Deep Learning Image Analysis for Angular Measurements in Wind Tunnels
José P. Ferreira, James H. Bell · AIAA Scitech 2020 Forum · 2020
Wind tunnel tests are part of a thorough process of development and certification for tomorrow’s aircraft and launch vehicles. The flight conditions are accurately simulated with a model in a controlled environment which includes the proper setting of its orientation in pitch, roll and yaw. Onboard sensors using accelerometers and conventional optical techniques such as photogrammetry are regularly used to determine the orientation of the model. An alternative approach is herewith presented, consisting of image analysis based on Convolutional Neural Networks to predict the continuous orientation of wind tunnel models. Several training data sets were constructed using a) synthetic images and photographs of a generic airplane rotating in pitch only; b) synthetic perspective views of CAD models and photographs of a model airplane at different pitch, yaw, and roll angles; and c) wind tunnel pictures of a half-span model in wind-on conditions. Results show that an accuracy of ± 0.5° is easily achieved for each Euler angle prediction, and an accuracy of ± 0.1° was achieved for pitch predictions in wind tunnel environment. It also proved effective to use synthetic views of CAD models as a training data set. This allows large training data sets to be rapidly generated without having to obtain photographs of the actual model in the wind tunnel.