ANALYSIS OF SPATIAL AND TEMPORAL FLUCTUATIONS IN SPRAY USING DEEP LEARNING METHOD FOCUSING ON CHARACTERISTICS OF INDIVIDUAL PARTICLES

This paper aims to investigate the efficacy and effectiveness of the spray, which are quantified based on the quantity and quality of spray particles produced by varying the atomization conditions. A deep learning-based particle detection method is employed to classify spray particles based on their geometric appearance. The outcome is used to perform analyses such as particle-producing capacity over time and estimating their orientation in a sequence of instantaneous frames of those cumulatively detected. This paper also focuses on characterizing the spray produced by a twin-fluid atomizer based on selective particles such as droplets and ligaments. The existing spray characterization methods are designed to analyze spherical droplets, which are ineffective for ligament analysis. However, a convolutional neural network (CNN)-enabled deep learning approach can detect desired spray particles with rectangular bounding boxes as regions of interest (ROI). A custom dataset was created to train the CNN model by labeling several droplets and ligaments in a frame using labeling rules made by domain experts, which are considered ground truth for verification. Based on the ground truth, we observed that spray particle detection by the CNN model is approximately 10% more accurate than the ImageJ-based conventional image processing approach. The width-to-height (W/H) ratio is analyzed using the dimensions of the bounding boxes of each detected particle and how the W/H ratio is affected by varying air-to-liquid ratios. During the temporal fluctuation process, particle deformations are examined at the peak points at the top and bottom. Numerous ligaments significantly deform horizontally when the particle count decreases and reaches its minimum. Conversely, most particles are circular when the particle count increases. These findings confirm the characterization of spray produced by twin fluid atomizers in terms of particle count and deformation resulting from the liquid breakup process.