4 as the fifth input data, first pass through a sixth residual module, and then pass through two seventh residual modules to obtain the fifth feature data of size 16×16×512; the sixth residual module first passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1 and standardization, and then adds it to the fifth input data, and then passes through the Relu activation function; the seventh residual module first passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1 and standardization, and then adds it to the fifth input data, and then passes through the Relu activation function.

[0120] The present invention constructs an aggregation network for fusing information of different scales, comprising the following steps:

[0121] B1. Concatenate the six fifth feature data with a size of 16×16×512 output by the residual network to obtain the sixth feature data with a size of 16×16×3072.

[0122] The 6 fifth feature data with sizes of 16×16×512 are texture data of five different scales with wavelengths of 0.1~0.5mm, 0.5~4mm, 4mm~16mm, 16mm~50mm and greater than 50mm in the asphalt pavement multi-scale two-dimensional texture dataset, as well as the texture data in the asphalt pavement two-dimensional texture dataset, which are output by the residual network respectively.

[0123] B2. The sixth feature data with a size of 16×16×3072 in step B1 is first standardized, and then passed through a convolution kernel with a size of 3×3, a number of channels of 512, and a stride of 1 and a Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the seventh feature data with a size of 8×8×512.

[0124] B3. The seventh feature data with a size of 8×8×512 in step B2 is first passed through a convolution kernel with a size of 3×3, a number of channels of 256, and a stride of 1 and a Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the eighth feature data with a size of 4×4×256.

[0125] B4. The eighth feature data with a size of 4×4×256 in step B3 is first passed through a convolution kernel with a size of 3×3, a number of channels of 128, and a stride of 1 and a Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the ninth feature data with a size of 2×2×128.

[0126] B5. The ninth feature data of size 2×2×128 in step B4 is first passed through a convolution kernel of size 3×3, number of channels 64, stride 1 and a Relu activation function, and then through a maximum pooling with a stride 2 to obtain the tenth feature data of size 1×1×64.

[0127] B6. The tenth feature data with a size of 1×1×64 in step B5 is passed through a convolution kernel with a size of 1×1, a number of channels of 32, and a step size of 1 and a Relu activation function to obtain an eleventh feature data with a size of 1×1×32.

[0128] The present invention constructs an output module for obtaining the friction prediction result of an asphalt pavement, comprising the following steps:

[0129] C1. Flatten the eleventh feature data of size 1×1×32 output by the aggregation network to obtain 32 twelfth feature data.

[0130] C2. The 32 twelfth feature data in step C1 are first passed through a fully connected layer with 128 nodes, then through a Relu activation function, and then through a fully connected layer with 1 node to obtain the friction prediction result of the asphalt pavement.

[0131] S33. Use the training set, test set and validation set in step S31 to train, test and validate the friction prediction model in step S32, obtain the trained friction prediction model, and integrate the trained friction prediction model into the inspection vehicle.

[0132] Specifically, the present invention determines the optimal hyperparameters when verifying the friction prediction model, and the optimal hyperparameters finally verified are: initial learning rate 0.001, Adam optimizer, the number of data samples captured in one training Btach_size=8, and the number of iterations Epochs=100.

[0133] like Figure 4 As shown in Figure 2, the loss values ​​of the training set and validation set of the friction prediction model are getting lower and lower with the training iteration process. The final test set accuracy reaches 0.890, and the R-square value reaches 0.855. The distribution of friction prediction results of the friction prediction model is shown in Figure 2. Figure 5 As shown, this method has excellent prediction results.

[0134] S4. Scan the asphalt pavement using the inspection vehicle integrated with the trained friction prediction model in step S3 to obtain the friction prediction result of the asphalt pavement.

[0135] In an optional embodiment of the present invention, a detection vehicle driving an integrated trained friction prediction model is driven on any asphalt road surface for scanning, and the friction prediction result of the current asphalt road surface is obtained synchronously in real time, so as to more effectively predict the anti-skid performance of the asphalt road surface.

