A meta-learning approach for predicting asphalt pavement deflection basin area
0 Abstract
To address the urgent need for accurate pavement performance modeling in pavement design, this study proposes a meta-learning-based few-shot learning method for predicting the Deflection Basin Area (DBA) of asphalt pavements. The method utilizes features such as pavement temperature and load pressure, and applies cyclic DBA data from various pavement types subjected to different pressures. The objective is to predict the trend of DBA changes over cycles at a specific pressure. By leveraging pre-training on diverse pavement datasets, the proposed meta-learning model reduces the training data required for target pavement DBA prediction, enabling better generalization to the target pavement. This approach enhances DBA prediction accuracy even with a small sample size. Compared to traditional machine learning and pre-training methods using data from a single pavement type, the proposed method achieves a Mean Square Error of 13.26 and a Mean Absolute Error of 2.85, demonstrating superior performance. Furthermore, it achieves high prediction accuracy with fewer iterations. Overall, the proposed method effectively predicts DBA across various pavement structures with a few data.
Keywords
1. INTRODUCTION
Highway infrastructure in China has undergone rapid development in recent years, with asphalt pavement (AP) structures being widely adopted across roads of all grades. The majority of these structures use semi-rigid base pavement designs[1]. However, the long-term service of operating highways, subjected to traffic loads and environmental factors, often leads to various degrees of structural damage, including cracking and subsidence. In this context, pavement deflection has emerged as a critical parameter for assessing the bearing capacity of subgrade pavements and plays a pivotal role in evaluating overall pavement conditions[2].
The Falling Weight Deflectometer (FWD) has become the standard tool for non-destructive pavement testing. While FWD detection results contain a wealth of pavement characteristic information, current engineering practices typically rely solely on the maximum deflection at the load center to characterize the surface deflection of the structural layer. This approach, however, underutilizes the comprehensive data provided by FWD testing[3]. Recognizing this limitation, researchers have increasingly focused on establishing correlations between the pavement Deflection Basin Area (DBA) and the structure's bearing capacity. This approach aims to predict pavement usage indicators and evaluate overall structural integrity more comprehensively[4-7]. The DBA, which considers the entire deflection profile rather than a single point, offers a more holistic representation of pavement response to loading.
However, the collection of FWD testing data presents significant challenges. The process is resource-intensive, time-consuming, and costly, making it difficult to conduct repeated measurements across large road networks. These limitations underscore the need for predictive models that can accurately estimate the DBA for various pavement structures without extensive field testing.
Recent advancements in artificial intelligence (AI) have opened new avenues for addressing this challenge. AI methods have been successfully applied across various domains of traffic engineering[8-10], demonstrating their potential for complex prediction tasks. In the context of pavement engineering, AI offers the promise of high-accuracy DBA prediction, potentially revolutionizing how we assess and manage road infrastructure. The motivation behind this research lies in the critical need for more efficient and cost-effective methods of pavement assessment. By developing accurate predictive models for DBA, we aim to enhance the ability of road authorities to monitor pavement conditions, plan maintenance activities, and optimize the allocation of resources. This approach promises to reduce the reliance on extensive field testing and provide a more nuanced understanding of pavement behavior under various conditions.
Despite the potential benefits, achieving high-accuracy DBA prediction remains a significant challenge. This research seeks to address this gap by exploring advanced AI techniques and their application to pavement engineering. By doing so, we aim to contribute to developing more robust and reliable methods for assessing pavement structural integrity, ultimately leading to improved road infrastructure management and longevity. This research aims to develop and design a meta-learning-based prediction model for DBA of APs, which uses measured cyclic data of pavement DBA under different pressures to predict the trend of rebound deflection basin area changes throughout all cycles at a specific pressure. The contributions of this research are as follows.
• This study applies meta-learning algorithms to predict the AP DBA over the entire cycle at a specific pressure for the first time, achieving better performance than traditional machine learning algorithms as measured by Mean Square Error (MSE) and Mean Absolute Error (MAE). These results demonstrate the effectiveness of meta-learning algorithms in small sample regression tasks.
• The proposed meta-learning prediction approach can reduce the need on the amount of target pavement data by reusing relevant data acquired from other pavement structures for training, which is crucial given the challenge of obtaining long-term cyclic data for target pavements in the early stages.
• Compared with the pre-training-based transfer learning method, the proposed meta-learning prediction uses a limited dataset specific to the target pavement to fine-tune the parameters from pre-training. This strategy not only ensures stability, but also produces better performance in predicting the DBA.
