AI-Driven Preventive Maintenance for Coastal Bridges in Marine Environments

The research objective is (1) to develop a proof-of-concept artificial intelligence (AI) model for optimal maintenance of coastal bridges in harsh marine environments, aimed at achieving the required performance and safety with the minimized overall maintenance cost; and (2) to pioneer a novel AI lifelong learning framework that incorporates historical records, domain knowledge, expert insights, and in-situ sensor data to transform maintenance practices across transportation systems. Problem Statement: Bridges along coastlines are inevitably subject to accelerated, corrosion-induced deterioration caused by sea salt and humidity. The maintenance of these bridges presents unique challenges but has received limited research focus. Preventive maintenance involves planned strategies of cost-effective treatments that retard deterioration and maintain or improve the function of these structures. The USDOT mandates a systematic approach to the preventive maintenance of bridges. Current maintenance often relies on engineers' judgment, which may not always maximize benefits or minimize costs. Optimized preventive maintenance can theoretically be addressed through reinforcement learning (RL). However, traditional data sources are currently insufficient for training AI models. Additionally, traditional RL struggles with handling uncertainty due to environmental variability. Recently, AI advancements, particularly in natural language processing and probabilistic deep RL, have provided new ways to address these challenges. This project will innovatively explore the use of historical textual data from bridge annual inspection reports, maintenance logs, construction documents, and national weather databases, as well as relevant research publications - data that have not previously been utilized with RL for preventive maintenance. It will also pioneer the use of probabilistic deep RL and a brand-new class of neural networks, GFlowNets, which have not yet been explored in preventive maintenance with limited and noisy datasets. Proposed Method: The method encompasses the following perspectives: (1) It involves collecting historical documents, including bridge annual inspection reports, maintenance records, construction documents, and relevant research publications related to deterioration in marine environments. This data will be utilized to fine-tune a large language model (LLM), such as BERT, to incorporate domain-specific knowledge, enabling the LLM to accurately represent text concerning deterioration and maintenance actions from relevant documents. The fine-tuned LLM can be used to generate text tokens for identifying and representing units of text clusters related to deterioration conditions, maintenance actions, costs, and the condition after maintenance from documents, exploring different text pooling and transforming methods to develop features that describe varying conditions, actions, costs, and consequences of actions across different sequences from historical documents in a similar way as the Named Entity Recognition. (2) The transformer-based AI, in a similar way to BERT, will be developed and trained to predict the masked conditions or consequences in a manner similar to how an LLM predicts the masked words in sentences, using sequences of text clusters in historical documents. This training allows the AI model to understand the underlying rules of deterioration and maintenance processes, akin to the language of natural laws, enabling it to generate simulated data for various conditions and maintenance scenarios. For instance, given a specific condition, if a certain maintenance action is taken, the model can probabilistically determine the cost of the action and the resulting condition of the bridge based on rules implied in the historical documents. (3) This simulated data will be employed to further train RL agents through probabilistic reinforcement learning (RL) or GFlowNets. The developed model will offer not just a single policy but a range of policy options whose probability distribution aligns with the cost function, allowing decision-makers to evaluate various actions and understand the associated costs and the renewed conditions. (4) The data will be continuously collected and utilized to refine the AI model further, incorporating Reinforcement learning from human feedback to include expert input within the loop, enabling it to operate in a lifelong learning mode as more data is gathered, thus enhancing the reliability of decision-making. This innovative approach will empower AI agents to make optimized decisions based on domain knowledge, historical data, current sensor data, and expert insights while accounting for uncertainty. (References will be provided upon request.)

    Language

    • English

    Project

    • Status: Active
    • Funding: $82500
    • Contract Numbers:

      69A3552348331

    • Sponsor Organizations:

      Office of the Assistant Secretary for Research and Technology

      University Transportation Centers Program
      Department of Transportation
      Washington, DC  United States  20590
    • Managing Organizations:

      Maritime Transportation Research and Education Center (MarTREC)

      University of Arkansas
      4190 Bell Engineering Center
      Fayetteville, AR  United States  72701
    • Performing Organizations:

      Jackson State University, Jackson

      Department of Civil and Environmental Engineering
      Jackson, MS  United States  39217-0168
    • Principal Investigators:

      Zheng, Wei

    • Start Date: 20250130
    • Expected Completion Date: 20260130
    • Actual Completion Date: 0
    • USDOT Program: University Transportation Centers

    Subject/Index Terms

    Filing Info

    • Accession Number: 01924917
    • Record Type: Research project
    • Source Agency: Maritime Transportation Research and Education Center (MarTREC)
    • Contract Numbers: 69A3552348331
    • Files: UTC, RIP
    • Created Date: Jul 23 2024 4:22PM