Data Augmentation Method for Asphalt Pavement Rutting Based on Radial Basis Neural Network

Abstract

The present invention discloses a method for expanding rutting data of asphalt pavement based on a radial basis neural network. First, the data of the performance detection of asphalt pavement and the influencing factors of the performance of asphalt pavement are collected, and the data is preprocessed. Then, according to the obtained data, feature generation is carried out, and the features highly correlated with rutting data are screened out according to the Pearson correlation coefficient. Subsequently, a radial basis neural network expansion model is constructed, the center of the Gaussian radial basis function is determined by the K-means algorithm, the radius of the Gaussian radial basis function is determined by the k-nearest neighbor algorithm, the weight matrix from the hidden layer to the output layer is determined, and the optimal model is achieved by adjusting the number of neurons in the hidden layer. The expanded data is predicted and verified. Finally, the data expansion model is extended to each pavement data, and based on its characteristic data, the expansion of rutting data of asphalt pavement based on a radial basis neural network is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of asphalt pavement performance prediction, and relates to a data augmentation method, in particular to a data augmentation method for asphalt pavement rutting based on a radial basis neural network. Background Art

[0002] In recent years, with the promotion of highway transportation by economic development, the types and quantities of vehicles have become increasingly diversified, and China's expressways have developed rapidly. However, due to the increase in traffic flow and traffic axle loads, traditional asphalt pavements have structural damage, mainly manifested as fatigue cracking and permanent deformation, with a short service life, and demolition and reconstruction have brought huge economic losses. In the future, in order to reduce costs such as maintenance, it will be more cost-effective to extend the design life of roads in busy traffic areas by at least 40 years without strengthening the structure.

[0003] However, due to limited and irregular data, the accuracy of prediction models based on neural networks is low, which limits periodic research. In 2017, in-situ load tests were carried out using RIOHTRACK to collect and study the evolution laws of multi-purpose performance under the conditions of nonlinear road structures and materials throughout their life cycles, and to verify and improve road structure design methods and materials. At different load levels, regular tests and data collection were carried out on 19 structural deflection basins with different stiffness levels, and their changes were analyzed. Among them, rutting is one of the important indicators for testing pavement damage. Therefore, establishing an accurate rutting prediction model can not only guide pavement design, but also provide an important basis for pavement maintenance and repair. Summary of the Invention

[0004] The present invention is precisely to establish a more accurate rutting prediction model, and provides a data augmentation method for asphalt pavement rutting based on a radial basis neural network. First, collect asphalt pavement performance detection data and data on factors affecting asphalt pavement performance, and preprocess the data; then, according to the obtained data, perform feature generation, and screen out features highly correlated with rutting data based on the Pearson correlation coefficient; subsequently, construct a radial basis neural network augmentation model, determine the centers of Gaussian radial basis functions by the K-means algorithm, determine the radii of Gaussian radial basis functions by the k-nearest neighbor algorithm, determine the weight matrix from the hidden layer to the output layer, and achieve the optimal model by adjusting the number of neurons in the hidden layer; perform prediction verification on the augmented data to illustrate the credibility and advantages of the model; finally, generalize the data augmentation model to various pavement data, and based on its feature data, realize data augmentation for asphalt pavement rutting based on a radial basis neural network.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a data augmentation method for asphalt pavement rutting based on a radial basis neural network, including the following steps:

[0006] S1, Data preprocessing: Collect the data of asphalt pavement performance detection and the data of influencing factors of asphalt pavement performance, and preprocess the data; the data of asphalt pavement performance detection is rut depth data, and the data of influencing factors of asphalt pavement performance include cycle, atmospheric temperature, surface temperature and cumulative axle load; the data preprocessing includes at least data screening and elimination, and data transformation.

[0007] S2, Feature generation: According to the data obtained in step S1, perform feature generation. The features are cycle, atmospheric temperature, surface temperature, cumulative axle load, single-load axle load and load change rate, and then screen out the features highly correlated with rut data according to the Pearson correlation coefficient.

[0008] S3, Construct a radial basis neural network expansion model: Divide the data set obtained in step S1 according to pavement structure, and construct a radial basis neural network expansion model; the neural network is a three-layer network, namely an input layer, a hidden layer and an output layer; in the model, the K-means algorithm determines the centers of Gaussian radial basis functions, where the number of all centers is the number of neurons in the hidden layer, the k-nearest neighbor algorithm determines the radius of the Gaussian radial basis function, determines the weight matrix from the hidden layer to the output layer, and then adjusts the number of neurons in the hidden layer to achieve the optimal model.