[0136] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A friction prediction method based on multi-scale decomposition of asphalt pavement texture, characterized in that: The following steps are involved: S1. Obtaining a two-dimensional texture dataset of an asphalt pavement and a friction dataset of an asphalt pavement; S2, performing denoising and multi-scale decomposition processing on the asphalt pavement two-dimensional texture dataset in step S1 to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement; Step S2 includes the following steps: S21, performing denoising processing on the asphalt pavement two-dimensional texture dataset in step S1 to obtain a denoised asphalt pavement two-dimensional texture dataset; Step S21 includes the following steps: S211, calculating the mean value of the two-dimensional texture data in the asphalt pavement two-dimensional texture data set, replacing the two-dimensional elevation point values ​​exceeding 2000 or less than -2000 with the mean value of the two-dimensional texture data, and obtaining an optimized asphalt pavement two-dimensional texture data set; S212, using a two-dimensional Fourier transform to decompose the optimized asphalt pavement two-dimensional texture data set in step S211 into asphalt pavement macro texture data and asphalt pavement micro texture data; S213, performing bilateral filtering and denoising on the asphalt pavement macro texture data in step S212 to obtain optimized asphalt pavement macro texture data, which is expressed as: Where: H(x, y) is the pixel value in the optimized macro texture data, and the optimized macro texture data is the set of all H(x, y), W p is the normalized weight, (x, y) is the coordinate of the current pixel, (i, j) is the coordinate of other pixels in the neighborhood, Ω is the pixel neighborhood, G s is the spatial weight function, P·|| is the norm 2, G r is the grayscale weight function, I(x,y) is the grayscale value of the current pixel, and I(i,j) is the grayscale value of other pixels in the neighborhood; S214, performing MAD method outlier denoising on the asphalt pavement micro texture data in step S212 to obtain first optimized asphalt pavement micro texture data, expressed as: in: is the pixel value of the first optimized micro texture data, and the first optimized micro texture data is all A collection of, M is the median of the micro texture data, X i is the pixel value in the micro texture data, k is the threshold multiple, MAD is the MAD method outlier of the micro texture data, otherwise is other cases; S215, performing wavelet transform denoising on the first optimized asphalt pavement micro-texture data in step S214 to obtain second optimized asphalt pavement micro-texture data, which is expressed as: in: is an image of the second optimized micro texture data, the second optimized micro texture data is all A set of, a is the scale parameter, b is the location parameter, is the wavelet coefficient after denoising, ψ a,b (x,y) is the wavelet basis function; S216, reconstructing the optimized asphalt pavement macro texture data in step S213 and the second optimized asphalt pavement micro texture data in step S215 to obtain a denoised asphalt pavement two-dimensional texture data set; S22, performing multi-scale decomposition processing on the denoised asphalt pavement two-dimensional texture dataset in step S21 by using discrete Fourier transform, obtaining texture information with wavelengths of 0.1-0.5 mm, 0.5-4 mm, 4 mm-16 mm, 16 mm-50 mm and greater than 50 mm, so as to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement; Step S22 includes the following steps: S221, converting the denoised asphalt pavement two-dimensional texture data set in step S21 into the frequency domain by discrete Fourier transform to obtain the frequency component of the two-dimensional texture of the asphalt pavement, expressed as: Where: X1[k,l] is the frequency component of the two-dimensional texture of the asphalt road surface, [k,l] is the horizontal and vertical coordinates of the spectrum domain, M is the number of rows, N is the number of columns, m1[x,y] is the pixel value of the two-dimensional texture in the denoised two-dimensional texture data set, [x,y] is the horizontal and vertical coordinates; S222, sequentially retaining the texture information of the target wavelengths of 0.1-0.5 mm, 0.5-4 mm, 4 mm-16 mm, 16 mm-50 mm and greater than 50 mm in the frequency components of the two-dimensional texture of the asphalt pavement in step S221, and filtering out the clutter outside the target wavelength to obtain the frequency components of the two-dimensional texture of the asphalt pavement after the clutter is filtered out; S223, converting the frequency components of the asphalt pavement two-dimensional texture after the clutter is filtered out in step S222 into the asphalt pavement two-dimensional texture through inverse Fourier transform to obtain a multi-scale two-dimensional texture dataset of the asphalt pavement, which is expressed as: Among them: m2[x,y] is the pixel value of the multi-scale two-dimensional texture of the asphalt pavement, the multi-scale two-dimensional texture dataset of the asphalt pavement is the set of all m2[x,y], X2[k,l] is the frequency component of the two-dimensional texture of the asphalt pavement after filtering out clutter; S3, constructing a friction prediction model, and using the asphalt pavement friction dataset in step S1 and the asphalt pavement multi-scale two-dimensional texture dataset in step S2 to train the friction prediction model, obtaining the trained friction prediction model, and integrating the trained friction prediction model into the inspection vehicle; S4. Scan the asphalt pavement using the inspection vehicle integrated with the trained friction prediction model in step S3 to obtain the friction prediction result of the asphalt pavement.

