Indexation of Bench Test and Flight Data

Abstract

An important amount of data is provided every day in Safran Aircraft Engines’ test benches. Produced by thousands of sensors, test and flight data represent a big interest for engineers but manual analysis of the information is too complex, if not impossible. As specific data are extracted from the database, it is not unusual to miss interesting information when focusing on a given problem. Defining the data as a succession of labels where each label appears as transient or stabilized phases would be one way to solve the problem. The start and stop points of the different phases will be computed by an offline change-point detection algorithm. In order to detect potential crucial changes of characteristic variables, it is relevant to develop powerful algorithms. The Pruned Exact Linear Time (PELT) method (Killick R. & Eckley, 2012) is a parametric change-point detection which searches the optimal partition of a monovariate signal (temporal series of one
variable) in our case (but can also be applied to a multivariate signal) with a linear complexity. This algorithm meets our expectations in many ways: robustness, fast computing and accurate results. Then on a multivariate aspect, patterns are built with parameters and initial conditions and, classified in a specific category with a map/reduce scheme. This clustering will allow different analysis: the comparison of different patterns with the definition of a distance and the research of a specific pattern in a large database. For example if an engine shows a specific engine temperature pattern after the test pilot changes the shaft rotation speed from one level to another, engineers may ask if this behavior is usual. If not, it should be very interesting to see if such pattern happens in the past on other engines or other tests and dig from the database the old documents related to those rare events and eventually the people concerned. The objective of this project is to progressively score and classify different patterns in an increasing database of labels. The first step was to implement the PELT algorithm. Then it is possible to identify the different transient phases extracted from small subsets of temporal measurements and compute models for each patterns. These codification of transient phase will lead to a classification into labels or topics. After defining enough patterns, each new record of measurements will be automatically classified.