[0009] S4, Data expansion: Expand the rut data according to pavement structure.

[0010] S5, Implementation: Use the expanded data for regression estimation and short-term prediction of rut depth.

[0011] As an improvement of the present invention, in step S1, the data is screened and eliminated according to the cycle. When the feature data of a certain cycle is missing, the feature data is directly deleted; when the order of magnitude of some features is too large, the data is logarithmically compressed and transformed.

[0012] As another improvement of the present invention, the calculation formula of the Pearson correlation coefficient in step S2 is

[0013]

[0014] where x and y respectively represent two feature data.

[0015] As another improvement of the present invention, the pavement structure in step S3 is divided into six categories, namely thin AC semi-rigid matrix structure, ordinary semi-rigid matrix structure, rigid composite matrix structure, inverted structure, thick AC matrix structure and full-depth AC structure. Divide the data set according to the six pavement structures and construct a radial basis neural network expansion model.

[0016] As a further improvement of the present invention, the method for determining the center of the Gaussian radial basis function by using the K-means algorithm in step S3 is specifically as follows: randomly select k objects from n sample data as the initial clustering centers, calculate the distance from each sample to each clustering center respectively, assign the objects to the cluster with the nearest distance, after all data are assigned, recalculate the centers of the k clusters, compare with the centers of the k clusters obtained in the previous calculation, if the center of any cluster changes, recalculate the distance from each sample to each clustering center, otherwise terminate the algorithm and output the clustering result.

[0017] As a further improvement of the present invention, in step S3, the specific method for determining the radius of the Gaussian radial basis function by using the k-nearest neighbor algorithm is as follows: select the number k of neighbors to be considered, calculate the distance from each node to the center, sort them in ascending order according to the magnitude of the distance values, select the first K rows from the sorted array, and finally calculate the root mean square distance between the current cluster and its k nearest neighbors, which is the radius of the Gaussian radial basis function:

[0018]

[0019] As a further improvement of the present invention, the weight ω i from the hidden layer to the output layer satisfies the following equation

[0020]

[0021] The compact form is φW = y, where φ ∈ R n×N , W = (ω1, ω2, …, ω N ), T so the weight matrix from the hidden layer to the output layer is:

[0022] W = (φ T φ) -1 φ T y.

[0023] As a further improvement of the present invention, the optimal number of neurons in the hidden layer in step S3 is 4 to 6 times the number of features.

[0024] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1) It can avoid the uncertainty error caused by using field data, and can increase the data quantity, and further improve the accuracy of the prediction model based on the neural network; 2) It can generate relatively regular data, and can try some new algorithms to study the periodic characteristics of the data; 3) It can effectively reduce the measurement cost, and can perform low-density measurement on rut data and high-density measurement on easily measurable data such as temperature and axle load. Brief Description of the Drawings

[0025] Figure 1 Schematic diagram of the RIOHTRACK test section in Embodiment 2 of the present invention;

[0026] Figure 2 Neural network model diagram of the radial basis function neural network of the present invention;

[0027] Figure 3 Framework diagram of the asphalt pavement rutting data augmentation model based on the radial basis function neural network of the present invention. Specific implementation manners

[0028] The present invention will be further clarified below in conjunction with the accompanying drawings and specific implementation manners. It should be understood that the following specific implementation manners are only used to illustrate the present invention and not to limit the scope of the present invention.

[0029] The asphalt pavement rutting data augmentation method based on the radial basis neural network includes the following steps:

[0030] S1, data preprocessing: Collect the asphalt pavement performance detection data and the asphalt pavement performance influencing factor data, and preprocess the data; the asphalt pavement performance detection data is mainly rut depth data, and the asphalt pavement performance influencing factor data includes cycle, atmospheric temperature, surface temperature and cumulative axle load;

[0031] Collect the asphalt pavement performance detection data and the asphalt pavement performance influencing factor data over the years, and preprocess the collected data, mainly including data screening and elimination, data conversion, etc. Due to irresistible factors such as equipment and environment, the collected data often has situations such as data missing and duplication. Therefore, first screen the data according to the cycle and delete the duplicate data within the same cycle; for the situation where the characteristic data of a certain cycle is missing, directly delete it instead of filling it with the average value, aiming to avoid the supplementary data having a negative impact on the subsequent data augmentation model; finally, since the order of magnitude of some characteristics is very large, in order to prevent some important characteristics from being ignored due to the calculation of distance during data augmentation, in this embodiment, the data is logarithmically compressed and converted.