2. The friction prediction method based on multi-scale decomposition of asphalt pavement texture according to claim 1 is characterized in that: Step S1 includes the following steps: S11, determining the test road section measuring points, and using a detection vehicle equipped with a high-speed laser scanner to scan the asphalt pavement texture at the test road section measuring points to obtain a two-dimensional texture data set of the asphalt pavement; Each individual data of the asphalt pavement two-dimensional texture data set is 512×512 two-dimensional elevation point data, and the horizontal spacing between the two-dimensional elevation points is 0.05 mm; S12. Use a dynamic friction tester to measure the friction performance of the test section measuring points in step S11 to obtain an asphalt pavement friction data set.

3. The friction prediction method based on multi-scale decomposition of asphalt pavement texture according to claim 1 is characterized in that: Step S3 includes the following steps: S31, matching the asphalt pavement friction dataset in step S1 with the asphalt pavement multi-scale two-dimensional texture dataset in step S2, and dividing the matched datasets into a training set, a test set, and a validation set in a ratio of 8:1:1; S32, constructing a friction prediction model including a residual network for extracting information of different scales, an aggregation network for fusing information of different scales, and an output module for obtaining a friction prediction result of an asphalt pavement; S33. Use the training set, test set and validation set in step S31 to train, test and validate the friction prediction model in step S32, obtain the trained friction prediction model, and integrate the trained friction prediction model into the inspection vehicle.

