Using Synthetic Data to Train an Accurate Real-World Fault Detection System

Avoidance of unscheduled downtime and costly secondary damage make the accurate prediction of equipment remaining useful life of enormous economic benefit to industry. The detection of faults is an important first step in building a prognostic reasoner. This paper describes an approach for improving the performance of fault detection systems that operate on time series data. The method generates synthetic data that closely matches the characteristics of the raw data. These synthetic data are used to develop and evaluate classification systems in the common situation where real labeled data is quite scarce. The real data can then be preserved for use as a test set to assess the performance of the classification system. Data collected from aircraft engine/airframe systems is used to assess the performance of the resulting classification system