Cartographing dynamic stall with machine learning

Once stall has set in, lift collapses, drag increases and then both of these forces will fluctuate strongly. The result is higher fatigue loads and lower energy yield. In dynamic stall, separation first develops from the trailing edge up the leading edge. Eventually the shear layer rolls up, and then a coherent vortex forms and then sheds downstream with its low-pressure core causing a lift overshoot and moment drop. When 50+ experimental cycles of lift or pressure values are averaged, this process appears clear and coherent in flow visualizations. Unfortunately, stall is not one clean process but a broad collection of processes. This means that the analysis of separated flows should be able to detect outliers and analyze cycle-to-cycle variations. Modern data science and machine learning can be used to treat separated flows. In this study, a clustering method based on dynamic time warping is used to find different shedding behaviors. This method captures the fact that secondary and tertiary vorticity vary strongly, and in static stall with surging flow the flow can occasionally reattach. A convolutional neural network was used to extract dynamic stall vorticity convection speeds and phases from pressure data. Finally, bootstrapping was used to provide best practices regarding the number of experimental repetitions required to ensure experimental convergence.