Deep Learning-Based Cone Angle Estimation Using Spray Sequence Images

Engine efficiency, combustion process, and gas emissions are greatly affected by spray strategies. Spray strategies are utilized in engines with internal combustion. Spray strategies are determined by parameters such as nozzle diameter, injection pressure, chamber pressure, cylinder type, and others. These parameters determine spray shape. Spray shape is established by three main spray macroscopic parameters which are cone angle, penetration length, and spray area. Spray cone angle, with other spray macroscopic parameters, is often used to describe the parameters of numerical simulations. In this paper, we propose two new methods for the estimation of spray cone angle which affects the air engulfing and mixing process. Spray images gathered during a single spray injection are highly correlated. To use this fact to our advantage we proposed two deep learning-based methods that use image sequence as input. StackNet is a regression neural network that stacks images and uses them as input. It also uses a feature extractor and a fully connected layer. CNN-LSTM is another regression neural network with a feature extractor, but it utilizes Long Short-Term Memory (LSTM) cells before a fully connected layer. Both of the methods were trained, validated, and tested on preprocessed sequence images. To achieve better generalization and more data diversity, data augmentation was used. Three state-of-the-art feature extractors were tested, VGG16, MobileNetV3, and EfficientNetB0. The proposed methods were compared with the baseline approach which uses a single image as an input. Experimental validation showed that StackNet with VGG as a feature extractor achieved the best result. The proposed method estimated cone angle with a mean absolute error of 0.505 degrees, which is more than two times more accurate than the best baseline approach.