Prediction of Combustion Phasing Using Deep Convolutional Neural Networks

A Machine Learning (ML) approach is presented to correlate in-cylinder images of early flame kernel development within a spark-ignited (SI) gasoline engine to early-, mid-, and late-stage flame propagation. The objective of this study was to train machine learning models to analyze the relevance of flame surface features on subsequent burn rates. Ultimately, an approach of this nature can be generalized to flame images from a variety of sources. The prediction of combustion phasing was formulated as a regression problem to train predictive models to supplement observations of early flame kernel growth. High-speed images were captured from an optically accessible SI engine for 357 cycles under pre-mixed operation. A subset of these images was used to train three models: a linear regression model, a deep Convolutional Neural Network (CNN) based on the InceptionV3 architecture and a CNN built with assisted learning on the VGG19 architecture. Analysis of these models showed that images of the early flame in the combustion cycle do contain information to train regression and CNN models on forthcoming states (i.e., CA10, CA50); however, there were significant limitations on how far into the burn process these predictions remained accurate due to the complex thermal physics (i.e., CA90). Future research should address the need to increase the quantity of training data or to introduce additional measurements, such as simultaneous images from multiple camera angles and flame propagation late in the combustion cycle.