Data-driven Condition Evaluation of Transportation Systems
Open Access
- Author:
- Sengupta, Agnimitra
- Graduate Program:
- Civil Engineering (PHD)
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- June 02, 2023
- Committee Members:
- Patrick Fox, Program Head/Chair
S. Ilgin Guler, Chair & Dissertation Advisor
Parisa Shokouhi, Outside Field Member
Mehrdad Mahdavi, Outside Unit Member
Sudeepta Mondal, Special Member
Vikash Gayah, Major Field Member - Keywords:
- nondestructive evaluation
impact echo
machine learning
infrastructure systems
traffic prediction
uncertainty
traffic dynamics
deep learning - Abstract:
- Transportation systems involve complex interactions with traffic demand, loading, and environmental factors, which result in non-linearities in system performance. The structural and functional conditions of a system determine its efficiency in meeting mobility demands. However, budget constraints impose serious limitations on actively monitoring these system responses to maintain reliability. Due to the intrinsic complexity of system responses and limited data availability, it is necessary to develop robust machine-learning models that can accurately characterize system performance and predict future states, based on which actions can be undertaken to maximize their performance under optimal settings. This dissertation focuses on the development and application of machine-learning strategies in evaluating and predicting the conditions of transportation systems like infrastructures using non-destructive evaluation (NDE) techniques, and road networks using real-time traffic data. Multi-dimensional NDE data that capture damage-specific signatures are interpreted to quantify the degree of damage and structural integrity in terms of condition ratings. Several spectral-based autonomous signal classification mechanisms and probabilistic sequential models like hidden Markov models, which perform well with limited data availability, have also been explored. Additionally, this dissertation contributes to the functional performance estimation of networks in terms of macroscopic traffic variables by analyzing real-time traffic datasets. In particular, it focuses on solving problems like traffic prediction and uncertainty quantification using advanced deep learning models, which are essential for efficient traffic operations and optimal control. Data-driven modeling-specific issues like data scarcity, synthetic data generation and transferability, and generalizability of the models on out-of-distribution datasets have been discussed in the context of both NDE and traffic data.