Convolutional neural networks for tracking coalescing waveforms in supersonic jet noise

Abstract

High-speed schlieren images of the field surrounding a supersonic jet provide a rich database for testing convolutional neural network (CNN) classification methods due to the presence of many waveform structures of interest. One such structure is waveform coalescence, where waves intersect at small angles, which can lead to increased steepening and nonlinear distortion in the jet near field [Willis et al., AIAA Journal (2023)]. Although waves exhibiting coalescence behavior have been identified in narrow field-of-view (FOV) schlieren images, new methods are needed for large FOV images to improve computational efficiency. This presentation will explore methods to track and classify waves observed in large FOV schlieren images as coalescing or noncoalescing using CNNs. Using transfer learning, pretrained networks can be retrained for this problem, with training data obtained from narrow FOV schlieren or from two-dimensional simulated waveforms using the Khokhlov–Zabolotskaya–Kuznetzov (KZK) equation. The impact of model hyperparameters, choice of pretrained network, image scaling, and other factors on the results of waveform classification by the CNN will be explored. [WAW is supported by the ARL:UT Chester M. McKinney Graduate Fellowship in Acoustics.]