Using Meta-Features to Boost the Performance of Classifier Fusion Schemes for Time Series Data

Accurate detection of faults and prediction of equipment remaining useful life result in considerable economic benefit to industry due to avoidance of unscheduled downtime and costly secondary damage. This paper described an approach to improving the performance of fault detection schemes that operate on time series data by fusing the results of an ensemble of classifiers that operate on the raw data. A meta-feature is designed to provide an indication of the historical state of other classifiers in the system, and to aid in fusion by allowing the individual classification results to be discounted appropriately. This meta-feature is shown to provide a substantial performance increase for fusion schemes. Data from actual aircraft engine/airframe systems are used to assess the performance of the approach.