| Author: | Lai, Li |
| Title: | AI-assisted decision-making system for infrastructure management incorporating inspection and monitoring |
| Advisors: | Dong, You (CEE) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Department: | Department of Civil and Environmental Engineering |
| Pages: | xvi, 215 pages : color illustrations |
| Language: | English |
| Abstract: | Due to deterioration over its service life, the performance of civil infrastructure could gradually fall below the threshold of its serviceability limit state. Failure of infrastructure may lead to catastrophic consequences both in terms of economic and human losses. Consequently, it is imperative to implement reasonable interventions to maintain the functions of infrastructure. Traditional infrastructure maintenance strategies typically rely on standards, involving periodic inspections of structures, preventive maintenance, and maintenance actions based on inspection reports. This passive pattern is conservative, resulting in higher maintenance costs over the infrastructural life cycle. To end this, the goal of the study is to build an Artificial Intelligence (AI) assistant decision-making system that enables efficient and economical managing infrastructure. To realize this AI-assisted management system, several pertinent modules and algorithms are integrated, including assessment of infrastructure performance, the evolution of infrastructure deteriorations, decision-making framework, and smart agent training for management policy. Unlike other studies, this work is dedicated to addressing practical infrastructure management issues using real-world data. It starts with the fatigue issue of welded details, progresses to the maintenance of hangers in concrete-filled steel tube arch bridges, and further expands to the maintenance of entire bridge components, as well as bridges maintenance within a transportation system. For the maintenance of a welded detail, the cumulative fatigue damage is estimated by the rain-flow method with dynamic strain data. As mentioned above, deriving a potential evolution model for fatigue is essential in life cycle management. For this, the Bayesian dynamic linear model (BDLM) is employed to predict stress cycling based on long-term monitoring strain data. To consider the inspected error and uncertainty in fatigue, the entire maintenance parameters are quantified into Partially Observable Markov Decision Processes (POMDPs). Due to the small scale of state and action, employing point-based algorithms is expected to yield convergence to the optimal maintenance policy. For arch bridge hangers' management, the data resource and management system become more complex. To accommodate the monitoring data change from dynamic strain to weigh-in-motion data, digital twins (DTs) are established to reconstruct the stress spectrum of hangers with vehicle loads. In addition, improved dynamic POMDPs are designed to adapt the hangers' different loading experiences. Lastly, to combat the large state space, the neural network is employed as an agent by creating a mapping between the state of hangers and the maintenance actions. Leveraging the predictive capabilities of DTs, the training set is produced through interaction within the DTs environment, and the neural network is updated by the Asynchronous Advantage Actor-Critic (A3C) algorithm. When management extends to all components of a bridge, three additional challenges need to be addressed. The first challenge involves the need for tailored deterioration models catering to the different types of components, such as the Weibull and Markov models. The POMDPs correspondingly develop to the Hybrid Markov decision processes (HMDPs) to better represent the bridge management system. The second challenge comes from the maintenance action space. The component-level management in a bridge result in the size of the action space increasing exponentially with the number of components. To avoid an oversized output layer in a neural network, a ranking model is supplemented to generate a prioritized list of maintenance tasks which simplifies the complexity of action space from 2n to n2. The third challenge is the difficulty of training smart agents. With the complexity in management and objective function, the policy often converges to suboptimal solutions. To address this issue, a novel training method incorporating imitation learning with Advantage Actor-Critic (A2C) is proposed. When the complexity of management is raised to the level of transportation networks, the states and maintenance behaviors of individual bridges can significantly affect the system. Any simplification of the action space may lead the agent to lose control of the system. To address these limitations, this study proposes a non-fully connected neural network that draws inspiration from the nervous system of an octopus. It incorporates synergetic branches and hierarchical rewards, factorizing the action space and diminishing system complexity from exponential to linear with the number of infrastructures. In addition, a novel training method which mimics biological learning behaviors has proved effective in finding optimal policy. The methodologies and frameworks developed in this work have been rigorously evaluated within real-world infrastructure cases, where their strengths, limitations, and expansive potential are thoroughly examined and compared with expert maintenance policies. It is impressive to find the superior performance of smart agents in the field of infrastructure management. Notably, as the complexity of maintenance challenges arises, the benefits of employing smart agents over traditional expert policies become increasingly evident. In summary, this work integrates smart agents in reinforcement learning with state-of-the-art technologies in structural health monitoring, bridge inspection, and digital twins, exploring its application in infrastructure management. It opens new scientific paths toward the realization of smart cities maintenance, the management of infrastructures, and minimizing the government budget for infrastructural services. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13685