Automatic circulation optimization method for pavement structure design based on machine learning

Abstract

The invention discloses a machine learning-based pavement structure design automatic loop optimization method, which comprises the following steps: S1, establishing a pavement structure finite element analysis result database, generating pavement mechanical response data under different loads, materials and structure parameters, and providing data labels for subsequent finite element-deconvolution neural network training; s2, establishing a finite element-deconvolution neural network, performing network training based on the pavement structure finite element analysis result database, and mapping a finite element analysis process to the finite element-deconvolution neural network; and S3, based on the trained finite element-deconvolution neural network, inputting pavement structure parameters which are not in a pre-selected range, and performing forward prediction. The problems that in the prior art, stress, strain and plastic deformation distribution matrixes of a pavement structure layer under different thickness combinations, material parameters and traffic loads cannot be rapidly solved, and the stress, strain and plastic deformation of a control point cannot be minimized by automatically optimizing the thickness of the structure layer are solved.

Description

Technical Field

[0001] This invention belongs to the field of pavement structure mechanics calculation technology, specifically a machine learning-based automatic iterative optimization method for pavement structure design. Background Technology

[0002] The properties of pavement materials and their structural combination have a significant impact on the service life of pavements. The main purpose of pavement design is to determine the types of materials used and their corresponding thicknesses in different layers, ensuring that the overall mechanical properties of the structure meet the expected requirements. For different pavement types and functional layers, mechanical calculations need to obtain mechanical response indicators such as stress, strain, and permanent deformation at different control points. Therefore, whether designing new pavements or performing major repairs, mechanical response analysis and verification of the structure are necessary. Currently, pavement structure design is mainly carried out through manual design and mechanical response verification.

[0003] Based on pre-selected materials and combined thicknesses, the mechanical response of pavement structures can be calculated using the theoretical formulas of elastic layered systems, yielding the stress and strain at each layer. Traditional elastic layered system mechanical models only calculate the mechanical response parameters at different points under two-dimensional conditions, and are generally only applicable to pavement structures with three or fewer layers. Finite element analysis (FEM) can automatically solve for stress and deformation at points across the entire spatial domain, and can be used for complex pavement structure combinations. However, FEM-based pavement structure mechanical calculations generally rely on high-performance computing equipment and are time-consuming, especially for complex geometries and fine meshes. Therefore, FEM cannot quickly verify massive pavement structure combinations, leading to difficulties in automated pavement structure design.

[0004] In response to the large computational load of finite element method, some scholars have proposed using deep learning neural networks to predict the mechanical properties of material structures. Deep learning-based prediction of the mechanical properties of material structures can basically achieve rapid calculation of mechanical parameters, but there are still the following problems: (1) It can only calculate mechanical response parameters at typical and key locations, and cannot predict stress and strain cloud maps across the entire domain, nor can it predict plastic deformation; (2) It can only calculate the mechanical response differences under variable load conditions for a single structure, and cannot input parameters of different structures and materials; (3) It cannot achieve automatic optimization of structural layer thickness, and cannot solve for the optimal combination of structural thicknesses with mechanical properties under multiple constraints such as construction cost and load level.

[0005] Therefore, an automatic iterative optimization method for pavement structure design based on machine learning is provided. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides an automatic iterative optimization method for pavement structure design based on machine learning. This method solves the problem that existing technologies cannot quickly solve the stress, strain, and plastic deformation distribution matrix of pavement structure layers under different thickness combinations, material parameters, and traffic loads, and automatically optimizes the structure layer thickness to minimize the stress, strain, and plastic deformation at control points.

[0007] The technical solution to achieve the above objectives is: An automated iterative optimization method for pavement structure design based on machine learning, comprising: Step S1: Establish a database of finite element analysis results for pavement structures, generate pavement mechanical response data under different loads, materials and structural parameters, and provide data labels for subsequent finite element-deconvolutional neural network training; Step S2: Establish a finite element-deconvolutional neural network, train the network based on the database of finite element analysis results of road structure, and map the finite element analysis process to the finite element-deconvolutional neural network. Step S3: Based on the trained finite element-deconvolutional neural network, input road structure parameters that are not in the pre-selected range, perform positive prediction, and evaluate the prediction error range through finite element analysis. Step S4: Based on the fast forward prediction of the finite element-deconvolutional neural network, define the objective function for pavement structure design, and search for the pavement structure layer thickness combination with the best mechanical performance through heuristic algorithm iterative optimization.