2. RELATED WORK
Machine learning methods can be widely used for data regression prediction tasks[11], and in recent years, they have been increasingly applied to pavement performance prediction tasks. The primary hurdle in pavement performance prediction arises from extracting crucial spatiotemporal features from multidimensional coupled data. Early work, including Artificial Neural Network (ANN)[12], stacked autoencoders[13], and Long Short-term Memory (LSTM)[14], predominantly focused on extracting temporal features but encountered challenges in unveiling latent spatial features within traffic data, leading to suboptimal performance. Yang et al.[15] use K-nearest neighbor to characterize crack type and crack width information derived from FWD test data for applied statistical techniques, including ANN and multiple nonlinear regression, to proficiently rut distress in AP using an AP analyzer[16].
Recently, sophisticated learning algorithms have been integrated into engineering modeling applications. Tang et al. use genetic algorithms to develop a reverse analysis program that combines finite element analysis and population-based optimization techniques to infer the modulus of the pavement layer[17]. Mabrouk et al. use ANNs to calculate the pavement layer modulus as a function of traffic speed deviation, thus predicting the pavement modulus[18]. Zhang et al. propose a brand-new framework for predicting time series of ruts, based on multi-level discrete wavelet decomposition and multivariate transfer entropy for feature selection, to achieve higher prediction accuracy[19]. Li et al. compare random forest and gradient boosting regression methods for predicting the pavement Asphalt Concrete (AC) layer and use grid search and cross-validation for optimization[20].
In addition, the frequent collection of pavement structure data is very time-consuming, disrupts traffic flow, and is costly, making it hard to duplicate. These factors might contribute to the oversight of pavement structure aspects in maintenance or repair decisions. Consequently, researchers have invested considerable effort in identifying relatively expedient alternative analysis methods to overcome these limitations. Li et al. propose a Chaotic Particle Swarm Optimization (CPSO) to optimize the Extreme Gradient Boosting (XGBoost) model, decreasing the frequency of deflection tests while maintaining estimation accuracy[21]. Zhang et al. propose a model for predicting ruts based on multi-source transfer entropy and graph neural networks, which can adapt to sufficient predictive performance and the generalization ability of various complex pavement design data[22]. Shen et al. employ LSTM to construct a predictive model for estimating the technical condition score of bridge components[23].
Nevertheless, traditional machine learning methods still face a practical challenge: the accuracy of deep learning regression prediction algorithms heavily depends on the availability of a substantial amount of training data[24]. For tasks such as pavement performance prediction, a large volume of pavement structure data is required, which presents a significant drawback. Collecting such extensive data is time-consuming and resource-intensive, but it remains essential for training deep learning models[25,26]. Furthermore, if the pavement structure used for training does not align with the target pavement structure, retraining the model becomes necessary, which again requires considerable time and effort. These issues highlight a major limitation of traditional machine learning methods: developing accurate performance prediction models for various pavement structures demands both significant time and financial investment.
Given these challenges, alternative approaches such as Model-Agnostic Meta-Learning (MAML)[27] have gained attention in the field. MAML is a powerful meta-learning algorithm designed for few-shot learning, and it has demonstrated superior performance in addressing problems with limited data[28-32]. As a member of optimization-based meta-learning algorithms, MAML differs from traditional methods by enabling models to learn in a way that generalizes quickly to new tasks with minimal data. Other meta-learning algorithms have introduced various modifications for learning weights in task-specific classifiers[33,34]. For example, methods in[35-38] first learn a function to embed the support set and target examples of a few-shot task, and then use the test support set to fine-tune task-specific weights for embedding the target examples. On the basis of meta-learning, Aguiar et al. propose selecting meta-features for extracting optimal dataset descriptions, which enhances multi-target regression with high predictive accuracy[39]. Additionally, Jeong et al. introduce a meta-learning approach for State-Of-Charge (SOC) estimation in batteries, aiming to reduce the amount of target data needed for training by leveraging deep learning[40]. By overcoming the data limitations inherent in traditional machine learning methods, meta-learning provides a more efficient framework for addressing pavement performance prediction tasks, particularly in situations with scarce or inconsistent data.
Building upon optimization-based meta-learning algorithms, MAML, in particular, has exerted significant influence, inspiring numerous direct extensions in[41-44]. Many of these extensions heavily depend on the foundational structure of the MAML algorithm, encompassing the outer loop (for meta-training) and the inner loop (for task-specific adaptation). However, scant prior research has thoroughly examined the reasons for the success of this core aspect of MAML. To investigate how and why MAML achieves effective few-shot learning[45-47], employs analytical tools such as centered kernel alignment to scrutinize the neural network representations learned by the MAML. This analysis also highlights the proficiency of this algorithm in acquiring practical few-shot learning features.
Meta-learning is sensitive to the choice of initial model parameters. In certain situations, different initial parameters may result in significant performance variations of the model across various tasks. Additionally, the performance of meta-learning may be affected by noise and distribution changes in the input data. When there is considerable noise or shifts in the data distribution within the meta-learning tasks, it can decline the model's generalization performance.
