A Framework for Generating Large Data Sets for Fatigue Damage Prognostic Problems

Prognostics and Health Management (PHM) relies on the availability of large amounts of data for a given system and allows to analyse this data and to draw conclusions as to the health state of the system, the identification of faults and failures, as well as the calculation of the remaining useful life time. Often, it is expected that this data is labelled, i.e. that the data has been pre-analysed and that for each data point an exact information is available as to what it is about, when it was measured, etc. In reality, this is not always easy and this labeled data is not always available. For example, on aerospace structures, complete labeled data until the end of their lifetime are not usually available. This may hamper for example the use of Deep Learning (DL) techniques for Predictive Maintenance, as they rely on the availability of large amounts of labeled sensor data. In this paper a framework and associated code
¹
is proposed to generate high dimensional data sets for a realistic fatigue damage prognostics problem, representative of fatigue cracks propagation in aeronautical fuselage panels. With this data, DL techniques can be trained, and we will illustrate this with a case study involving several of the most commonly used DL models to address failure prognostics.
¹
Seehttps://github.com/ansak95/FrameworkFDPP