Clustering investigation of scramjet combustion processes based on contrastive learning

Abstract

The transition between combustion processes or modes of scramjet engines is high-speed and dynamic, which may lead to unstable or even runaway. It is crucial to identify the combustion state of scramjet engine for controlling and maintaining stable operation. Based on the methods of contrastive learning, a clustering network model of combustion processes is developed to deal with the schlieren image of a scramjet combustor without labels. Firstly, manual classification is performed to define and classify the combustion processes of scramjet combustor under different experimental conditions. The datasets of scramjet combustion processes are constructed by using schlieren images. Meanwhile, the unsupervised clustering of combustion processes is realized to learn its change rule upon the contrastive learning framework of examples. Subsequently, the results of manual classification and model clustering are compared and analyzed to evaluate the identification performance. The results suggest that the clustering results of the model are basically the same as those of manual classification. All evaluation indicators show a good performance using the clustering model, and the Adjusted Rand Index and Accuracy are over 0.94 and 0.89, respectively. This work indicates that the application of an unsupervised contrastive learning method is impressively efficient in the combustion state identification of scramjet.