In addition, some robust meta-learning algorithms have been developed for noise rejection. The most effective robust training methods include adversarial training and robustness which establishes a theoretical foundation between accuracy and robustness, and many of their variants such as fast adversarial training methods[48-50], semi-supervised robust training[51,52], adversarial transfer learning. Furthermore, recent research[53-55] has explored the transferability of robustness within the realms of transfer learning and representation learning. Nevertheless, the conventional robust training methods mentioned earlier need to be better suited for MAML in few-shot learning, given the dual optimization characteristics inherent in MAML.
In summary, meta-learning exhibits significant advantages over traditional machine-learning methods when handling small-sample classification and regression tasks. Firstly, meta-learning achieves rapid model adaptation by learning across multiple relevant tasks, enabling it to adjust its parameters to accommodate new tasks quickly. Secondly, meta-learning enhances sample efficiency by acquiring general task-related knowledge, allowing the model to utilize limited samples for learning more effectively. Thirdly, emphasizing the transfer learning concept in meta-learning enables models to share knowledge between tasks, facilitating robust generalization performance. These capabilities represent a notable advantage that traditional machine-learning algorithms lack.
3. FUNDAMENTAL ALGORITHMS
3.1. Gradient boosting decision trees and random forests
Breiman[56] proposed the Random Forests (RF) algorithm, which builds upon the Bagging algorithm by incorporating Bootstrap sampling to construct multiple decision trees. The final prediction of the random forest is made through a majority voting mechanism. The use of Bootstrap sampling enhances the algorithm's accuracy and helps mitigate the overfitting issues commonly associated with single decision trees. The structure of RF is shown in Figure 1.
Figure 1. The structure of RF.
The Gradient Boosting Decision Tree (GBDT)[57] is an ensemble machine learning algorithm based on the Boosting strategy proposed by Friedman. The core idea is to sequentially train multiple weak learners to improve performance iteratively. In each iteration, except for the first decision tree, the objective is to minimize the loss function of the current learner by updating it in the direction of the gradient, ensuring that the loss function decreases with each step. Through continuous iterations, the residuals approach zero, and the results of all trees are combined to generate the final prediction. The specific implementation process of GBDT is as follows:
(1) Initialize weak learner
where
(2) Calculate negative gradient
For each tree
where
The obtained residual is used as the true value of the new sample, and the data
Update the strong learner, then we have
Get the final learner
3.2. Convolutional neural networks for solving regression problems
The network structure of Convolutional Neural Networks (CNNs) consists of an input layer, convolutional layers, pooling layers, and output layers, among others. CNNs are widely used for image classification tasks due to their ability to effectively capture spatial hierarchies in image data. However, when adapted for regression tasks involving numerical data, significant differences arise in various aspects, such as input data handling, the design of convolution and pooling layers, output layer configuration, and the choice of loss function.
In regression tasks, the input data typically takes the form of a numerical feature vector or matrix, denoted as
where
Activation functions introduce non-linear transformations, increasing the expressive ability of the model. Commonly used activation functions include the rectified linear unit (ReLU) function. The output of the activation function is given by:
where
Fully connected layers flatten the output of the pooling layer into a vector and perform linear transformations through matrix multiplication and bias terms. The output of the fully connected layer is given as
where
Finally, the output layer maps the output of the fully connected layer to the range of the predicted values through linear transformations and activation functions, resulting in the output
where
In regression tasks, the MSE is a commonly employed loss function. The network parameters are then updated through the backpropagation algorithm based on this loss function to minimize the overall loss. This iterative process allows the CNN algorithm to learn the relationship between input features and regression targets, enabling it to make numerical regression predictions.
Traditional machine learning methods rely on a substantial amount of data for training. When faced with small sample data, it is often unable to train a model with excellent performance. The model parameter update process for the service performance prediction of different pavement structures is shown in Figure 2, where each type of pavement structure must be trained separately as a model to perform regression tasks.
Figure 2. General deep-learning training methods.
3.3. Pre-training for the prediction of pavement deflection basin area
Pre-training method is an effective approach for few-shot learning, which is an unsupervised learning method. Before predicting the DBA of the target pavement structure data, we use other types of pavement surface data for the usual optimization pre-training. This approach is applicable to deep learning models trained through gradient descent. The comprehensive pre-training process is given in
where
Figure 3. Pre-training methods.
The purpose of pre-training models is to determine the optimal initial parameters
3.4. MAML for the prediction of pavement deflection basin area
The meta-learning algorithm stands out as one of the most effective approaches for addressing few-shot learning challenges. Figure 4 shows the overall process of using meta-learning methods to predict the DBA. The meta-model leverages the related datasets of 18 pavement structures, excluding the target pavement structure, for the training task, while the target pavement data serves as the testing task. Furthermore, each task is subdivided into a query set and a support set, utilized for training and testing in the inner loop, respectively. Algorithm 1 showcases the entire pre-training procedure of meta-learning.
