Deep Learning-based Image Analysis Method for Estimation of Macroscopic Spray Parameters

Fran Huzjan, Filip Juric, Sven Loncaric, Milan Vujanovic

Abstract: Spray strategies contribute to the overall engine efficiency, combustion process, and reduction of pollutant formation in internal combustion engines with direct rail injection systems. Spray shape is determined by several parameters such as nozzle diameter, injection pressure, injector geometry, and cylinder type, which influence spray macroscopic parameters. The spray macroscopic parameters are commonly used to describe the input parameters of numerical simulations. Three main spray macroscopic parameters are spray cone angle, penetration length, and spray area. In this paper, we propose a method for spray macroscopic parameter estimation that achieves the state-of-the-art results. To obtain these results, image segmentation was performed to separate the spray from the rest of the image. Then the spray macroscopic parameters are estimated from the segmented image. To perform image segmentation Min U-Net is used. Min U-Net is a novel lightweight deep learning neural network based on U-Net. Min U-Net achieves the state-of-the-art segmentation results while having more than 500 times fewer parameters and being at least twice as fast as other learning-based methods. To evaluate the proposed method, an available dataset containing a variety of images with various spray shapes and orientations. The experiments performed on the dataset showed that Min U-Net achieved a mean dice coefficient of 0.95 with an inference time of 11.94 ms/image. The spray macroscopic parameters estimation is also highly accurate, with spray cone angle having an error of 1.08\(^{\circ }\), spray penetration length having a relative error of 5.95%, and spray area having a relative error of 4.05%.