Applying Outbreak Detection Algorithms to Prognostics

Abstract

Fleet maintenance and supply management systems are challenged to increase the availability and reliability of equipment. Prognostics can help. This paper examines the utility of selected statistical data mining algorithms, originally developed for bio-surveillance applications, in achieving fleet prognostics. Preliminary experimental evaluation suggests that it is possible, useful and practical to apply such algorithms to rapidly detect emerging patterns of systematic failures of equipment or support processes, and to continuously monitor relevant data for indications of specific types of failures. The key remaining technical challenge is to tame down a potentially large number of plausible pattern detections without compromising high detectability rates. The key practical consequences to maintenance and supply managers include the ability to be notified about emergence of a possible problem substantially earlier than before, the ability to routinely screen incoming data for indications of problems of all conceivable types even if their number is very large, and the ability to pragmatically prioritize investigative efforts according to the statistical significance of the detections.