Figure 4. Meta learning algorithm flowchart
First, meta-learning initializes a model with random parameters
where
For each class of AP data, the data is further divided into
In the outer loop, the MAML algorithm improves the initial parameters of the model by learning through inner loop iterations on multiple small-sample tasks, enabling it to quickly adapt to new tasks. For each update in the inner loop, the obtained parameters are used as the initial parameters of the new model. The model is then tested using the corresponding test set
Using the sum of the loss
During the prediction phase, we use a portion of the target pavement data as training data (
Figure 5. Meta learning method for the performance prediction of the pavement.
Algorithm 1 Meta-learning Input:
1: Initialize
2: Divide
3: while not done do
4: for all
5: Compute adapted parameters :
end for
6: Calculate the loss value :
7: Update
end while
As observed, in the training process of the MAML model, compared to the impact of
4. RIOHTRACK STRUCTURE AND DATA SET
4.1. RIOHTrack structure
The data used by the meta-learning model in this experiment is the AP DBA. The measured data about the DBA is derived from the Ministry of Transport's full-scale pavement test loop project. This project is situated in Beijing and encompasses 25 AP structures. It has a full-scale on-site acceleration road test track called RIOHTrack, which is 2.038 km long. The arrangement of these pavement structures is illustrated in Figure 6.
Figure 6. RIOHTRACK test loop pavement.
The data used in this article was measured from different pavement structures. A total of 19 primary experimental pavement structures were established on the test circuit to investigate and compare the long-term performance and evolution of AC structures with varying combinations of structural stiffness. The asphalt concrete layer thickness for these pavement structures ranges from 12 to 48 cm (or 52 cm), covering the spectrum of asphalt concrete layer thicknesses found in highways across China, including the thickness of flexible base layers for thick AP.
4.2. Acquisition of datasets
The DBA is measured using a FWD, which mainly consists of a heavyweight (falling weight) and a measuring instrument (usually a level or optical instrument). The FWD is dropped freely from a certain height, causing it to impact the pavement. After that, the pavement undergoes a slight deformation, forming a depression area, the settlement basin. Based on the shape and size of the observed depression area, mathematical formulas or calculation methods can be used to determine the area of the settlement basin.
The measured settlement basin area data has the following characteristics: axle load times, load level, pavement temperature, atmospheric temperature, etc. The data used in the article comes from the DBA measured at pressures of 5, 7, 9, and 11 tons in multiple cycles, respectively. Taking the data of STR1 pavement structure under a pressure of 5 tons as an example, the partial data of the settlement basin area measured by the FWD are shown in Table 1. These variables were selected based on their direct relevance to pavement performance and degradation. Axle load times and load level are crucial as traffic load frequency and intensity affect pavement durability. Pavement temperature influences the stiffness and cracking susceptibility of asphalt, while atmospheric temperature affects thermal expansion and environmental wear, all contributing to the pavement's long-term performance.
The pavement feature data(part)
| Cycle number | Axle load times | Load level | Pavement temperature (℃) | Atmospheric temperature (℃) |
| N1 | 4.4533 | 56.47 | 12.77 | 9.17 |
| N2 | 5.0184 | 56.47 | 8.05 | 11.34 |
| N3 | 5.4729 | 56.47 | 11.66 | 7.81 |
| N4 | 5.6846 | 56.47 | 26.92 | 21.85 |
| N5 | 5.8265 | 56.47 | 30.2 | 25.32 |
| … | … | … | … | … |
| N111 | 7.7491 | 48.09 | 1.75 | 3.34 |
This text leverages measurement data from the pavement of 19 types of AP as the primary data source. The objective is to regressively predict the area of a particular pavement type's DBA. Due to the brief construction time of the pavement, the data is limited. The deflection basin measurement data is categorized under four distinct pressure conditions, amounting to only 1,792 data cycles. To address the scarcity, we utilize the data under three pressure conditions as the training set to predict pavement data under the remaining pressure conditions. Traditional machine learning methods are less effective for this regression task due to the limited dataset. To overcome the challenge, this paper employs a meta-learning method, using an additional 18 pavement datasets for meta-training, effectively expanding the dataset. The trained meta-learning model is subsequently used to predict the target pavement, achieving regression prediction of the DBA with only a small quantity of training data. Model evaluation is performed using MSE, Root Mean Squared Error (RMSE), and MAE. These metrics gauge the algorithm's performance by reflecting the differences between actual and predicted values, serving as commonly used performance indicators in regression tasks. Smaller values of these indicators indicate more accurate predictions. The metrics are defined in:
5. EXPERIMENTAL RESULTS
This section uses meta-learning to experimentally estimate the DBA of pavement to verify its estimation performance compared to pre-training. It also compares the performance of pre-training with other deep learning models that do not use pre-training. It verifies the impact of pre-training on predicting the DBA of the pavements. The purpose of the experiments in this paper is to utilize measured cyclic data of pavement DBA under different pressures to predict the trend of rebound DBA changes and its variation trend throughout all cycles at a specific pressure. Specifically, the training involves using DBA cyclic data under pressures of 5, 9, and 11 tons, and the goal is to predict the DBA and its variation trend over the entire cycle at a pressure of 7 tons.
