Pavement structure depth prediction method based on heterogeneous modeling and machine learning

Abstract

The invention belongs to the technical field of pavement detection, and particularly relates to a pavement structure depth prediction method based on heterogeneous modeling and machine learning. According to the method, firstly, a road plane image is obtained, a deep learning semantic segmentation method is adopted to carry out segmentation marking on aggregate, coarse aggregate grading is estimated, then a random aggregate modeling method is adopted to carry out aggregate model putting, gravity is applied to the aggregate, virtual compaction is carried out on an aggregate pile, and an asphalt pavement heterogeneous model is generated. The method comprises the following steps: firstly, calculating the average pavement structure depth for multiple times by using MATLAB software, dividing multiple groups of data obtained by calculation into a training set and a test set, constructing a neural network prediction model, and finally, inputting estimated coarse aggregate grading data into the neural network prediction model, and predicting the pavement structure depth in a plane image. According to the method, the prediction of the tectonic depth can be realized without relying on an entity test, the research cost and the time investment are greatly reduced, and support can be provided for rapid decision-making in an engineering design stage.

Description

Technical Field

[0001] The present invention belongs to the technical field of pavement detection, and particularly relates to a pavement texture depth prediction method based on heterogeneous modeling and machine learning. Background Art

[0002] In the construction of transportation infrastructure, the performance evaluation and prediction of pavement structures are key links to ensure road safety and durability. Among them, the texture depth of asphalt pavements, as an important indicator to measure the anti-skid performance of pavements, directly affects driving safety and comfort. However, due to the complexity of asphalt pavement material composition and the uncertainty of the construction process, the accurate prediction of texture depth has always been a difficult problem in engineering practice.

[0003] Traditional texture depth evaluation methods mainly rely on on-site tests, such as the sand paving method, laser detection, etc. For example, in the patent application with the publication number CN116956584A, a full-scale pavement identical to the structure and materials of the road to be predicted is built in an open natural environment, and a vehicle-mounted laser texture depth meter is used to detect the texture depth of the full-scale pavement. These methods are not only time-consuming and laborious, but also easily affected by environmental factors and human operations, resulting in insufficient reliability and representativeness of the measurement results. Although some studies have proposed empirical formulas and numerical simulation methods based on material composition characteristics, these methods often overly simplify the heterogeneous characteristics of pavement structures and cannot accurately reflect the influence law of aggregate gradation on texture depth, thus limiting the prediction accuracy. Currently, the emerging image recognition method based on deep learning still needs to rely on physical specimens or existing pavements for prediction, and this dependence on actual road image samples greatly limits its application in the engineering design stage. Summary of the Invention

[0004] The present invention provides a pavement texture depth prediction method based on heterogeneous modeling and machine learning. This method can not only simulate the real pavement structure through random aggregate generation technology, but also accurately describe material characteristics through heterogeneous modeling. Finally, a high-precision prediction model is established with the help of a neural network, which can accurately predict the texture depth without relying on physical tests. This not only greatly reduces the research cost and time investment, but also provides strong support for rapid decision-making in the engineering design stage, and has important theoretical value and engineering significance for improving the pavement engineering quality control level and promoting the development of intelligent construction technology.

[0005] The method includes the following steps:

[0006] S1. Use a high-precision camera to obtain a planar image of the road, segment and label the aggregates in the planar image by using the semantic segmentation method of deep learning, and estimate the coarse aggregate gradation adopted by the road.

[0007] S2. Based on the estimated coarse aggregate gradation, use the random aggregate modeling method to place the aggregate model and obtain an initial heterogeneous model;

[0008] S3. Apply gravity to the aggregates in the initial heterogeneous model to make them fall and form an aggregate pile. Apply a vertical downward virtual load to the top of the aggregate pile for virtual compaction to generate a heterogeneous model of the asphalt pavement, and output the digital image of the heterogeneous model of the asphalt pavement;

[0009] S4. According to the heterogeneous model of the asphalt pavement, use MATLAB software to calculate the average pavement texture depth;

[0010] S5. Repeat the operations in steps S2 to S4 to obtain multiple sets of data of the coarse aggregate gradation and the corresponding data of the average pavement texture depth;

[0011] S6. Divide the multiple sets of data obtained in step S5 into a training set and a test set, and construct a neural network prediction model;

[0012] S7. Use the data of the estimated coarse aggregate gradation in step S1 as the input data of the neural network prediction model to predict the texture depth of the pavement in the planar image.