[0032] S2, feature generation: According to the data obtained in step S1, perform feature generation, and then screen out the features highly correlated with the rut data according to the Pearson correlation coefficient.

[0033] Recently, the rut prediction model based on the RIOHTrack mechanical monitoring data has well fitted the expression of the mechanical empirical constitutive equation (M_E model). This expression shows that rutting is affected by factors such as temperature, load, deflection, and material parameters. Based on this, in order to avoid the loss and repetition of features during the modeling process, in this embodiment, first, feature generation is performed on the existing data, and then the highly correlated features are removed according to the correlation coefficients obtained from the correlation analysis. The full-scale pavement loop is called RIOHTrack. As Figure 1 shown, 19 different forms of asphalt pavement structures (STR1-STR19) are laid on the main test section of the loop. So far, a continuous 5-year collection work has been carried out, accumulating relatively complete stress and strain response information inside different pavement structures. The existing characteristic data of the loop are period, atmospheric temperature, surface temperature, and cumulative axle load. Taking the pavement data STR1 as an example, six features are generated. The new features include single-loading axle load and load change rate; according to the Pearson correlation coefficient calculation formula

[0034]

[0035] where x and y respectively represent two data to be discussed, which can be rut data and characteristic data, or any two characteristic data. Calculate the correlation coefficients between each feature and between rut and feature. The present invention believes that if the correlation coefficient is higher than 0.8, it is considered that there is a high correlation between the two data. Finally, four important features for data expansion are selected, namely atmospheric temperature, cumulative axle load, single-loading axle load, and load change rate.

[0036] S3. Construct a radial basis neural network expansion model: Divide the data set obtained in step S1 into six categories according to the pavement structure, namely thin AC semi-rigid matrix structure, ordinary semi-rigid matrix structure, rigid composite matrix structure, inverted structure, thick AC matrix structure, and full-depth AC structure. Divide the data set and construct a radial basis neural network expansion model, as Figure 2 shown;

[0037] First, divide the dataset. Divide the dataset obtained in step S1 into six categories according to the pavement structure. Among them, the thin AC semi-rigid matrix structure includes STR 1–3, the ordinary semi-rigid matrix structure includes STR 6–9, the rigid composite matrix structure includes STR4 and STR 5, the inverted structure includes STR 10 and STR 12, the thick AC matrix structure includes STR 11, 13–17, and the full-depth AC structure includes STR 18 and STR 19. Select a representative dataset for each category and divide it at intervals, evenly into a training set and a test set. The training set is used to train the radial basis neural network model, continuously optimizing the number of neurons in the hidden layer until the optimal data augmentation model is obtained. The test set is used to test the radial basis neural network and observe its augmentation effect.

[0038] Furthermore, determine the centers of the Gaussian radial basis functions through the K-means algorithm. First, randomly select k objects from n sample data as the initial cluster centers; then calculate the distance from each sample to each cluster center respectively, and assign the object to the cluster with the closest distance; after all data are assigned, recalculate the centers of the k clusters; compare with the centers of the k clusters obtained in the previous calculation. If the center of any cluster changes, recalculate the distance from each sample to each cluster center, otherwise terminate the algorithm and output the clustering result.

[0039] Then, determine the radius of the Gaussian radial basis function through the K-nearest neighbor algorithm. First, select the number k of neighbors to consider, then calculate the distance from each node to the center, sort them in ascending order according to the size of the distance value, and select the first k rows from the sorted array; finally, calculate the root mean square distance between the current cluster and its k nearest neighbors.

[0040]

[0041] This value is the radius of the Gaussian radial basis function, where the distances involved in the calculation are all Euclidean distances.

[0042] Next, determine the weight matrix from the hidden layer to the output layer. After the parameters of the Gaussian radial basis function, the next step is to determine the weight ω from the hidden layer to the output layer i , which satisfies the following equation

[0043]

[0044] The compact form is φW = y, where φ ∈ R n×N , W = (ω1, ω2, …, ω N ) T , so the weight matrix W from the hidden layer to the output layer = (φ T φ) -1 φ T y.

