The Effect of Knowledge Based Feature Extraction on Failure Detection of Control Surface Failures of Fighter Aircraft

Abstract

<p>While the area of maintenance is developing from scheduled maintenance toward the condition based maintenance also the failure detection that utilizes system operational data becomes more important. The failure detection from system data can be done in many manners but a process of feature extraction is present more or less almost when the system data is high dimensional, that is the case also with fighter aircraft systems. In this study we examine an effect of system knowledge based feature extraction on the further performance of an algorithmic tasks of failure detection on the operational flight data of a fighter aircraft. The failures for validating the results are several flight control surface failures from the flight data. The feature extraction has been done by using system knowledge of fighter aircraft experts. The failure detection algorithms are comprehensive set of algorithms from the field of anomaly detection, novelty detection, one class classification and unsupervised machine learning. This study demonstrates that some specific failure detection algorithms are more robust for feature extraction and can perform well even with low level of feature extraction when detecting the flight control surface failures. This result can be further used for selecting algorithms for failure detection tasks for other subsystems of aircraft in cases when the system knowledge and expertise are lacking and thus the feature extraction that can be done is little.</p>