Machine Learning for Modeling and Control of an Unsteady Wing Motion using Active Aerodynamic Bleed

The complex temporal dynamics of the aerodynamic effects of distributed air bleed actuation on a 2-D airfoil that is undergoing unsteady pitching and plunging motions in wind tunnel experiments is modeled using an architecture based on machine learning. The model was developed using measurements of the aerodynamic loads and the angular motion characteristics for a dynamically pitching airfoil along with its corresponding static configurations. Two machine learning architectures were investigated a fully connected feedforward neural network and a long- and short-term time-series network (LSTNet). The fully connected network consisted of 4 hidden layers and just over 18,000 tunable parameters. The LSTNet is a deep learning framework for predicting time series with multiple layers including a convolution neural network, two recurrent neural networks, a fully connected dense layer, and a linear autoregressive component. It is shown that machine learning architectures are capable of modeling the complex wing unsteady aerodynamic characteristic including its lift during unsteady motion in the presence of temporal bleed and that the effects of the bleed are adequately modeled.