A Data-Driven Rutting Depth Short-Time Prediction Model with Metaheuristic Optimization for Asphalt Pavements Based on RIOHTrack

Zhuoxuan Li, Iakov Korovin, Xinli Shi, Sergey Gorbachev, Nadezhda Gorbacheva, Wei Huang, Jinde Cao

TL;DR

This paper introduces a novel data-driven approach, IAPSO-RELM, for accurately predicting rutting depth in asphalt pavements. By combining complex network analysis with an optimized extreme learning machine, the model significantly improves prediction accuracy and reduces the need for extensive field testing, offering a more efficient way to assess pavement health.

What Problem Does This Paper Address?

Rutting, a critical pavement defect, significantly impacts road quality and safety, leading to substantial maintenance costs. Traditional methods for predicting rutting depth often struggle with limited, expensive field data and the inherent complexity of asphalt pavement behavior. This paper addresses the challenge of developing a robust and accurate prediction model that can effectively utilize available data, even when it's scarce for individual pavement types, and provide timely warnings of potential damage.

Methods

The core of this research lies in a multi-stage approach:

  • Complex Network Construction: The authors first construct a complex network where each asphalt pavement type is a node. The similarity in rutting depth variation between pavements is calculated using grey correlation, forming the edges of the network. This allows for a visual representation of how different pavement types behave similarly or differently.
  • Community Detection (Louvain Algorithm): The Louvain algorithm is applied to this complex network to identify "communities" of asphalt pavements that exhibit similar rutting behavior. This is a novel application of the Louvain algorithm in this domain.
  • Equivalent Training Set Creation: By grouping similar pavements into communities, the researchers create an "equivalent training set." This allows the model to learn from a broader range of similar behaviors, overcoming the limitation of having small datasets for individual pavement types.
  • RELM (Residual Correction Extreme Learning Machine): An enhanced Extreme Learning Machine (ELM) called RELM is proposed. RELM incorporates a second ELM to learn the residuals (errors) of the initial ELM, thereby improving stability and generalization.
  • IAPSO (Independent Adaptive Particle Swarm Optimization): The RELM model's parameters are optimized using an Independent Adaptive Particle Swarm Optimization (IAPSO) algorithm. This metaheuristic approach aims to find a better global optimum compared to standard PSO by adaptively adjusting parameters.

Datasets

The primary dataset used in this study comes from the RIOHTrack (Road Track Institute) full-scale accelerated pavement test track in Tongzhou, Beijing. This track features 25 types of asphalt pavement structures, and the study specifically utilizes data from 19 main test pavement structures (STR1-19). The data spans from 2017-2020 for training and 2021 for testing. No new benchmark datasets were introduced; rather, existing field data was leveraged.

Key Insights and Ideas

  • Community Detection for Data Augmentation: The most significant insight is the successful application of the Louvain algorithm for community detection to create an "equivalent training set." This elegantly addresses the problem of limited data for individual pavement types by grouping similar behaviors, leading to improved model performance.
  • RELM for Enhanced Prediction: The residual correction mechanism in RELM proves effective in improving the stability and generalization of ELM, making it more robust for complex pavement data.
  • IAPSO for Superior Optimization: The adaptive nature of IAPSO allows it to escape local optima more effectively than traditional PSO, leading to better parameter tuning for the RELM model and consequently higher prediction accuracy.
  • Data-Driven Pavement Engineering: The study reinforces the power of data-driven approaches in pavement engineering, demonstrating how AI can provide accurate predictions and reduce the reliance on costly and time-consuming physical tests.

Key Results

The proposed IAPSO-RELM model demonstrated superior performance compared to various other machine learning and deep learning algorithms. Specifically:

  • The model achieved an Average Root Mean Squared Error (RMSE) of 1.742, an Average Mean Absolute Error (MAE) of 1.363, and an Average Mean Absolute Percentage Error (MAPE) of 1.94% across 19 asphalt pavements.
  • The IAPSO-RELM algorithm showed significant improvements over other algorithms, with MAPE improvements ranging from 28.38% to 79.05% on average.
  • The community detection approach led to substantial improvements in prediction accuracy, with MAPE improvements of up to 84.97% compared to using single-type or all-kinds datasets.
  • The model's residuals were found to follow a normal distribution, indicating good reliability.

Limitations and Future Work

  • Limitations: The paper acknowledges that the weights in the standard ELM are randomly generated, leading to potential uncertainty in predictions. While RELM and IAPSO-RELM address this, the inherent randomness of initial ELM weights is a starting point for uncertainty. The study also notes that increasing the training sample size can further improve prediction performance.
  • Future Work: The authors propose extending their work to consider the combination of metaheuristic optimization and machine learning algorithms for oil and gas systems. They aim to predict key parameters in these systems by accounting for simulated environments with coupled environmental and internal system changes, providing a basis for intelligent scheduling.

Summary updated

BibTeX
@article{li2023data,
  author = {Zhuoxuan Li and Iakov Korovin and Xinli Shi and Sergey Gorbachev and Nadezhda Gorbacheva and Wei Huang and Jinde Cao},
  title = {{A Data-Driven Rutting Depth Short-Time Prediction Model with Metaheuristic Optimization for Asphalt Pavements Based on RIOHTrack}},
  journal = {IEEE CAA J. Autom. Sinica},
  year = {2023},
  doi = {10.1109/jas.2023.123192},
}
Figure