Abstract
A machine learning-based approach is presented, which allows to detect persistent engine faults after a single flight. It utilizes transient in-flight measurements and a transient engine model. The time series of the residuals between the measured data and the data resulting from performance synthesis is evaluated using moving windows containing at least one transient segment. A continuous wavelet transformation and a pretrained convolutional neural network are utilized on the residuals for feature extraction. The fault detection is carried out via a one-class support vector machine, trained exclusively on nominal engine operation data. Therefore, the approach requires no a-priory knowledge of the effects of engine faults on the in-flight measurements. Under the assumption of persistent faults, all windows of a single flight, which contain at least one transient segment, are considered in order to improve the reliability of the fault detection. This approach is validated using measured data of a small helicopter engine that replicates the dynamic flight of the corresponding model helicopter on a ground test bed. Consequently, step changes as well as complex variations of the shaft power output are considered. Four standard gas path sensors are considered. The one-class support vector machine is used successfully to detect two types of total pressure sensor anomalies. Assuming a typical number of transient segments for an average short haul flight, it turns out that persistent faults can be detected within one flight with a probability of above .