[0045] Finally, parameter optimization. The number of neurons in the hidden layer is an adjustable parameter of the RBF neural network model. Using RMSE, MAE, MAPE, and R 2 as evaluation metrics to adjust the parameters until the augmentation result reaches the optimal.

[0046] S4. Data augmentation: Provide the rutting data of 19 types of structural pavements for augmentation;

[0047] Perform regression estimation and short-term prediction on the data before and after augmentation respectively, and analyze and verify the augmentation model by comparing the accuracies of the two.

[0048] S5. Implementation: Use the augmented data for regression estimation and short-term prediction of rut depth.

[0049] Regression estimation: For the features selected in S2, augment the data from 2017 to 2020, train the augmented data with a radial basis neural network, predict the rutting data in 2021 by regression, and compare it with the source data to obtain the final prediction error.

[0050] Short-term prediction: For the features selected in S2, augment the data from 2017 to 2020, use a time sliding window with a time window of 5 and a step size of 1 to fragment the original data, use the processed data as the training set to train the LSTM algorithm, obtain the short-term prediction data of rutting in 2021, and compare it with the source data to obtain the final prediction error.

[0051] Split the dataset. The 19 pavement datasets obtained in step S1 are divided at intervals and evenly divided into a training set and a test set. The training set is used to train the radial basis neural network model, continuously optimize the number of neurons in the hidden layer until the optimal data augmentation model is obtained, and the test set is used to test the radial basis neural network to observe its augmentation effect;

[0052] Initialize parameters. The weights of the RBF neural network model are initialized randomly. Through the neural network parameters of six types of pavement data, it is found that the optimal number of neurons in the hidden layer is about 4 to 6 times the number of features. Therefore, set the initial number of neurons to 20;

[0053] Parameter optimization. Using root mean square error RMSE, mean absolute error MSE, mean absolute percentage error MAPE, and R 2 as evaluation metrics, continuously adjust the parameters until the augmentation result reaches the optimal.

[0054] Example 2

[0055] The rutting data measured in this embodiment comes from the full-scale pavement test loop road project of the Ministry of Transport. The full-scale pavement loop includes 25 kinds of asphalt pavement structures, which are overall in a runway layout, with a total length of 2,039 meters, known as RIOHTrack. As Figure 1 shown, nearly 1,200 dynamic stress and strain sensors are buried in 19 different forms of asphalt pavement structures laid on the main test section of the loop road to collect the mechanical response state inside the asphalt pavement structure under the coupling action of load and environment in real time. So far, a continuous 5-year collection work has been carried out, accumulating relatively complete stress and strain response information inside different pavement structures.

[0056] In this embodiment, the measurement data of the full-scale pavement structures of 19 kinds of asphalt pavements are used as the data source. Each pavement has a total of 103 data samples. The goal is to seek the features related to the rut depth, and then based on these features, use the radial basis function neural network to effectively expand the rutting data. MSE, MAE, and R are used to reflect the difference between the true value and the predicted value. The smaller the index, the more accurate the prediction.

[0057] Table 1 Source data of rutting features

[0058]

[0059] In this embodiment, taking six main pavement structures as examples, for STR2, STR5, STR8, STR11, STR12, and STR18, a data expansion model is established based on the radial basis function neural network. The framework diagram of the asphalt pavement rutting data expansion model based on the radial basis function neural network is as Figure 3 shown. (1) Divide the data set: Each pavement structure data has 103 data, which are divided into a training set and a test set. Among them, 52 are used for training and 51 are used for testing. Part of the source data is shown in Table 1. (2) Initialize the parameters: The initialization of the weights of the radial basis function neural network model adopts the random initialization method. Through the neural network parameters of the six types of pavement data, it is found that the optimal number of neurons in the hidden layer is about 4 to 6 times the number of features. Therefore, the initial number of neurons is set to 20; (3) Parameter optimization: Using the root mean square error RMSE, mean absolute error MSE, mean absolute percentage error MAPE, and R 2 as the evaluation indexes, continuously adjust the parameters until the expansion result reaches the optimal. The number of neurons in the hidden layer of the radial basis neural network for the six types of pavements are 28, 22, 17, 30, 22, and 28 respectively. The test results show that when the expansion results of the six types of pavement structures are compared with the true values, the average RMSE is 3.95, and the average R 2It is 0.957. The experimental results of this method were compared with those of the classical neural network method, and the results showed that the method based on the radial basis function neural network is superior to the classical neural network method in data augmentation. Finally, through the example verification of rut data prediction, it is shown that the augmented data can indeed improve the prediction accuracy.