4. The friction prediction method based on multi-scale decomposition of asphalt pavement texture according to claim 3 is characterized in that: In step S32, constructing a residual network for extracting information of different scales includes the following steps: A1. The multi-scale two-dimensional texture data of asphalt pavement with a size of 512×512×1 is used as the first input data of the friction prediction model. It first passes through a convolution kernel with a size of 7×7, a number of channels of 64, and a stride of 2, and then passes through normalization and ReLU activation function, and then passes through the maximum pooling with a stride of 2 to obtain the first feature data with a size of 128×128×64; A2, taking the first feature data of size 128×128×64 in step A1 as the second input data, and obtaining the second feature data of size 128×128×64 through three first residual modules; the first residual module first passes through a convolution kernel of size 3×3, number of channels 64, and stride 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 64, and stride 1 and standardization, and then adds it to the second input data, and then passes through the Relu activation function; A3, taking the second feature data of size 128×128×64 in step A2 as the third input data, first passing through a second residual module, and then passing through three third residual modules to obtain the third feature data of size 64×64×128; the second residual module first passes through a convolution kernel of size 3×3, number of channels 128, and stride 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 128, stride 1 and standardization, and then adds it to the third input data, and then passes through the Relu activation function; the third residual module first passes through a convolution kernel of size 3×3, number of channels 128, and stride 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 128, stride 1 and standardization, and then adds it to the third input data, and then passes through the Relu activation function; A4, taking the third feature data of size 64×64×128 in step A3 as the fourth input data, first passing through a fourth residual module, and then passing through five fifth residual modules to obtain fourth feature data of size 32×32×256; the fourth residual module first passes through a convolution kernel of size 3×3, number of channels 256, and stride 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 256, stride 1 and standardization, and then adds it to the fourth input data, and then passes through the Relu activation function; the fifth residual module first passes through a convolution kernel of size 3×3, number of channels 256, and stride 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel of size 3×3, number of channels 256, stride 1 and standardization, and then adds it to the fourth input data, and then passes through the Relu activation function; A5. Use the fourth feature data of size 32×32×256 in step A4 as the fifth input data, first pass through a sixth residual module, and then pass through two seventh residual modules to obtain the fifth feature data of size 16×16×512; the sixth residual module first passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 2, and then passes through standardization and Relu activation function, and then passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1 and standardization, and then adds it to the fifth input data, and then passes through the Relu activation function; the seventh residual module first passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1, and then passes through standardization and Relu activation function, and then passes through a convolution kernel with a size of 3×3, a number of channels of 512, and a step size of 1 and standardization, and then adds it to the fifth input data, and then passes through the Relu activation function.

5. The friction prediction method based on multi-scale decomposition of asphalt pavement texture according to claim 3 is characterized in that: In step S32, constructing an aggregation network for fusing information of different scales includes the following steps: B1. Concatenate the six fifth feature data with a size of 16×16×512 output by the residual network to obtain the sixth feature data with a size of 16×16×3072; B2, the sixth feature data of size 16×16×3072 in step B1 is first standardized, and then passed through a convolution kernel of size 3×3, number of channels 512, stride 1 and Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the seventh feature data of size 8×8×512; B3, the seventh feature data with a size of 8×8×512 in step B2 is first passed through a convolution kernel with a size of 3×3, a channel number of 256, and a stride of 1 and a Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the eighth feature data with a size of 4×4×256; B4, the eighth feature data with a size of 4×4×256 in step B3 is first passed through a convolution kernel with a size of 3×3, a number of channels of 128, and a stride of 1 and a Relu activation function, and then through a maximum pooling with a stride of 2 to obtain the ninth feature data with a size of 2×2×128; B5, the ninth feature data of size 2×2×128 in step B4 is first passed through a convolution kernel of size 3×3, number of channels 64, stride 1 and Relu activation function, and then through maximum pooling with stride 2 to obtain the tenth feature data of size 1×1×64; B6. The tenth feature data with a size of 1×1×64 in step B5 is passed through a convolution kernel with a size of 1×1, a number of channels of 32, and a step size of 1 and a Relu activation function to obtain an eleventh feature data with a size of 1×1×32.

6. The friction prediction method based on multi-scale decomposition of asphalt pavement texture according to claim 5 is characterized in that: In step B1, the 6 fifth feature data with a size of 16×16×512 are texture data of five different scales with wavelengths of 0.1-0.5mm, 0.5-4mm, 4mm-16mm, 16mm-50mm and greater than 50mm in the asphalt pavement multi-scale two-dimensional texture dataset, as well as the texture data in the asphalt pavement two-dimensional texture dataset, which are output by the residual network respectively.

7. The friction prediction method based on multi-scale decomposition of asphalt pavement texture according to claim 3 is characterized in that: In step S32, constructing an output module for obtaining the friction prediction result of the asphalt pavement includes the following steps: C1. Flatten the eleventh feature data of size 1×1×32 output by the aggregation network to obtain 32 twelfth feature data; C2. The 32 twelfth feature data in step C1 are first passed through a fully connected layer with 128 nodes, then through a Relu activation function, and then through a fully connected layer with 1 node to obtain the friction prediction result of the asphalt pavement.