5.1. Performance after a small number of iterations
A meta-learning prediction model trained on multiple pavement data (excluding the target pavement) can predict the target pavement DBA by fine-tuning a small quantity of pavement data in several gradient steps. To assess the effectiveness of meta-learning in DBA prediction, this study conducted a performance comparison based on gradient steps. A limited dataset specific to the target pavement was employed to fine-tune pre-trained DBA prediction models constructed using DNNs. In addition, this study also uses calibrated linear unit activation functions and the Adam optimizer. Multiple error measurement indicators, including the coefficient of determination (
Comparison of model performance after 10 to 100 iterations
| Iterations | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
| MSE | 172.38 | 107.33 | 106.61 | 55.66 | 68.56 | 36.04 | 63.76 | 46.35 | 49.52 | 40.10 |
| MAE | 9.12 | 9.41 | 8.17 | 5.65 | 6.62 | 4.74 | 6.17 | 5.09 | 5.41 | 4.73 |
Table 2 illustrates a noteworthy reduction in MSE and MAE after ten to 100 iterations, underscoring that DNN models with meta-learning can rapidly adapt to accurately predict the pavement DBA.
5.2. Prediction of the area of pavement deflection basin
The CNN model used in this experiment consists of one convolutional layer and two fully connected layers. The convolutional layer has 1 input channel, 10 output channels, and a kernel size of 2. The first linear layer has an input dimension of 10 and an output dimension of 10, containing ten neurons. The second linear layer has an input dimension of 10 and an output dimension of 1, containing 1 neuron. The learning rate is 0.001. The GBDT model uses a parameter
Model Parameters Summary
| Model | Parameter | Value |
| CNN | Convolutional layer | 1 input channel, 10 output channels, kernel size = 2 |
| First linear layer | Input dimension = 10, Output dimension = 10 | |
| Second linear layer | Input dimension = 10, Output dimension = 1 | |
| Learning rate | 0.001 | |
| GBDT | Number of estimators | 500 |
| Learning rate | 0.01 | |
| RF | Number of estimators | 200 |
| Loss function | MSE (Mean Squared Error) | |
| Max depth | none | |
| Min samples split | 2 | |
| DNN | Number of neurons | 40 (3 hidden layers) |
The performance of meta-learning in predicting the area of deflection basins using data from different pavement structures. To test the exactitude of meta-learning in predicting the area of deflection basins, experiments are conducted on pre-training data involving various pavement conditions without considering the similarity between targets. Table 4 summarizes the DBA estimation errors for each method. As observed, among the studied methods, the deflection basin area estimation using meta-learning achieves the most accurate results (MSE = 13.262, MAE = 2.854) despite using a pre-trained pavement structure different from the target pavement. Among the compared ML methods, the RF model has the least accurate prediction results, with MSE = 372.982 and MAE = 17.161, as it did not receive sufficient training (only using 424 target data points).
Comparison of model performance
| Model | MSE | MAE | |
| RF | 372.982 | 17.161 | - |
| GBDT | 228.281 | 14.651 | - |
| CNN | 151.741 | 10.144 | 0.745 |
| Pre-training | 16.645 | 3.253 | 0.902 |
| MAML | 13.262 | 2.854 | 0.922 |
Taking the STR19 road as an example, the MAML algorithm uses the data of STR19 as the target road and performs meta-training on other roads. We intend to forecast the DBA of STR19 within one cycle. The MAML model uses parameters of 0.01 for the inner loop learning rate, 0.001 for the outer loop learning rate, and 1 for the inner steps. The experimental results are shown in Figure 7.
Figure 7. (A) The training loss values of the MAML model, (B) The model prediction performance after different steps.
5.3. Predictions with various pavement types
This section outlines the prediction performance of meta-learning when utilizing pavement data with diverse structures. To validate the exactitude of meta-learning in predicting the DBA, the subsequent experiments are conducted under the condition that the meta-training data is composed of pavement data with diverse structures without considering the similarity between targets.
This experiment aims to verify that the meta-learning algorithm can have a good estimation performance when facing pavement data with different structures rather than being only applicable to specific data. The specific process of the experiment is similar to Section 5.2. The experiment is conducted on 19 pavements, with one target structure to be predicted and the other 18 pavements as the training data for meta-learning.
