Extension of the Learning Domain in Monitoring Turbofan Start Capability System
Abstract
The presented system monitors a turbofan start sequence using indicators and operating conditions to detect abnormal behavior. It is based on the analysis of the residuals between the measured indicator values and the corresponding estimated values assuming healthy state. Estimation uses regression models trained on a database. However, as in many monitoring problems, the amount of data is limited due to application issues and covers only a limited region of the feature space. Thus, the models are trained in a limited domain defined implicitly by the available learning data and their efficiency is not controlled outside this implicit domain. This paper deals with the definition and the extension of the models validity region while keeping the extension effect on the monitoring process under control. A methodology based on one-class SVM is proposed and is applied to the presented monitoring system. Practical and methodological conclusions are drawn from the proposed experiments.