Abstract
Flash boiling spray improves the quality of fuel atomization while also increasing the occurrence of spray collapse. The spray collapse phenomenon significantly increases the likelihood of unintended spray-wall impingement, thus reducing combustion stability. Therefore, this study focuses on the spray collapse states under flash boiling conditions and implements machine learning methods to enhance spray analysis including collapse pattern recognition and prediction. In this study, the macroscopic spray cross-patterns of multiple-component fuel mixtures under different fuel temperatures and pressures were first captured by a high-speed planar laser Mie-scattering imaging system. Then, a supervised residual network (ResNet) classification model was applied for spray collapse pattern recognition. The well-trained ResNet model automatically classified the spray collapse patterns into three categories including non-collapse state, transition state, and collapse state with an accuracy of over 99%. This data-driven model provides a novel idea for the identification of spray collapse states. Based on these recognition results, a decision tree classifier (DTC) was built to evaluate the importance of test and fuel parameters that influence the spray collapse behavior. The results indicated that the superheated index modified by the ratio of ambient pressure to bubble point pressure was capable of determining the spray collapse states. With such trained DTC model and the modified superheated index, it is anticipated that the spray collapse states of multi-component fuels could be predicted with high accuracy even under untested conditions.
