Robust Pavement Modulus Prediction Using Time-Structured Deep Models and Perturbation-Based Evaluation on FWD Data.

Area of Science:

• Civil Engineering
• Geotechnical Engineering
• Data Science

Background:

• Accurate pavement structural modulus prediction is vital for infrastructure maintenance and lifecycle assessment.
• Deep learning models show promise but struggle with time-series data structuring and measurement noise.

Purpose of the Study:

• To develop an integrated framework for pavement modulus prediction using deep learning.
• To investigate the impact of input sequencing strategies on model performance and robustness.
• To evaluate the model's reliability under simulated sensor uncertainty.

Main Methods:

• Developed five input sequencing strategies (Plan A-E) for time-series data.
• Utilized a hybrid Wide & Deep ResRNN architecture (SimpleRNN, GRU, LSTM) for multi-layer modulus prediction.
• Injected Gaussian noise (±3% variance) to assess robustness and estimated confidence intervals via Monte Carlo simulation.

Main Results:

• Input time-step design critically influences prediction accuracy and robustness.
• Plan D sequencing strategy demonstrated the optimal balance between accuracy and stability.
• The proposed framework effectively handles noisy data and provides confidence intervals.

Conclusions:

• Systematic time-step modeling and perturbation-based evaluation enhance deep learning for pavement engineering.
• The framework offers a practical and generalizable solution for pavement modulus prediction in uncertain field conditions.