[0013] In a specific embodiment, step S2 includes the following specific steps:

[0014] S2.1. Create a region with a specified aspect ratio in MATLAB software, randomly generate points in the region, and use the convex hull algorithm to generate a closed polygon to obtain multiple random aggregate templates with different two-dimensional shapes, and save the information of the generated aggregate templates;

[0015] S2.2. Import the aggregate template information into the PFC software, determine the two-dimensional boundary of the heterogeneous modeling, calculate the volume fraction of the coarse aggregates according to the estimated coarse aggregate gradation and the mix ratio of the road, and randomly generate circular particles with corresponding particle sizes within the two-dimensional boundary according to the estimated coarse aggregate gradation and the volume fraction of the coarse aggregates, and ensure that the circular particles do not overlap with each other;

[0016] S2.3. Then randomly replace each circular particle with a different-shaped aggregate template, calculate the scaling ratio of the aggregate template according to the particle size of the circular particle, so that the longest diagonal of the aggregate template is equal to the diameter of the circular particle, complete the random placement of the aggregates, and obtain the initial heterogeneous model.

[0017] In a specific embodiment, step S3 includes the following specific steps:

[0018] S3.1. Apply gravity to the aggregates in the initial heterogeneous model in the PFC software to make the aggregates fall and ensure that the aggregates naturally accumulate under the action of gravity.

[0019] S3.2. Directly apply a vertical virtual load to the top of the aggregate pile in the PFC software, so that the coarse aggregate model gradually arranges and reaches a stable state under the compaction of gravity and the virtual load.

[0020] S3.3. Generate a heterogeneous model of the asphalt pavement and output the digital image of the heterogeneous model of the asphalt pavement.

[0021] In a specific embodiment, step S4 includes the following specific steps:

[0022] S4.1. Save all the geometric and spatial information of the heterogeneous model of the asphalt pavement generated in step S3 as a data packet. The content of the data packet includes the horizontal coordinates and height coordinates of each aggregate, and import the data packet into the MATLAB software.

[0023] S4.2. The MATLAB software extracts the maximum height coordinate value at the same horizontal coordinate value in the data packet, calculates the reference height of the heterogeneous model of the asphalt pavement, screens out the height coordinate values of the aggregates exceeding the reference height in the data packet, and calculates the average difference of these height coordinate values, that is, the texture depth of the heterogeneous model of the asphalt pavement is obtained.

[0024] In a specific embodiment, step S6 includes the following specific steps:

[0025] S6.1. Divide the data of multiple groups of coarse aggregate gradations and the corresponding average pavement texture depth data obtained in step S5 into a training set and a test set. Among them, 80% of the data groups are used for training and 20% of the data groups are used for testing.

[0026] S6.2. Build a feedforward neural network model for regression prediction.

[0027] S6.3. After the training of the feedforward neural network model is completed, verify the performance of the feedforward neural network model on the test set, and evaluate the prediction effect by calculating the mean square error and the coefficient of determination.

[0028] The present invention has at least the following beneficial effects:

[0029] 1. Compared with the traditional modeling method, the heterogeneous model of the present method can construct the corresponding mesoscopic model of the pavement structure according to different gradation information, laying a foundation for calculating the pavement texture depth.

[0030] 2. This method predicts the texture depth of the road surface based on heterogeneous modeling and neural network prediction models. Compared with traditional road surface texture depth test methods, this method can obtain the texture depth of different gradations through heterogeneous modeling, easily obtain a large number of data samples for the training of the neural network prediction model, and avoid a large number of experiments.

[0031] 3. This method uses deep learning semantic segmentation technology to segment road surface images and estimate the gradation of coarse aggregates, which has the characteristics of low cost and high efficiency.

[0032] 4. This method, through the combination of machine learning and heterogeneous modeling, is applicable to the early prediction of road surface texture depth. When the gradation is known, the texture depth can be directly predicted. When the gradation is unknown, only by obtaining geometric features through image information, the prediction of road surface texture depth can be carried out. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] Please refer to Figure 1 , a method for predicting the texture depth of a road surface based on heterogeneous modeling and machine learning provided by the present invention specifically includes the following steps:

[0035] S1. Use a high-precision camera to obtain a planar image of the road, and use the semantic segmentation method of deep learning to segment and calibrate the aggregates in the planar image to obtain the contour information of the road surface aggregate image, including perimeter, area, major and minor axis values. Then, identify and mark the aggregates with particle sizes between 4.75 mm and 9.5 mm, 9.5 mm and 16 mm, and 16 mm and 19 mm, and calculate their contents respectively, so as to estimate the gradation of the coarse aggregates used in this road.