[0008] Preferably, step S1 includes: Step S11: Based on typical pavement structure and material properties, determine the parameter range of the number of pavement structure layers, thickness, material and load. Within the pre-selected parameter range, arrange and combine the number of layers, layer thickness, material and load parameters. Step S12 involves creating components, defining material sections, assembling components, creating analysis steps, defining boundaries and loads to establish finite element models of road structures for different scenarios and a database of finite element analysis results. Step S13: Mesh the established finite element models of road surface structures for different scenarios and solve the problem. Step S14: Output the mechanical response calculation results of road surface structures in different scenarios, including the stress, strain and plastic deformation matrices of global control points. Match the model input parameters and the output results of the finite element analysis for subsequent finite element-deconvolutional neural network training.

[0009] Preferably, in step S11, the pavement structure layer modeled by finite element modeling includes one or more base layers and asphalt surface layers, with a total number of 2-5 layers, each layer having a thickness range of 4-20 cm, and the material types including cement-stabilized crushed stone, graded crushed stone, asphalt crushed stone, asphalt concrete and asphalt mastic concrete, and the load range being 200,000 to 1,000,000 cycles of BZZ-100 standard axle load. In step S12, finite element models of road surface structures for different scenarios are established based on ABAQUS or ANSYS analysis software. In step S14, the stress, strain, and plastic deformation matrix output by the finite element analysis is 393×322.

[0010] Preferably, step S2 includes: Step S21: Construct a finite element-deconvolutional neural network; Step S22: Determine the loss function and hyperparameters, use the arithmetic mean of dice coefficient loss and cross-entropy loss to represent the error, and use dynamic decreasing learning rate and early stopping strategy to search for the optimal training cycle; Step S23: Train the finite element-deconvolutional neural network. Using the finite element analysis result database of the road structure generated in step S1, divide it into a training set and a test set. During the training process, a validation set is set to calculate the validation loss. The ratio of the sample size used in the training set to the test set is 8:2, and the sample size of the validation set is 10% of that in the training set.

[0011] Preferably, in step S21, constructing a finite element-deconvolutional neural network includes: The network input is a (6N+1)×1 vector, which includes the load axis and the thickness of N structural layers, material modulus, Poisson's ratio and three creep parameters, where N is the total number of pavement structural layers, and N is between 2 and 5. The input features are expanded to a 1024×1 feature vector through a fully connected layer and then reorganized to generate an initial feature matrix of 32×32×1. The initial feature matrix is ​​expanded into 64 36×36 feature matrices through a 5×5 convolution kernel and a 1×1 stride deconvolution operation. The feature matrix is ​​further upsampled by using a 3×3 convolution kernel and a 1×1 stride deconvolution operation to generate 64 38×38 feature matrices. The number of feature matrices is reduced by using a 3×3 convolution kernel and a 2×2 stride deconvolution operation, resulting in 32 77×77 feature matrices. The feature matrix size is further expanded by using a 3×3 convolution kernel and a 5×4 stride deconvolution operation to generate 32 feature matrices of 385×308. By using a 3×3 convolution kernel and a 1×1 stride deconvolution, the number of channels is compressed to 16, generating 16 feature matrices of 387×310. Finally, three 393×322 matrices are output through a 7×13 convolution kernel and a 1×1 stride deconvolution: the stress matrix, the strain matrix, and the plastic deformation matrix, corresponding to the stress, strain, and plastic deformation values ​​at different locations of the pavement structure.

[0012] Preferably, in step S23, each set of training or testing data includes 6N+1 road structure input parameters and 3 393×322 mechanical response matrix labels. Error is backpropagated by the loss value between the prediction matrix and the label matrix to achieve network training.

