Optimization And Machine Learning Approaches For Constrained Intelligent Maintenance Planning

Abstract

Modern life relies on complex, highly interconnected production, distribution, and logistics systems for energy, goods, and services. These systems are designed to operate with minimal interruption, ensuring exceptionally high reliability and rapid responsiveness to disruptions caused by unforeseen events such as extreme weather, natural disasters, and pandemics. Their failure can be extremely costly, making it increasingly crucial for organizations to efficiently manage equipment, enhance asset availability, and reduce operational costs. Many of these systems follow an alternating cycle of missions and breaks. Selective maintenance (SM) is widely used in these contexts and aims to optimize asset performance during the next mission by allocating limited maintenance resources.

This dissertation explores four key themes, each addressing critical knowledge gaps in the modeling and application of SM. The first theme explores the application of SM to fleets of geographically distributed systems and networked systems with geographically dispersed components. The models developed in this theme have far-ranging applicability and require the development of tailored decomposition algorithms. The second theme develops a data-driven predictive SM framework where deep neural networks are utilized for remaining useful life (RUL) prediction. To quantify uncertainty in predictions, a Bayesian approach is applied to generate a distribution of RUL opposed to a single point prediction. The third theme establishes a theoretical link between the SMP and the resource-constrained project scheduling problem (RCPSP). This connection leads to the development of a more practical maintenance model that is able to handle maintenance planning under resource constraints and multiple maintenance capacities. Finally, the fourth theme explores the application of SM to mission-critical systems implementing mission abort policies. An optimization model is developed that seeks to strike a balance between mission success and system survival probability. The developed models are applied to several illustrative examples and case studies, demonstrating their applicability and added value.