[0036] S2. According to the estimated gradation of the coarse aggregates, use the random aggregate modeling method to place the aggregate model. The specific operation is as follows:

[0037] S2.1. Define the target aspect ratio r = w / h in MATLAB software, and specify the number of vertices N of the polygon. Based on a rectangular area, fix its height h as 1, and calculate its width w according to the target aspect ratio. Randomly generate N points within this rectangular area. The abscissa x of each point is randomly taken within the range [0, w], and the ordinate y is randomly taken within the range [0, h], ensuring that the points are distributed within the specified rectangular area. Then, use the convex hull algorithm to process the random points. The convex hull algorithm will connect these points into a closed polygon, such that all vertices of the polygon are located on the outer boundary of the polygon and there are no overlapping areas inside the polygon, thereby creating a single aggregate template with the specified aspect ratio. Repeat this operation to generate random aggregate templates of different two-dimensional shapes, and store the shape information of these aggregate templates in a file format suitable for PFC software, such as an STL format file.

[0038] S2.2. Import the aggregate template in STL format into PFC software. Determine the two-dimensional boundary size for heterogeneous modeling as 1m × 3m. Calculate the volume fraction of coarse aggregates according to the estimated coarse aggregate gradation and the mix ratio of this road. Randomly generate circular particles with corresponding particle sizes within the above two-dimensional boundary according to the estimated coarse aggregate gradation and the volume fraction of coarse aggregates, and ensure that the circular particles do not overlap with each other.

[0039] 2.3. Then randomly replace each circular particle with an aggregate template of a different shape, and calculate the scaling ratio of the aggregate template according to its particle size, such that the longest diagonal of the aggregate template is equal to the diameter of the circular particle, completing the random placement of aggregates. Perform a secondary overall scaling according to the target porosity and the total volume of aggregates to obtain the initial heterogeneous model.

[0040] S3. Apply gravity to the aggregates in the initial heterogeneous model in PFC software through the set gravity instruction, causing the aggregates to fall and ensuring that the aggregates naturally accumulate under the action of gravity. Then, in PFC software, apply a virtual load that gradually increases with time to the top area of the aggregate pile by directly applying a vertical pressure to the top of the aggregate pile. After the compaction process under the action of gravity and the virtual load, the aggregates will gradually arrange and reach a stable state, finally generating a heterogeneous model of the asphalt pavement and outputting a digital image of the heterogeneous model of the asphalt pavement.

[0041] S4. Save all geometric and spatial information of the heterogeneous model of the asphalt pavement generated in step S3 as a data packet. The content of this data packet includes the coordinate positions of each aggregate, where x represents the horizontal position of the aggregate and y represents the height position of the aggregate. Import the data packet into MATLAB software. MATLAB software extracts the maximum height value y corresponding to the same x value in the data packet to form an array H n , representing the elevation coordinate array of the road surface. Find the array Hn The arithmetic mean value is used to obtain the reference height value h of the heterogeneous model of the asphalt pavement b , search through the array H n , extract all coordinate values greater than h b to form an array h 1 . For each value h 1 in the array h 1(i) , calculate the difference deth b between it and the reference height h (i) = h 1(i) - h b , to form a new array detH. Calculate the average value of the numerical values in the array detH to obtain the texture depth of the heterogeneous model of the asphalt pavement.

[0042] S5. Repeat the operations in steps S2 to S4 to obtain 500 groups of data on the coarse aggregate gradation and the corresponding data on the average pavement texture depth.

[0043] S6. Divide the 500 groups of data on the coarse aggregate gradation and the corresponding data on the average pavement texture depth obtained in step S5 into a training set and a test set. Among them, 80% of the groups of data are used for training, and 20% of the groups of data are used for testing. Construct a feedforward neural network model for regression prediction to establish the corresponding relationship between the coarse aggregate gradation and the texture depth of the heterogeneous model of the asphalt pavement under this gradation. After the training of the feedforward neural network model is completed, verify the performance of the feedforward neural network model on the test set, and evaluate the prediction effect by calculating the mean square error and the coefficient of determination.

[0044] S7. Input the data on the coarse aggregate gradation estimated in step S1 into the neural network prediction model, and the texture depth of the pavement in the planar image can be predicted.