[0013] Preferably, step S3 includes: Step S31: Determine the road structure parameters that are not in the pre-selection range, and determine the input parameters by interpolation or extrapolation methods; Step S32: Based on the pavement structure parameters that are not in the pre-selected range, input the finite element-deconvolutional neural network for forward prediction, output the stress, strain and plastic deformation matrices, and compare them with the finite element simulation results to calculate the prediction matrix error; In step S33, if the positive prediction result is within the allowable error range, proceed to step S4; if the error requirement is not met, return to step S2 to retrain.

[0014] Preferably, in step S32, the road material parameters input to the network's positive prediction are within 50% of the upper and lower limits of common asphalt material properties, used to simulate abnormal material properties; the load axle cycle range is increased to 1.5 million cycles, used to simulate extremely heavy loads. The prediction matrix error is calculated using Euclidean distance or Mahalanobis distance.

[0015] Preferably, step S4 includes: Step S41: For the pavement structure design optimization problem, define the number of pavement structure layers, material range, and constraints. Simultaneously, set an optimization objective to minimize stress, strain, or plastic deformation at different locations, and construct the objective optimization function. ,variable The thickness of each layer; Step S42: A heuristic algorithm is used to search for the optimal solution in a loop. First, an initial solution set of road thickness combinations is randomly generated. The mechanical response under different thickness combinations is quickly calculated by the finite element-deconvolutional neural network in step S3. The solution set is ranked by constraint penalty and optimization objective function. The solution set is adjusted in a loop until the convergence condition is met and the optimization ends. Step S43: After optimization using the finite element-deconvolutional neural network and heuristic algorithm, the optimal combination of structural thicknesses and the mechanical response of each control point are obtained. Step S44: Finally, the optimal thickness combination is verified by conventional finite element analysis to determine that the mechanical response of each control point is within the range required by the specification, thereby obtaining the best combination of pavement structure layer thicknesses.

[0016] Preferably, in step S41, the objective optimization function is to minimize the permanent deformation at the load center location, the tensile strain at the bottom of the surface layer, or the tensile stress at the bottom of the base layer. In step S42, the constraints include the total thickness of the pavement structure, construction cost, and environmental emissions. In step S43, the heuristic algorithm is one of the following: genetic algorithm, ant colony algorithm, or simulated annealing algorithm.

[0017] Compared with the prior art, the beneficial effects of the present invention are: the present invention can quickly predict the mechanical response of different pavement materials and structures under variable loads, replacing complex finite element calculations and greatly saving computing power; it can predict the stress, strain and plastic deformation matrix of the pavement structure in the whole domain and obtain the instantaneous and long-term mechanical response characteristics of the pavement structure; it can predict the permanent deformation of the pavement caused by repeated loads; it can realize the rapid optimization of the pavement structure and automatically solve the combination of structural layer thicknesses with the best mechanical performance. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an automatic iterative optimization method for road structure design based on machine learning, according to the present invention. Figure 2 This is a flowchart of the process of establishing a database of finite element analysis results for pavement structures and generating pavement mechanical response data under different loads, materials and structural parameters in this invention. Figure 3 This is a flowchart of the process of establishing a finite element-deconvolutional neural network in this invention, training the network based on the database of finite element analysis results of road structure, and mapping the finite element analysis process to the finite element-deconvolutional neural network. Figure 4 This is a flowchart of the present invention, which is based on a trained finite element-deconvolutional neural network, inputs road structure parameters that are not in the pre-selected range, performs positive prediction, and evaluates the prediction error range through finite element analysis. Figure 5 This is a flowchart of the present invention, which defines the objective function for road structure design based on the fast forward prediction of finite element-deconvolutional neural network, and searches for the combination of road structure layer thicknesses with the best mechanical performance through heuristic algorithm iterative optimization. Figure 6This is a schematic diagram of the finite element database for road surface mechanical response in this invention; Figure 7 This is a schematic diagram of the finite element-deconvolutional neural network structure in this invention; Figure 8 This is a flowchart of the finite element-deconvolutional neural network training process in this invention; Figure 9 This is a flowchart of the iterative optimization process for road structure design based on genetic algorithm in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, an automatic iterative optimization method for pavement structure design based on machine learning includes: Step S1: Establish a database of finite element analysis results for pavement structures, generate pavement mechanical response data under different loads, materials and structural parameters, and provide data labels for subsequent finite element-deconvolutional neural network training.

