Intermediate feature space approach for anomaly detection in aircraft engine data

Change detection is an important task for remote monitoring, fault diagnostics and system prognostics. When a fault occurs, it will often times cause changes in measurable quantities of the system. Early detection of changes in system measurements that indicate abnormal conditions helps the diagnostics of the fault so that appropriate maintenance action can be taken before the fault progresses, causing secondary damage to the system and system downtime. This paper presents two approaches for fusing the output of multiple change detection algorithms using random forests. What is novel and interesting about the work presented here is that the partitioning of the data into different change scenarios before training the classifier fusion approach results in a significant improvement over even a straightforward fusion approach.