Adaptive, State-Based Modeling Applied to Prognostics and Health Management of Industrial Rotating Equipment

Abstract

Unplanned downtime due to equipment health problems imposes significant costs on many manufacturing operations in the form of maintenance expenditures and lost production. To reduce these costs, companies have become interested in leveraging the latest smart manufacturing technologies to implement predictive maintenance strategies. The field of prognostics and health management research seeks to support this transition by developing modeling and analysis tools that provide manufacturers with real-time feedback on the health of their equipment throughout its life cycle. Based on this information, companies can minimize production interruptions by administering maintenance procedures only when necessary and at opportune times.

Predictive maintenance strategies rely on machine health models to derive actionable insights on the current and future health of machines from sensor measurements. The ability to quickly develop and deploy accurate health models is then critical for the advancement of predictive maintenance across the manufacturing industry. However, creating these models in manufacturing applications can be challenging due to numerous disturbances that commonly impact machines, lack of sufficient data to characterize machine behavior leading up to health events, and uncertainty surrounding machine degradation processes.

This dissertation proposes health modeling frameworks that address these challenges to facilitate more widespread adoption of predictive maintenance strategies. Novel, state-based modeling methods are developed to allow machine health models to maintain a memory of recent sensor measurements and prior modeling results when analyzing system health. This approach is well-suited for dynamic manufacturing environments and supports representations of multi-stage degradation processes. An extensible digital twin framework provides the tools to model the health of complex systems with hierarchies of standardized, reusable digital twins. An adaptive modeling framework is also proposed to probabilistically detect and diagnose ongoing degradation processes based on machine signal trends while simultaneously monitoring for unforeseen degradation modes. Multiple case studies with different types of rotating equipment demonstrate how the contributions of this dissertation allow manufacturers to develop standardized machine health models for a broad range of applications to fully realize the cost and production benefits of predictive maintenance.