[0021] like Figure 2 , 6 As shown, step S1 includes: Step S11: Based on typical pavement structure and material properties, determine the parameter range of the number of pavement structure layers, thickness, materials, and loads. Within the pre-selected parameter range, arrange and combine the number of layers, layer thickness, materials, and load parameters.

[0022] In this embodiment, the pavement structure layer modeled by finite element method includes one or more base layers and asphalt surface layers. The total number of structural layers is 2-5, and the thickness of each layer ranges from 4 to 20 centimeters. The material types include cement-stabilized crushed stone, graded crushed stone, asphalt crushed stone, asphalt concrete, and asphalt mastic concrete. The load range is 200,000 to 1,000,000 cycles of BZZ-100 standard axle load.

[0023] Step S12 involves creating components, defining material sections, assembling components, creating analysis steps, defining boundaries and loads to establish finite element models of road structures for different scenarios and a database of finite element analysis results.

[0024] In the embodiments, finite element models of road surface structures for different scenarios are established based on ABAQUS or ANSYS analysis software.

[0025] Step S13: Mesh the established finite element models of road surface structures for different scenarios and solve the problems.

[0026] Step S14: Output the mechanical response calculation results of road surface structures in different scenarios, including the stress, strain and plastic deformation matrices of global control points. Match the model input parameters and the output results of the finite element analysis for subsequent finite element-deconvolutional neural network training.

[0027] In this embodiment, within a pre-selected parameter range, the number of layers, layer thickness, material, and load parameters are combined and arranged. Finite element models of pavement structures for different scenarios are established by creating components, defining material sections, assembling components, creating analysis steps, and defining boundaries and loads. The established finite element models of pavement structures for different scenarios are meshed and solved, outputting the mechanical response calculation results of the pavement structures for different scenarios, including global control point stress, strain, and plastic deformation matrices and corresponding contour maps. The output matrix size is 393×322. The model input parameters and finite element analysis output results are matched to generate a label data table containing more than 300 sets of simulation data for subsequent finite element-deconvolutional neural network training.

[0028] Step S2: Establish a finite element-deconvolutional neural network. Train the network based on the database of finite element analysis results of road structure and map the finite element analysis process to the finite element-deconvolutional neural network.

[0029] like Figure 3 , 7 As shown in Figure 8, step S2 includes: Step S21: Construct a finite element-deconvolutional neural network.

[0030] In this embodiment, constructing a finite element-deconvolutional neural network includes: The network input is a (6N+1)×1 vector, which includes the load axis and the thickness of N structural layers, material modulus, Poisson's ratio and three creep parameters, where N is the total number of pavement structural layers, and N is between 2 and 5. The input features are expanded to a 1024×1 feature vector through a fully connected layer and then reorganized to generate an initial feature matrix of 32×32×1. The initial feature matrix is ​​expanded into 64 36×36 feature matrices through a 5×5 convolution kernel and a 1×1 stride deconvolution operation. The feature matrix is ​​further upsampled by using a 3×3 convolution kernel and a 1×1 stride deconvolution operation to generate 64 38×38 feature matrices. The number of feature matrices is reduced by using a 3×3 convolution kernel and a 2×2 stride deconvolution operation, resulting in 32 77×77 feature matrices. The feature matrix size is further expanded by using a 3×3 convolution kernel and a 5×4 stride deconvolution operation to generate 32 feature matrices of 385×308. By using a 3×3 convolution kernel and a 1×1 stride deconvolution, the number of channels is compressed to 16, generating 16 feature matrices of 387×310. Finally, three 393×322 matrices are output through a 7×13 convolution kernel and a 1×1 stride deconvolution: the stress matrix, the strain matrix, and the plastic deformation matrix, corresponding to the stress, strain, and plastic deformation values ​​at different locations of the pavement structure.