Table 5 lists the DBA prediction results for 18 different pavement structures. The MSE and MAE are significantly lower than those of the CNN, GBDT, and other models used in the comparative experiments. The estimation performance of the meta-learning algorithm is superior. Apart from the different data on the target pavement, the same conditions are applied in this experiment. As shown in Figure 8A-R, the fitting and estimation effects for different structures of the pavement are generally sound. Therefore, it is concluded that the meta-learning method could achieve good estimation performance for different types of pavement structures.
Comparison of model performance
| Pavement | STR1 | STR2 | STR3 | STR4 | STR5 | STR6 | STR7 | STR8 | STR9 |
| MSE | 20.132 | 13.830 | 57.710 | 63.206 | 50.844 | 63.216 | 60.360 | 14.453 | 36.942 |
| MAE | 3.394 | 2.805 | 5.677 | 3.680 | 3.757 | 4.622 | 3.895 | 3.075 | 4.765 |
| 0.933 | 0.843 | 0.791 | 0.622 | 0.604 | 0.630 | 0.696 | 0.855 | 0.776 | |
| Pavement | STR10 | STR11 | STR12 | STR13 | STR14 | STR15 | STR16 | STR17 | STR18 |
| MSE | 56.620 | 11.873 | 52.740 | 50.941 | 69.873 | 20.132 | 39.535 | 23.012 | 19.111 |
| MAE | 5.658 | 2.788 | 5.480 | 4.917 | 5.045 | 3.394 | 4.946 | 3.863 | 5.335 |
| 0.840 | 0.888 | 0.644 | 0.781 | 0.787 | 0.933 | 0.895 | 0.885 | 0.914 |
Figure 8. The results of estimating the deflection basin area of the remaining 18 pavements through meta-learning.
6. CONCLUSIONS
This paper proposes a meta-learning-based approach to predict the DBA of AP. By leveraging data from multiple pavement structures, excluding the target pavement, the proposed method significantly reduces the required training data volume, enabling accurate DBA prediction with small sample sizes. Traditional machine learning methods, which rely on training with data from a single pavement type, are less effective in estimating the DBA of the target pavement. In contrast, the proposed meta-learning method effectively reuses data from different pavement structures to pre-train deep learning models, improving the model’s generalization capabilities.
Through experiments, we demonstrated that the meta-learning approach outperforms traditional machine learning, neural networks, and pre-training methods in predicting the DBA of APs. The results show that the proposed method is effective and resource-efficient for estimating pavement performance, especially in the early stages of pavement development.
However, the study also has limitations. One potential limitation is the impact of sample size on the model’s prediction accuracy. Further research could explore how variations in sample size affect model performance and investigate ways to mitigate the effects of limited data. Additionally, the applicability of the proposed method to different pavement types and conditions remains an area for future exploration. Future studies could also focus on extending the model’s capabilities to include real-time data integration and testing in various environmental conditions.
DECLARATIONS
Authors’ contributions
Methodology, validation, visualization, and writing-original draft: Li Z, Jin X
Conceptualization, writing-reviewing, supervision, and editing: Shi X, Cao J
Availability of data and materials
The data cannot be shared publicly as the partner (company) does not permit public disclosure. It is available from the corresponding author upon reasonable request.
Financial support and sponsorship
This work was supported by the National Key Research and Development Project of China under Grant (No. 2020YFA0714300) and the Open Project of Nanjing Modern Multimodal Transportation Laboratory (MTF2023004).
Conflicts of interest
Cao J is an Advisory Board Member of the journal Complex Engineering Systems. Shi X is an Junior Editorial Board Member of the same journal. Cao J and Shi X were not involved in any steps of editorial processing, notably including the selection of reviewers, manuscript handling and decision-making, while the other authors have declared that they have no conflicts of interest.
Ethical approval and consent to participate
Not applicable.
Consent for publication
Not applicable.
Copyright
© The Author(s) 2024.
REFERENCES
1. Wang X. Design of pavement structure and material for full-scale test track. J Highw Transp Res Dev 2017;34:30-7.
2. Fu G, Zhao Y, Zhou C, Liu W. Determination of effective frequency range excited by falling weight deflectometer loading history for asphalt pavement. Constr Build Mater 2020;235:117792.
3. Habbouche J, Hajj EY, Sebaaly PE, Morian NE. Damage assessment for M-E rehabilitation design of modified asphalt pavements: challenges and findings. Transp Res Rec 2018;2672:228-41.
4. Huang W, Liang S, Wei Y. Surface deflection-based reliability analysis of asphalt pavement design. Sci China Technol Sci 2020;63:1824-36.
5. Vyas V, Singh AP, Srivastava A. Prediction of asphalt pavement condition using FWD deflection basin parameters and artificial neural networks. Road Mater Pavement Des 2021;22:2748-66.
