Neural network prediction and control of three-dimensional unsteady separated flowfields

Using artificial neural networks (ANN), one approach to the control of unsteady aerodynamics is to develop real-time models which, given the actuator control signals, anticipate the unsteady flowfield wing interactions. These models of flow-wing interactions can then be used as the foundation upon which to develop adaptive control systems. This article supports this concept using three-dimensional unsteady surface pressure topologies collected from a rectangular wing pitched through the static stall angle at seven nondimensional pitch rates. A neural network model of the unsteady surface pressures was developed by training an ANN on five of these seven data sets. Following training, the only inputs required for the model were instantaneous angle of attack and angular velocity. These network-predicted unsteady surface pressure time histories were compared directly to the experimental pressure data. Then, a neural network controller for the wing motion history was developed using the pressure model. The results indicated that the controller actuator signals reliably yielded motion histories that generated the measured lift to drag ratio (LID) time histories. Further, the results suggest that for any desired LID requirement optimized motion histories can be generated.