[0060] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For those of ordinary skill in the art of this technology, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements all fall within the protection scope of the claims of the present invention.

Claims

1. A method for expanding rut data of asphalt pavement based on a radial basis neural network, characterized in that It includes the following steps: S1. Data preprocessing: Collect the data of asphalt pavement performance detection and the data of influencing factors of asphalt pavement performance, and preprocess the data; the data of asphalt pavement performance detection is rut depth data, and the data of influencing factors of asphalt pavement performance at least includes cycle, atmospheric temperature, surface temperature and cumulative axle load; the data preprocessing at least includes data screening and elimination, and data transformation; S2. Feature generation: According to the data obtained in step S1, perform feature generation. The features include cycle, atmospheric temperature, surface temperature, cumulative axle load, single-loading axle load and load change rate, and then screen out the features highly correlated with rut data according to the Pearson correlation coefficient; S3. Construct a radial basis neural network expansion model: Divide the data set obtained in step S1 according to pavement structures, and construct a radial basis neural network expansion model. The neural network is a three-layer network, namely an input layer, a hidden layer and an output layer; in the model, the K-means algorithm determines the centers of Gaussian radial basis functions, where the number of all centers is the number of neurons in the hidden layer, the k-nearest neighbor algorithm determines the radius of Gaussian radial basis functions, determines the weight matrix from the hidden layer to the output layer, and then adjusts the number of neurons in the hidden layer to achieve the optimal model; S4. Data expansion: Expand the rut data according to pavement structure types; S5. Implementation: Use the expanded data for regression estimation and short-term prediction of rut depth.

2. The method for augmenting rut data of asphalt pavement based on a radial basis neural network according to claim 1, wherein: In step S1, the data is screened and eliminated according to the cycle. When the feature data of a certain cycle is missing, the feature data is directly deleted; when the order of magnitude of some features is too large, the data is logarithmically compressed and transformed.

3. The method for augmenting rut data of asphalt pavement based on a radial basis neural network according to claim 2, wherein: The calculation formula of the Pearson correlation coefficient in step S2 is where x and y respectively represent two feature data; When the Pearson correlation coefficient is higher than 0.8, it is considered that there is a high correlation between the two data, and then the features highly correlated with rut data are screened out.

4. The method for augmenting rut data of asphalt pavement based on a radial basis neural network according to claim 3, wherein: In step S3, the pavement structures are divided into six categories, namely thin AC semi-rigid matrix structure, ordinary semi-rigid matrix structure, rigid composite matrix structure, inverted structure, thick AC matrix structure and full-depth AC structure. Divide the data set according to the six pavement structures and construct a radial basis neural network expansion model.

5. The method for augmenting rut data of asphalt pavement based on a radial basis neural network according to claim 3, wherein: The method of using the K-means algorithm to determine the centers of Gaussian radial basis functions in step S3 is specifically as follows: Randomly select k objects from n sample data as the initial clustering centers, calculate the distances from each sample to each clustering center respectively, and assign the objects to the nearest cluster. After all data are assigned, recalculate the centers of the k clusters, compare with the centers of the k clusters obtained in the previous calculation. If the center of any cluster changes, recalculate the distances from each sample to each clustering center, otherwise terminate the algorithm and output the clustering result.

6. The method for expanding rut data of asphalt pavement based on radial basis neural network according to claim 3, wherein: In the step S3, the specific method for determining the radius of the Gaussian radial basis function by using the k-nearest neighbor algorithm is as follows: select the number k of neighbors to be considered, calculate the distance between each node and the center, sort them in ascending order according to the magnitude of the distance value, select the first K rows from the sorted array, and finally calculate the root mean square distance between the current cluster and its k nearest neighbors, which is the radius of the Gaussian radial basis function:

7. The method for augmenting rut data of asphalt pavement based on a radial basis neural network according to claim 3, characterized in that: The weight ω from the hidden layer to the output layer in the step S3 i , satisfies the following equation The compact form is φW = y, where φ ∈ R n×N , W = (ω1, ω2, …, ω N ) T , so the weight matrix from the hidden layer to the output layer is: W = (φ T φ) -1 φ T y.

8. The method for augmenting rut data of asphalt pavement based on radial basis neural network according to claim 4 or 5 or 6 or 7, characterized in that: In the step S3, the optimal number of neurons in the hidden layer is 4 to 6 times the number of features.

Images