For an improved control of the production process of aircraft engines, a new databasis has been built containing a huge number of geometrical characteristics of the engine’s parts as well as performance measurements obtained during the final acceptance test of the engine. The goal is to use these data to build a new prediction model of the performance that has high accuracy and detects drifts at an early stage of the production process. More precisely, we are interested in the prediction of the engine’s total airflow and data are mainly about characteristics of the fan blades and conditions of the test bench.

In this work we present the database and illustrate its specific characteristics like strong correlations among some of the features, missing values and general temporal drifts. In addition, we explore classical machine learning models to determine the relationship between fan blade production and the engine’s total airflow. In a second step we apply different time series models to better model the dependencies, such as ARIMA and LSTM, for the prediction of the airflow. It turns out that it is particularly difficult to disentangle the impact of the geometric measurements and the production drifts on the performance. This calls for methods that go beyond conventional machine learning models and that are specifically designed for our data.

This content is only available via PDF.
You do not currently have access to this content.