Prognosis: Challenges, Precepts, Myths and Applications

• Published in: 2022 IEEE Aerospace Conference (AERO) pp. 1 - 13
• Main Authors: Vachtsevanos, George, Zahiri, Feraidoon
• Format: Conference Proceeding
• Language: English
• Published: IEEE 05.03.2022
• Subjects: Data mining; Data models; Prediction algorithms; Prognostics and health management; Temperature; Training; Uncertainty
• DOI: 10.1109/AERO53065.2022.9843823

This paper addresses the design, testing and evaluation of rigorous and verifiable diagnostic and prognostic algorithms integrated into a holistic framework for integrity management of aircraft components/systems that are subjected to incipient failure modes. Prognosis has captivated the world from ancient times. Pythia's pronouncements at Delphi were a deciding factor in war and peace. The uncertainty associated with her pronouncements was legendary and set the stage for current uncertainty representations. It remains the major challenge in many application domains. Three pillars of technological advances are affected by prognosis: Autonomy - prognosis of the remaining useful life of failing components/systems is an essential attribute for autonomous operation of critical assets. Prognostics and Health Management - the major constituent technology for component/system health management enabling aircraft, industrial and manufacturing processes to predict the Remaining Useful Life (RUL) of failing components allowing the application of mitigating strategies when needed. Life Cycle Management - Long-term prediction of the RUL allows for optimum component/system maintenance management strategies when no fault is detected, and the system is under stress/usage patterns. This paper focuses on the utility of a systems integrating methodology for Condition Based Maintenance (CBM) of complex systems that are subjected to unanticipated disturbances. Prognosis is the Achilles' heel of the CBM architecture due to the inherent uncertainty in prediction models. The designer of a comprehensive and verifiable prognostics architecture is faced with significant challenges: data availability (baseline and fault data that is correlated and time stamped, sampled at appropriate rates) is always a concern; uncertainty representation, propagation and management, inherent in prognosis; high fidelity modeling of critical components/systems is lacking; data mining tools for feature extraction and selection are hand-engineered; finally, diagnostic and prognostic algorithms must address accurately the designer's specifications. There is a need to differentiate between prognosis and trending, i.e., the practice of regressing linearly a system variable (temperature, for example) until it reaches a specified threshold. Another differentiating characteristic relates to health-based vs usage-based prognostics with the former predicting the RUL of failing components on-line in real time as data is streaming in after a fault or incipient failure has been detected and identified. The latter refers to long-term prediction of the RUL exploited for reliability and life cycle management studies. We address both prognostic notions in this paper and view each one within the overall context and scope of the health-based vs usage-based framework. The theoretical underpinnings borrow from the emerging fields of Prognostics and Health Management (PHM), Condition Based Maintenance (CBM+) and novel data mining, reasoning, and work package optimization to expedite maintenance actions at the depot level.