6. Liu H, Cao J, Huang W, Shi X, Zhou X, Li Z. A multidimensional framework for asphalt pavement evaluation based on multilayer network representation learning: a case study in RIOHTrack. Expert Syst Appl 2024;237:121370.
7. Wang W, Zhang H, Li T, et al. An interpretable model for short term traffic flow prediction. Math Comput Simul 2019;171:264-78.
8. Li Z, Korovin I, Shi X, Gorbachev S, Gorbacheva N, Huang W. A data-driven rutting depth short-time prediction model with metaheuristic optimization for asphalt pavements based on RIOHTrack. IEEE/CAA J Autom Sin 2023;10:1918-32.
9. Li Z, Cao J, Shi X, Huang W. QPSO-AHES-RC: a hybrid learning model for short-term traffic flow prediction. Soft Comput 2023;97:9347-66.
10. Yang Y, Li Z, Chen J, Liu Z, Cao J. TRELM-DROP: an impavement non-iterative algorithm for traffic flow forecast. Physica A 2024;633:129337.
11. Pan JS, Zhang Z, Chu SC, Zhang SQ, Wu JMT. A parallel compact marine predators algorithm applied in time series prediction of Backpropagation neural network (BNN) and engineering optimization. Math Comput Simulat 2024;220:65-88.
12. Saltan M, Terzi S. Modeling deflection basin using artificial neural networks with cross-validation technique in backcalculating flexible pavement layer moduli. Adv Eng Softw 2008;39:588-92.
13. Lv Y, Duan Y, Kang W, Li Z, Wang FY. Traffic flow prediction with big data: a deep learning approach. IEEE Trans Intell Transp Syst 2015;16:865-73.
14. Ma X, Tao Z, Wang Y, Yu H, Wang Y. Long short-term memory neural network for traffic speed prediction using remote microwave sensor data. Transp Res Part C Emer Technol 2015;54:187-97.
15. Yang Q, Deng Y. Evaluation of cracking in asphalt pavement with stabilized base course based on statistical pattern recognition. Int J Pavement Eng 2019;20:417-24.
16. Karballaeezadeh N, Ghasemzadeh Tehrani H, Mohammadzadeh Shadmehri D, Shamshirband S. Estimation of flexible pavement structural capacity using machine learning techniques. Front Struct Civ Eng 2020;14:1083-96.
17. Tang X, Yang X. Inverse analysis of pavement structural properties based on dynamic finite element modeling and genetic algorithm. Int J Transp Sci Technol 2013;2:15-30.
18. Mabrouk GM, Elbagalati OS, Dessouky S, Fuentes L, Walubita LF. Using ANN modeling for pavement layer moduli backcalculation as a function of traffic speed deflections. Constr Build Mater 2022;315:125736.
19. Zhang J, Cao J, Huang W, Shi X, Ji X, Zhou X. A hybrid framework for asphalt pavement rutting prediction modeling and influencing factors analysis based on multilevel wavelet decomposition and transfer entropy. Appl Math Model 2023;121:714-30.
20. Li M, Dai Q, Su P, You Z, Ma Y. Surface layer modulus prediction of asphalt pavement based on LTPP database and machine learning for Mechanical-Empirical rehabilitation design applications. Constr Build Mater 2022;344:128303.
21. Li Z, Shi X, Cao J, Wang X, Huang W. CPSO-XGBoost segmented regression model for asphalt pavement deflection basin area prediction. Sci China Technol Sci 2022;65:1470-81.
22. Zhang J, Cao J, Huang W, Shi X, Zhou X. Rutting prediction and analysis of influence factors based on multivariate transfer entropy and graph neural networks. Neural Netw 2023;157:26-38.
23. Shen C, Jiang B, Yue L. LSTM combined with BIM technology in the management of small and medium-sized span highway concrete beam bridges. Results Eng 2023;20:101539.
24. Sharma A, Sachdeva SN, Aggarwal P. Predicting IRI using machine learning techniques. Int J Pavement Res Technol 2023;16:128-37.
25. Tommasi T, Orabona F, Caputo B. Learning categories from few examples with multi model knowledge transfer. IEEE Trans Pattern Anal Mach Intell 2013;36:928-41.
27. Finn C, Abbeel P, Levine S. Model-agnostic meta-learning for fast adaptation of deep networks. arXiv 2017. pp. 1126-35.
28. Vinyals O, Blundell C, Lillicrap T, Kavukcuoglu K, Wierstra D. Matching networks for one shot learning. arXiv 2016.
30. Santoro A, Bartunov S, Botvinick M, Wierstra D, Lillicrap T. Meta-learning with memory-augmented neural networks; 2016. Available from: https://proceedings.mlr.press/v48/santoro16.pdf [Last accessed on 28 Dec 2024].