[0031] Step S22: Determine the loss function and hyperparameters. Use the arithmetic mean of dice coefficient loss and cross-entropy loss to represent the error. Use dynamic decreasing learning rate and early stopping strategy to search for the optimal training cycle.

[0032] Step S23: Train the finite element-deconvolutional neural network. Using the finite element analysis result database of the road structure generated in step S1, divide it into a training set and a test set. During the training process, a validation set is set to calculate the validation loss. The ratio of the sample size used in the training set to the test set is 8:2, and the sample size of the validation set is 10% of that in the training set.

[0033] In this embodiment, each set of training or testing data includes 6N+1 road structure input parameters and 3 393×322 mechanical response matrix labels. Error is backpropagated by the loss value between the prediction matrix and the label matrix to achieve network training.

[0034] Step S3: Based on the trained finite element-deconvolutional neural network, input road structure parameters that are not in the pre-selected range, perform positive prediction, and evaluate the prediction error range through finite element analysis.

[0035] like Figure 4 As shown, step S3 includes: Step S31: Determine the road structure parameters that are not in the pre-selected range, and determine the input parameters by interpolation or extrapolation methods.

[0036] Step S32: Based on the pavement structure parameters that are not in the pre-selected range, input the finite element-deconvolutional neural network for forward prediction, output the stress, strain and plastic deformation matrices, and compare them with the finite element simulation results to calculate the prediction matrix error.

[0037] In this embodiment, the road material parameters input to the network's positive prediction are within 50% of the upper and lower limits of common asphalt material properties, used to simulate abnormal material properties; the load axle cycle range is increased to 1.5 million cycles, used to simulate extremely heavy loads. The prediction matrix error is calculated using Euclidean distance or Mahalanobis distance.

[0038] In step S33, if the positive prediction result is within the allowable error range, proceed to step S4; if the error requirement is not met, return to step S2 to retrain.

[0039] Step S4: Based on the fast forward prediction of the finite element-deconvolutional neural network, define the objective function for pavement structure design, and search for the pavement structure layer thickness combination with the best mechanical performance through heuristic algorithm iterative optimization.

[0040] like Figure 5 As shown, step S4 includes: Step S41: For the pavement structure design optimization problem, define the number of pavement structure layers, material range, and constraints. Simultaneously, set an optimization objective to minimize stress, strain, or plastic deformation at different locations, and construct the objective optimization function. ,variable The thickness of each layer.

[0041] In this embodiment, the objective optimization function is to minimize the permanent deformation at the load center location, the tensile strain at the bottom of the surface layer, or the tensile stress at the bottom of the base layer.

[0042] Step S42: A heuristic algorithm is used to iteratively search for the optimal solution. First, an initial solution set of road thickness combinations is randomly generated. Then, the mechanical response under different thickness combinations is quickly calculated using the finite element-deconvolutional neural network in step S3. The solution set is ranked by constraint penalties and optimization objective function. The solution set is iteratively adjusted until the convergence condition is met, and the optimization ends.

[0043] In the embodiments, the constraints include the total thickness of the pavement structure, construction cost, and environmental emissions.

[0044] Step S43: After optimization using the finite element-deconvolutional neural network and heuristic algorithm, the optimal combination of structural thicknesses and the mechanical response of each control point are obtained.

[0045] In this embodiment, the heuristic algorithm is one of the following: genetic algorithm, ant colony algorithm, or simulated annealing algorithm; for example, such as Figure 9 As shown, if a genetic algorithm is used to iteratively search for the optimal solution, an initial solution set of road thickness combinations is first randomly generated. The mechanical response under different thickness combinations is then quickly calculated using a finite element-deconvolutional neural network. The fitness of the solution set is evaluated by the structural mechanical response and constraint penalty. Selection, crossover, and mutation operations are performed on the initial solution set to generate a new generation of solution sets. It is then determined whether the new generation of solution sets converges. If they do not converge, the fitness of the solution set is evaluated again, and the selection, crossover, and mutation operations are repeated.

