Train on synthetic - Test on real: Domain adaptation for strain-based damage detection on an aircraft wing

Kurzfassung

There is an increasing development towards sustainable aviation. Thus, there are multiple approaches to decrease emissions. One of the possible solutions is mass reductions of structural elements. For these structural elements, mass reduction leads to an increasing need for precise maintenance. To fulfill this need, present research focuses on predictive or prescriptive maintenance. Nevertheless, predictive maintenance requires the as-is condition of a component for effective planning. Therefore, the train on synthetic - test on real approach is considered for structural health monitoring of the components. The approach is examined on the prediction of crack-like damages on an aircraft wing using strain-based data. Hence, a methodology is outlined that is used for the present paper. This methodology covers the development of a suitable test scenario. In the next step, a virtual and a physical representation of the scenario is built to generate a data set for each domain. Virtually, structural finite element model simulations are prepared and automated in Salome Meca using the Code_Aster solver. Physically, a test stand consisting of a cantilever with applied strain gauges is constructed. The objective is to train a machine learning model with the virtual data set and test it with the physical data set. For this, different supervised machine learning models from the Python library Scikit-Learn are compared. The classification of a damaged or undamaged structure works well among all models and all data sets. For the regression of the damage position success is recognized inside the virtual and the physical data set. Training the models in the virtual and testing them in the physical domain leads to problems for this initial investigation. A similar behavior appears for the multi-output regression. This requires a deeper understanding of both data set characteristics as well as continuing to exploit further machine learning opportunities like neural networks and dedicated domain adaptation methods.