31. Ravi S, Larochelle H. Optimization as a model for few-shot learning; 2016. Available from: https://openreview.net/forum?id=rJY0-Kcll [Last accessed on 28 Dec 2024].
32. Nichol A, Schulman J. Reptile: a scalable metalearning algorithm. arXiv 2018. Available from: https://yobibyte.github.io/files/paper_notes/Reptile___a_Scalable_Metalearning_Algorithm__Alex_Nichol_and_John_Schulman__2018.pdf [Last accessed on 28 Dec 2024].
33. Chu X, Cai F, Cui C, Hu M, Li L, Qin Q. Adaptive recommendation model using meta-learning for population-based algorithms. Inf Sci 2019;476:192-210.
34. Yu H, Zhang Q, Liu T, Lu J, Wen Y, Zhang G. Meta-ADD: a meta-learning based pre-trained model for concept drift active detection. Inf Sci 2022;608:996-1009.
35. Lee Y, Choi S. Gradient-based meta-learning with learned layerwise metric and subspace. arXiv 2018. pp. 2927-36.
36. Gordon J, Bronskill J, Bauer M, Nowozin S, Turner RE. Meta-learning probabilistic inference for prediction. arXiv 2018.
37. Bertinetto L, Henriques JF, Torr PHS, Vedaldi A. Meta-learning with differentiable closed-form solvers. arXiv 2018.
38. Lee K, Maji S, Ravichandran A, Soatto S. Meta-learning with differentiable convex optimization. In Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 15-20 June 2019; Long Beach, CA, USA. pp. 10657-65.
39. Aguiar GJ, Santana EJ, de Carvalho ACPFL, Barbon Junior S. Using meta-learning for multi-target regression. Inf Sci 2022;584:665-84.
40. Jeong D, Bae S. Estimating battery state-of-charge with a few target training data by meta-learning. J Power Sources 2023;553:232238.
43. Grant E, Finn C, Levine S, Darrell T, Griffiths T. Recasting gradient-based meta-learning as hierarchical bayes. arXiv 2018.
44. Rusu AA, Rao D, Sygnowski J, et al. Meta-learning with latent embedding optimization. arXiv 2018.
45. Morcos AS, Raghu M, Bengio S. Insights on representational similarity in neural networks with canonical correlation. arXiv 2018.
46. Kornblith S, Norouzi M, Lee H, Hinton G. Similarity of neural network representations revisited. arXiv 2019. pp. 3519-29.
47. Yang J, Wang X, Luo Z. Few-shot remaining useful life prediction based on meta-learning with deep sparse kernel network. Inf Sci 2024;653:119795.
49. Wong E, Rice L, Kolter JZ. Fast is better than free: Revisiting adversarial training. arXiv 2020.
50. Andriushchenko M, Flammarion N. Understanding and improving fast adversarial training. arXiv 2020.
51. Carmon Y, Raghunathan A, Schmidt L, Liang P, Duchi JC. Unlabeled data improves adversarial robustness. arXiv 2019.
52. Uesato J, Alayrac JB, Huang PS, Stanforth R, Fawzi A, Kohli P. Are labels required for improving adversarial robustness? arXiv 2019.
53. Chan A, Tay Y, Ong YS. What it thinks is important is important: Robustness transfers through input gradients. In Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 13-19 June 2020; Seattle, WA, USA. pp. 332-41.
54. Utrera F, Kravitz E, Erichson NB, Khanna R, Mahoney MW. Adversarially-trained deep nets transfer better: Illustration on image classification. arXiv 2020.
55. Salman H, Ilyas A, Engstrom L, Kapoor A, Madry A. Do adversarially robust imagenet models transfer better? arXiv 2020.
Cite This Article
Open AccessHow to Cite
Li, Z.; Jin, X.; Shi, X.; Cao, J. A meta-learning approach for predicting asphalt pavement deflection basin area. Complex Eng. Syst. 2024, 4, 26. http://dx.doi.org/10.20517/ces.2024.64
Download Citation
Export Citation File:
Type of Import
Tips on Downloading Citation
Citation Manager File Format
Type of Import
Direct Import: When the Direct Import option is selected (the default state), a dialogue box will give you the option to Save or Open the downloaded citation data. Choosing Open will either launch your citation manager or give you a choice of applications with which to use the metadata. The Save option saves the file locally for later use.
Indirect Import: When the Indirect Import option is selected, the metadata is displayed and may be copied and pasted as needed.
About This Article
Copyright
Data & Comments
Data
0
Comments
Comments must be written in English. Spam, offensive content, impersonation, and private information will not be permitted. If any comment is reported and identified as inappropriate content by OAE staff, the comment will be removed without notice. If you have any queries or need any help, please contact us at support@oaepublish.com.