Segmentation and tracking: A deep learning-based method for analyzing dynamic features of aluminum agglomerates

Abstract

The phenomenon of aluminum agglomeration, which occurs during the combustion of aluminized solid propellants, can have a series of adverse effects on solid rocket motors. High-speed microimaging has been demonstrated to be an effective technique for investigating this phenomenon. Consequently, an efficient and accurate analysis of aluminum agglomerates in high-speed video is a crucial aspect of this research. However, traditional analysis methods often produce unsatisfactory segmentation results in complex images and show low efficiency in extracting the features of the same agglomerate across multiple frames. To address these issues, this study proposes an online analysis method based on deep learning, which allows for the real-time segmentation and cross-frame tracking of aluminum agglomerates. Comparative experiments with the classical threshold-based method demonstrate the effectiveness and superior accuracy of the proposed method, with the AP50 metric improving significantly from 0.546 to 0.940, achieving impressive segmentation performance. Subsequently, the method was employed to analyze the overall velocity characteristics, features of individual agglomerates, and the second mergence phenomenon of aluminum agglomerates. The analysis captured the complete dynamic process of agglomerates from formation to exiting the frame and yield a linear correlation between the projected area and maximum vertical velocity. The proposed method markedly simplifies the analysis process for aluminum agglomerates, furnishes more detailed dynamic information, and provides robust support for further studies on aluminum particle combustion and agglomeration mechanisms.