ニューラルネットワークを用いた過給機性能予測

抄録

The volumetric flow rate at surge of compressor of turbocharger is an important index to evaluate its performance. It is difficult to obtain this index by computational fluid dynamics (CFD) because fluid flow inside a turbocharger is complex and unsteady. Therefore, we need to make and test prototypes again and again until target performances are achieved. Increasing cost and time by those tests is one of the important problems. To solve the problem, we develop prediction models for the volumetric flow at a surge point using past experimental data and neural networks, which is one of the machine learning techniques. When training and testing neural networks, we used indices that are obtained from simple one-dimensional simulation models as input parameters in addition to parameters of shape and operational conditions. The mean relative error of the prediction model is roughly 5%, which is better than other machine learning models. The neural networks are useful to the application, and reduce cost and time to predicting the index.