[0046] Step S44: Finally, the optimal thickness combination is verified by conventional finite element analysis to determine that the mechanical response of each control point is within the range required by the specification, thereby obtaining the best combination of pavement structure layer thicknesses.

[0047] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic iterative optimization method for pavement structure design based on machine learning, characterized in that, include: Step S1: Establish a database of finite element analysis results for pavement structures, generate pavement mechanical response data under different loads, materials and structural parameters, and provide data labels for subsequent finite element-deconvolutional neural network training; Step S2: Establish a finite element-deconvolutional neural network, train the network based on the database of finite element analysis results of road structure, and map the finite element analysis process to the finite element-deconvolutional neural network. Step S3: Based on the trained finite element-deconvolutional neural network, input road structure parameters that are not in the pre-selected range, perform positive prediction, and evaluate the prediction error range through finite element analysis. Step S4: Based on the fast forward prediction of the finite element-deconvolutional neural network, define the objective function for pavement structure design, and search for the pavement structure layer thickness combination with the best mechanical performance through heuristic algorithm iterative optimization.

2. The automatic iterative optimization method for pavement structure design based on machine learning according to claim 1, characterized in that, Step S1 includes: Step S11: Based on typical pavement structure and material properties, determine the parameter range of the number of pavement structure layers, thickness, material and load. Within the pre-selected parameter range, arrange and combine the number of layers, layer thickness, material and load parameters. Step S12 involves creating components, defining material sections, assembling components, creating analysis steps, defining boundaries and loads to establish finite element models of road structures for different scenarios and a database of finite element analysis results. Step S13: Mesh the established finite element models of road surface structures for different scenarios and solve the problem. Step S14: Output the mechanical response calculation results of road surface structures in different scenarios, including the stress, strain and plastic deformation matrices of global control points. Match the model input parameters and the output results of the finite element analysis for subsequent finite element-deconvolutional neural network training.

3. The automatic iterative optimization method for road structure design based on machine learning according to claim 2, characterized in that, In step S11, the road structure layer modeled by finite element modeling includes one or more base layers and asphalt surface layers. The total number of structural layers is 2-5 layers, the thickness of each layer is 4-20 cm, the material types include cement stabilized crushed stone, graded crushed stone, asphalt crushed stone, asphalt concrete and asphalt mastic concrete, and the load range is 200,000 to 1,000,000 cycles of BZZ-100 standard axle load. In step S12, finite element models of road surface structures for different scenarios are established based on ABAQUS or ANSYS analysis software. In step S14, the stress, strain, and plastic deformation matrix output by the finite element analysis is 393×322.

4. The automatic iterative optimization method for pavement structure design based on machine learning according to claim 1, characterized in that, Step S2 includes: Step S21: Construct a finite element-deconvolutional neural network; Step S22: Determine the loss function and hyperparameters, use the arithmetic mean of dice coefficient loss and cross-entropy loss to represent the error, and use dynamic decreasing learning rate and early stopping strategy to search for the optimal training cycle; Step S23: Train the finite element-deconvolutional neural network. Using the finite element analysis result database of the road structure generated in step S1, divide it into a training set and a test set. During the training process, a validation set is set to calculate the validation loss. The ratio of the sample size used in the training set to the test set is 8:2, and the sample size of the validation set is 10% of that in the training set.

