Parameter sensitivity analysis for diesel spray penetration prediction based on GA-BP neural network

Abstract

Machine learning has started to be used in engine research to optimize combustion and predict fuel spray characteristics. This paper presents the development of a machine learning model using a Genetic Algorithm-Backpropagation (GA-BP) neural network to predict spray penetration. The GA-BP neural network was selected for its ability to optimize neural network weights and thresholds, thereby improving model convergence and avoiding local minima, which are common challenges in complex, non-linear problems such as spray prediction. The model was trained using experimental data from diesel injector spray tests, and its accuracy was evaluated through parametric sensitivity analysis, examining the influence of various input factors. A comparison between the machine learning model and the traditional empirical formulas of spray penetration revealed that the machine learning model achieved greater accuracy. In terms of the sensitivity to inputs, it is interesting to find that the cognition of machines is different from that of humans. When an input parameter does not have any functional relationship with other input parameters, the absence of this input parameter will lead to a significant decrease in the accuracy of the output result. The results demonstrate that the machine learning approach offers higher accuracy and better generalizability compared to traditional empirical methods. This study recommends the ways to get better results of penetration prediction with BP neural networks, which is efficient in training and utilizing Artificial Neural Networks (ANNs).