Title: Automatic anomaly detection in jet engines
Authors: David Leslie - University of Lancaster (United Kingdom)
Nicos Pavlidis - Lancaster University (United Kingdom)
Harjit Hullait - Lancaster University (United Kingdom) [presenting]
Abstract: Jet Engines contain hundreds of sensors measuring various engine processes in continuous time. Building sophisticated statistical methods to process and make informative decisions from the sensor data offers huge opportunities. We will start by looking at sensor data from Pass-Off tests. Each test comprised an engine performing various pre-defined manoeuvres. There are two types of manoeuvres: piecewise-linear manoeuvres, containing sudden changes in behaviour and functional manoeuvres, which are comprised of smooth accelerations and decelerations. There are a number of challenges in using this data; firstly, the manoeuvres performed have not been labelled and secondly, manoeuvres can be partially performed. The first aim is to build an efficient classification method to identify the different manoeuvres. We have used the PELT changepoint algorithm to extract manoeuvre segments. We then use Needleman-Wunsch and Functional Principal Component Analysis to compare the manoeuvre segments to a number of model templates, which gives us a similarity score with respect to each template. We use the Needleman-Wunsch algorithm, as it is capable of identifying missing segments, so is particularly robust in scoring partially performed manoeuvres. These scores characterise the manoeuvres effectively, enabling the manoeuvres to be classified with a high level of accuracy.