5. The automatic iterative optimization method for pavement structure design based on machine learning according to claim 4, characterized in that, In step S21, constructing a finite element-deconvolutional neural network includes: The network input is a (6N+1)×1 vector, which includes the load axis and the thickness of N structural layers, material modulus, Poisson's ratio and three creep parameters, where N is the total number of pavement structural layers, and N is between 2 and 5. The input features are expanded to a 1024×1 feature vector through a fully connected layer and then reorganized to generate an initial feature matrix of 32×32×1. The initial feature matrix is ​​expanded into 64 36×36 feature matrices through a 5×5 convolution kernel and a 1×1 stride deconvolution operation. The feature matrix is ​​further upsampled by using a 3×3 convolution kernel and a 1×1 stride deconvolution operation to generate 64 38×38 feature matrices. The number of feature matrices is reduced by using a 3×3 convolution kernel and a 2×2 stride deconvolution operation, resulting in 32 77×77 feature matrices. The feature matrix size is further expanded by using a 3×3 convolution kernel and a 5×4 stride deconvolution operation to generate 32 feature matrices of 385×308. By using a 3×3 convolution kernel and a 1×1 stride deconvolution, the number of channels is compressed to 16, generating 16 feature matrices of 387×310. Finally, three 393×322 matrices are output through a 7×13 convolution kernel and a 1×1 stride deconvolution: the stress matrix, the strain matrix, and the plastic deformation matrix, corresponding to the stress, strain, and plastic deformation values ​​at different locations of the pavement structure.

6. The automatic iterative optimization method for pavement structure design based on machine learning according to claim 4, characterized in that, In step S23, each set of training or testing data includes 6N+1 road structure input parameters and 3 393×322 mechanical response matrix labels. Error is backpropagated by the loss value between the prediction matrix and the label matrix to achieve network training.

7. The automatic iterative optimization method for pavement structure design based on machine learning according to claim 1, characterized in that, Step S3 includes: Step S31: Determine the road structure parameters that are not in the pre-selection range, and determine the input parameters by interpolation or extrapolation methods; Step S32: Based on the pavement structure parameters that are not in the pre-selected range, input the finite element-deconvolutional neural network for forward prediction, output the stress, strain and plastic deformation matrices, and compare them with the finite element simulation results to calculate the prediction matrix error; In step S33, if the positive prediction result is within the allowable error range, proceed to step S4; if the error requirement is not met, return to step S2 to retrain.

8. The automatic iterative optimization method for pavement structure design based on machine learning according to claim 7, characterized in that, In step S32, the road material parameters input to the network's forward prediction are within 50% of the upper and lower limits of common asphalt material performance, used to simulate abnormal material performance; the load axle cycle range is increased to 1.5 million cycles, used to simulate extremely heavy loads. The prediction matrix error is calculated using Euclidean distance or Mahalanobis distance.

9. The automatic iterative optimization method for pavement structure design based on machine learning according to claim 1, characterized in that, Step S4 includes: Step S41: For the pavement structure design optimization problem, define the number of pavement structure layers, material range, and constraints. Simultaneously, set an optimization objective to minimize stress, strain, or plastic deformation at different locations, and construct the objective optimization function. ,variable The thickness of each layer; Step S42: A heuristic algorithm is used to search for the optimal solution in a loop. First, an initial solution set of road thickness combinations is randomly generated. The mechanical response under different thickness combinations is quickly calculated by the finite element-deconvolutional neural network in step S3. The solution set is ranked by constraint penalty and optimization objective function. The solution set is adjusted in a loop until the convergence condition is met and the optimization ends. Step S43: After optimization using the finite element-deconvolutional neural network and heuristic algorithm, the optimal combination of structural thicknesses and the mechanical response of each control point are obtained. Step S44: Finally, the optimal thickness combination is verified by conventional finite element analysis to determine that the mechanical response of each control point is within the range required by the specification, thereby obtaining the best combination of pavement structure layer thicknesses.

10. The automatic iterative optimization method for pavement structure design based on machine learning according to claim 9, characterized in that, In step S41, the objective optimization function is to minimize the permanent deformation at the load center location, the tensile strain at the bottom of the surface layer, or the tensile stress at the bottom of the base layer. In step S42, the constraints include the total thickness of the pavement structure, construction cost, and environmental emissions. In step S43, the heuristic algorithm is one of the following: genetic algorithm, ant colony algorithm, or simulated annealing algorithm.

Images