Anomaly detection for aircraft electrical generator using machine learning in a functional data framework

To reduce the number of aircrafts on ground, the electrical design engineers are interested in predicting the oil temperature of the generator during a flight. Changes on the temperature value may indicate an incorrect functioning of the generator. An abnormal behavior can be identified by using machine learning algorithms trained on flights free from any anomalies to predict the generator oil temperature. These predictions can then be compared to the observed values, here the sensor data collected from the aircraft during flight. If the observed value is far from the predicted value, a failure warning is raised and a maintenance action shall be performed. In this paper, we build a digital twin of the electrical generator which predicts the oil generator temperature at a given time thanks to the history of features. We compare several machine learning algorithms and the most promising appears to be a neural network which was implemented as part of the electrical generator digital twin. The digital twin is tested by using real flight data containing generator failures and it is verified that the algorithm is able to detect an anomaly prior to the failure events (early failure detection).