Deep learning based techniques for flame identification in optical engines

Abstract

The mandatory migration from fossil to renewable energy sources requires the characterization of new alternative fuels. One important step in fuel characterization is the test in optical engines, which allows the morphological characterization of flames. This analysis requires the post treatment of images by using segmentation. In many cases, an automatic threshold presents shortcomings as the flames may present different regions with variable luminosity, as also reflections from valves and cylinder liner. Consequently, a time-consuming manual image processing is required and, therefore, an automatic procedure would be welcome. The use of deep learning techniques for image segmentation is a promising alternative for such task, which has showed excellent results in several applications. In this study, two different models were trained to identify flames in images obtained from an optical engine operating at various conditions. The dataset used to train the models was generated by using images from tests with several types of fuels and combustion modes. The effects of image resolution and the generalization capabilities for different fuels and combustion operation were investigated. After analyzing the results, the use of deep learning methods to identify and characterize flames was validated as a mean for improving processing time.