[0045] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. It cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions and substitutions can still be made, which should all be regarded as belonging to the protection scope of the present invention.

Claims

1. A pavement structure depth prediction method based on heterogeneous modeling and machine learning, characterized in that: The steps include: S1. Using a high-precision camera to obtain a plane image of the road, using a deep learning semantic segmentation method to segment and mark the aggregates in the plane image, and estimating the coarse aggregate gradation used in the road; S2. Based on the estimated coarse aggregate gradation, a random aggregate modeling method is used to place the aggregate model to obtain an initial heterogeneous model; S3, applying gravity to the aggregates in the initial heterogeneous model to make them fall, forming an aggregate pile, applying a vertically downward virtual load to the top of the aggregate pile for virtual compaction, generating an asphalt pavement heterogeneous model, and outputting a digital image of the asphalt pavement heterogeneous model; S4. Calculate the average pavement structural depth using MATLAB software according to the asphalt pavement heterogeneity model; S5, repeating the operations of steps S2 to S4 to obtain multiple sets of data on coarse aggregate gradation and corresponding data on average pavement structure depth; S6, dividing the multiple groups of data obtained in step S5 into a training set and a test set, and constructing a neural network prediction model; S7, using the data of the coarse aggregate gradation estimated in step S1 as input data of the neural network prediction model to predict the structural depth of the road surface in the plane image.

2. The pavement structure depth prediction method based on heterogeneous modeling and machine learning according to claim 1 is characterized in that: Step S2 includes the following specific steps: S2.1, by creating a region with a specified aspect ratio in MATLAB software, randomly generating points in the region, and using a convex hull algorithm to generate a closed polygon, a plurality of random aggregate templates with different two-dimensional shapes are obtained, and the generated aggregate template information is saved; S2.2, importing the aggregate template information into the PFC software, determining the two-dimensional boundary of the heterogeneous modeling, calculating the coarse aggregate volume fraction according to the estimated coarse aggregate gradation and the mix ratio of the road, and randomly generating circular particles of corresponding particle sizes within the two-dimensional boundary according to the estimated coarse aggregate gradation and coarse aggregate volume fraction, and ensuring that the circular particles do not overlap each other; S2.

3. Then randomly replace each circular particle with an aggregate template of a different shape, calculate the scaling ratio of the aggregate template according to the particle size of the circular particle, so that the longest diagonal of the aggregate template is equal to the diameter of the circular particle, complete the random placement of aggregate, and obtain the initial heterogeneous model.

3. The pavement structure depth prediction method based on heterogeneous modeling and machine learning according to claim 2 is characterized in that: Step S3 includes the following specific steps: S3.

1. Apply gravity to the aggregates in the initial heterogeneous model in the PFC software to make the aggregates fall and ensure that the aggregates are naturally accumulated under the action of gravity; S3.

2. Apply a vertical virtual load directly to the top of the aggregate pile in the PFC software, so that the coarse aggregate model is gradually arranged and reaches a stable state through the compaction of gravity and virtual load; S3.

3. Generate an asphalt pavement heterogeneous model and output a digital image of the asphalt pavement heterogeneous model.

4. The pavement structure depth prediction method based on heterogeneous modeling and machine learning according to claim 3 is characterized in that: Step S4 includes the following specific steps: S4.

1. Save all geometric and spatial information of the asphalt pavement heterogeneous model generated in step S3 as a data package, wherein the content of the data package includes the horizontal coordinates and height coordinates of each aggregate, and import the data package into MATLAB software; S4.

2. MATLAB software extracts the maximum height coordinate value under the same horizontal coordinate value in the data package, calculates the benchmark height of the asphalt pavement heterogeneous model, screens out the height coordinate values ​​of aggregates exceeding the benchmark height in the data package, and calculates the difference average of these height coordinate values, that is, the construction depth of the asphalt pavement heterogeneous model is obtained.

5. The pavement structure depth prediction method based on heterogeneous modeling and machine learning according to any one of claims 1 to 4, characterized in that: Step S6 includes the following specific steps: S6.1, dividing the multiple groups of data on coarse aggregate gradation and the corresponding data on average pavement structure depth obtained in step S5 into a training set and a test set, wherein 80% of the data are used for training and 20% of the data are used for testing; S6.2, construct a feedforward neural network model to perform regression prediction; S6.

3. After the feedforward neural network model training is completed, the performance of the feedforward neural network model is verified on the test set, and the prediction effect is evaluated by calculating the mean square error and determination coefficient.

Images