Gas Flow Structures Detection on Shadowgraph Images and Their Matching to CFD Using Convolutional Neural Networks
Shadowgraph imaging has been widely used to study flow fields in experimental fluid dynamics. Nowadays high-speed cameras allow to obtain millions of frames per second. Thus, it is not possible to analyze and process such large data sets manually and automatic image processing software is required. In the present study a software for automatic flow structures detection and tracking was developed based on the convolutional neural network (the network architecture is based on the YOLOv2 algorithm). Auto ML techniques were used to automatically tune model and hyperparameters and speed-up model development and training process. The neural network was trained to detect shock waves, thermal plumes, and solid particles in the flow with high precision. We successfully tested out software on high-speed shadowgraph recordings of gas flow in shock tube with shock wave Mach number M = 2-4.5. Also, we performed CFD to simulate the same flow. In recent decades, the amount of data in numerical simulations has grown significantly due to the growth in performance of computers. Thus, machine learning is also required to process large arrays of CFD results. We developed another ML tool for experimental and simulated by CFD shadowgraph images matching. Our algorithm is based on the VGG16 deep neural network for feature vector extraction and k-nearest neighbors algorithm for finding the most similar images based on the cosine similarity. We successfully applied our algorithm to automatically find the corresponding experimental shadowgraph image for each CFD image of the flow in shock tube with a rectangular obstacle in the flow channel.