Artificial Intelligence for Prognostics: Papers from the AAAI Fall Symposium
George Vachtsevanos, Serdar Uckun, and Kai Goebel, Cochairs
Over the last ten years, there has been substantial interest and investment in prognostics in aerospace, transportation, and other industries. The field of prognostics focuses on methods and tools to determine functional degradation of components and systems and to estimate remaining useful life. The ultimate goal of prognostics is to manage the remaining useful life of systems such that maintenance actions can be performed “just in time” prior to failure, thus increasing safety as well as reducing maintenance expenses due to unscheduled downtime or unnecessary “preventive” maintenance. Given accurate remaining life estimates, prognostics also aims to manage the accumulation of further damage through control actions, for example, by either redistributing the load onto other components or changing the mission profile by trading off secondary mission goals.
These days, the hype around prognostics rivals the early days of artificial intelligence. Nevertheless, in practice, accurate prognostics has proven rather difficult to accomplish. There are numerous issues that still need to be resolved before prognostics is adopted as standard practice in the industry. These issues include the following:
- How does one successfully manage uncertainty of the prediction without incurring inadmissibly large confidence bounds that would wipe out the benefits?
- How does one validate prognostic techniques for new and expensive systems for which no historical data exist?
- How does one treat post-prognostic decision making that involves a tradeoff between numerous criteria including cost, risk, and logistics while some of these criteria change dynamically?
- How does one provide comprehensive coverage for a large number of fault modes that make a detailed materials-based approach infeasible for all fault modes?
- How does one communicate the information between different components and arrive at a complete system-wide health status that properly accounts for the interactions between different components?
- How does one perform online (or offline) reconfiguration that avert impending system failure within the framework of complex systems?