Engine data analysis using decision trees

Monitoring systems on aircraft engines typically record many parameters, sampled frequently, over the duration of a flight. On some newer engines such data have been recorded, and stored, over hundreds of flights. But data alone are not sufficient, and are too unwieldy, to help the analyst understand the state of the engine's health. Automated data mining, for knowledge discovery, is being successfully used in several fields. One well-known technique is a decision tree, which is induced from a subset of a known training set of engine data and outcome. Since engine parameters are continuous-valued a splitting criterion, known as the minimum description length principle, is used to define the branch points of the tree. The decision tree is induced from a subset of a known training set of engine data and outcome. Issues such as overfitting the training data, and the problem of large, or randomly-chosen, training sets are also discussed in this paper. Copyright ©2000 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved. BACKGROUND Many airlines have invested in engine monitoring systems (EMS) on their newer aircraft. These systems record many parameters, such as temperatures, pressures, rotor speeds, fuel flow and others. The hope is that, by tracking the engines over hundreds of flights, one can discern component degradations which lead to diminished performance or safety. The large numbers of parameters, and the large volumes of time-series data, pose a challenge to the analyst sifting through the data. Simple checks such as minimum/maximum parameter bounds often lead to false alarms. Relaxing these bounds could lead to missed alerts. Several NASA Ames research efforts are directed at exploring ways of using computers to monitor the health of physical systems. In a previous paper [8] surveying aircraft engine health monitoring systems, various machine learning-based techniques, that are being researched for engine data analysis, were outlined. In this paper we focus on using the decision tree method for learning to classify engine's health based on its sensor data. (c)2000 American Institute of Aeronautics & Astronautics or Published with Permission of Author(s) and/or Author(s)' Sponsoring Organization.