Study of innovative sensor configurations for evaluating of the loads acting on the aircraft structure

In sizing a flight control system, one of the main required parameter is the hinge moment. The hinge moment of a control surface is given by the forces acting on the surface itself multiplied by the mechanical transmissions arm. This moment is assumed to be evaluated on the hinge axis, which is generally the only rotating component of an aircraft wing. Due to its high rate of variation depending on flight attitude and conditions, hinge moment boundaries are often numerically evaluated during design phase. The present study aims to introduce a methodology and draft guidelines to collect and analyze a great number of data entry, in order to fill a dataset to be used for building a simple Deep Learning Network (DLN) model. The model will actually represent a ”virtual sensor” able to estimate wing hinge moments during some flight scenarios. Even if it has been chosen a real case study in terms of geometry and structure, to limit the complexity not all data have been measured by own from beginning. In fact, every approach for the calculation of a wing hinge moment needs the engineer to be informed at least about the aerodynamic characteristics of the wing, not to mention flight conditions. Moreover, through the analysis of these characteristics it is also possible to emulate a particular sensor detection, then to consider it as a useful input for the aforementioned network. Aerodynamic analyses have been performed through Xflr5, an analysis tool for airfoils, wings and planes operating at low Reynolds Numbers. MatLab has been used to capture and collect data about hinge moment value and parameters (aileron deflection, flight speed, angle of attack) which produced it. Several simulations have been run to prepare a significant dataset and reach a dimension for DLN model to be well trained. All about data exploratory analysis and DNL modelling (building, training, results evaluation) has been performed through “Jupyter”, a non-profit, open-source project evolved to support interactive data science and scientific computing. After having trained the model on how to forecast hinge moment values starting from some inputs, obtained predictions have been evaluated to better understand performances and capabilities of the model itself. The DLN has been optimized to maximize forecasting “goodness” without complicate too much the adopted structure. Indeed, only if prediction error is under a certain threshold, the model may be considered “ready to deploy”. The design has been driven by a simple but fully modular approach, in which each component is open for future developments, providing at the same time a guide line about the expected output to move towards the DLN. It is also important to highlight how the choice of some parameters can affect hinge moment magnitude and a simple model makes this possible and easier.