Deep learning-based pavement performance modeling using multiple distress indicators and road work history

Technical Summary: Deep Learning-Based Pavement Performance Modeling Using Multiple Distress Indicators and Road Work History

Problem Statement
Pavement deterioration is a complex process driven by design, environmental, material, and unobserved variables. Accurate prediction of pavement conditions is critical for transportation agencies to optimize maintenance and rehabilitation budgets. Traditional performance models, such as Markov Chain approaches and various statistical regression methods, often rely on transition probabilities or require extensive manual feature extraction and selection. While machine learning models like Artificial Neural Networks (ANN) have been applied to pavement performance, they still often necessitate engineered features. This paper addresses the challenge of modeling pavement deterioration using raw, multi-dimensional time-series data without conventional feature engineering, leveraging the capabilities of deep learning to capture complex deterioration rates and the impact of maintenance history.

Methodology
The study employs three deep learning architectures to model pavement performance using data from the Texas Department of Transportation (TxDOT) Pavement Management Information System (PMIS) spanning 18 years (2000–2018).

• Data: The dataset comprises over 100,000 pavement sections (approximately 0.5 miles each) and includes 21 flexible pavement condition indicators. These indicators cover rutting (visual and depth), cracking (block, alligator, longitudinal, transverse), patching, failures, raveling, flushing, and roughness (IRI, Ride Score, Distress Score). Additionally, the model incorporates road work history, including 20 types of maintenance and rehabilitation treatments (e.g., seal coats, overlays, rehabilitation) and the time elapsed since the last treatment.
• Input Representation: Each data point is reshaped into a 2D matrix representing a "time-image," where rows correspond to years and columns represent the 21 condition indicators plus maintenance dummy variables and time-since-treatment variables.
• Models:
  1. Convolutional Neural Network (CNN): Utilized for its ability to capture local predictors and encode structural information through shared parameters and sparse connections. The architecture includes three convolutional layers (32, 64, 128 filters) followed by max-pooling and fully connected layers.
  2. Long Short-Term Memory (LSTM): A recurrent neural network variant designed to handle time-series data and preserve structural information over time steps. It utilizes a memory cell and gating mechanisms (forget, input, and output gates) to manage long-term dependencies.
  3. Hybrid CNN-LSTM: A framework where CNN layers perform feature extraction on the input data at each time step, and the resulting features are processed by an LSTM layer to model the sequence.
• Evaluation: The models were trained on 80% of the data and tested on 20%. Performance was evaluated using $R^2$ scores for regression tasks (continuous variables) and Accuracy for classification tasks (discrete variables like raveling and flushing codes).

Key Results
The study developed and tested models for all 21 condition indicators. The results yielded the following observations:

• Overall Performance: The CNN models generally outperformed the LSTM and CNN-LSTM models for more than half of the condition indicators.
• High-Performance Indicators: Five indicators achieved $R^2$ scores greater than 0.70: Rut Failure, Ride Score, Left IRI, Right IRI, and Average IRI. The authors attribute this success to the consistency and reliability of roughness data collection compared to other distress types.
• Low-Performance Indicators: Several indicators, including severe rutting, patching, failure, block cracking, and transverse cracking, resulted in large negative $R^2$ scores. The paper notes this is primarily due to the target variables having very little variation in the dataset, causing the sum-of-squares denominator in the $R^2$ calculation to be near zero, which renders the metric sensitive to small prediction errors.
• Visual Validation: Plots of actual versus predicted values for a randomly selected section showed that the models produced reasonable predictions for rut failure, rut depth, longitudinal cracking, raveling, flushing, and ride quality indicators. However, for indicators where historical data remained constant (e.g., deep rutting, severe rutting), predictions diverged significantly from true values.
• Training Dynamics: Analysis of training history indicated that for high-performing indicators (e.g., Ride Score, IRI), the $R^2$ scores were still rising at the 100th epoch, suggesting the models had not yet overfit and could potentially achieve higher accuracy with further training.

Significance and Claims
The paper claims that the proposed deep learning framework demonstrates effectiveness and robustness in predicting pavement performance, particularly for roughness-related indicators at the network level. By utilizing raw data without manual feature extraction, the study highlights the potential of deep learning to automate the feature engineering process inherent in traditional modeling.

The authors state that the developed models can assist engineers and administrators in:

1. More effectively managing infrastructure systems through improved performance prediction.
2. Optimizing budget allocation and needs analysis at the network level.
3. Extending the methodology to characterize the performance of other infrastructure facilities beyond pavement.

The study concludes that while the models show promise for specific indicators (roughness), the variability in other distress data types presents challenges that result in lower predictive accuracy, indicating that the applicability of the model varies by the specific type of pavement distress being modeled.