CIAM-UTC-REG41
Research Team
PI: Kostas Papakonstantinou, Penn State
Co-PI: Ilgin Guler, Penn State
Co-PI: Vikash Gayah, Penn State
Funding Sources
Penn State Share: $125,000
PennDOT Match $126,090.28
Total Project Cost: $251,090.28
Agency ID or Contract Number
69A3551847103
Start and End Dates
01/03/2022 — 02/03/2024
Project Description
Accurate prediction of infrastructure component condition and performance is essential to support optimal planning of life-cycle maintenance and inspection actions. Although relevant available data are often abundant and describe several features and characteristics, they are not always accurate and/or complete. Data can often be completely missing from databases, for example due to a regular inspection not being properly recorded or, more importantly, maintenance activities not being documented, and various monitoring devices’ sensitivities and ratings given by different inspectors can lead to significant data variability and uncertainty. To this end, three important goals of this project are (i) to offer solutions for data-based detection of missing maintenance and repair database entries, (ii) quantification of the impacts of uncertain and incomplete data on deterioration modeling, and (iii) assessment of all of the above in the context of life-cycle decision-making. Advanced statistical analysis and machine learning methodologies will be utilized to identify trends in the data and detect potential outliers, such as condition ratings that last statistically longer than expected or statistically prolonged intervals between inspections. Frameworks already developed by the PIs can be leveraged and extended to address these concerns. As examples: Markov models (with or without latent states) can support structural deterioration predictions based on noisy and incomplete data, under further development; survival analysis models can parametrize and map noisy condition data to important metrics of interest (e.g. remaining service life); and Partially Observable Markov Decision Processes (POMDPs) powered by AI concepts of Deep Reinforcement Learning (DRL) can be extended to integrate the above with decision-making optimization powered by AI concepts of Deep Reinforcement Learning (DRL) can be extended to integrate the above with decision-making optimization.