Investigation of ANN and SVM based on limited samples for performance and emissions prediction of a CRDI-assisted marine diesel engine. (25th January 2017)

• Record Type: Journal Article
• Authors: Niu, Xiaoxiao; Yang, Chuanlei; Wang, Hechun; Wang, Yinyan
• Is Part Of: Applied thermal engineering. Volume 111(2017:Jan.)
• Page Start: 1353
• Page End: 1364
• Publication Date: 2017-01-25
• Subjects: Marine diesel engine -- CRDI -- Prediction -- ANN -- SVM -- Taguchi orthogonal array; Heat engineering -- Periodicals; Heating -- Equipment and supplies -- Periodicals; Periodicals; 621.40205
• DOI: 10.1016/j.applthermaleng.2016.10.042
• Languages: English
• ISSNs: 1359-4311

Abstract

Highlights: Performances of ANN and SVM vary with different application areas. Predictive accuracy of ANN differs from the network architecture. Predictive accuracy of ANN is unstable due to random initial parameters. SVM has an excellent performance with a small amount of training data. Abstract: Artificial Neural Network (ANN) and Support Vector Machine (SVM), due to their accuracy and ability to analyze nonlinear problems, have been applied in many research areas. However, their performances vary with the application area. This paper investigates the performances of these two approaches for the responses prediction of a common rail direct injection system (CRDI)-assisted marine diesel engine. Moreover, considering that the experiments of marine diesel engines are always time, money and energy consuming, the main purpose of this study is to determine the better predictive approach based on a small amount of training data. The Taguchi orthogonal array is employed for the operating points determination of training data; then, based on the same training data, which contain only 25 samples, the predictive performances of ANN and SVM are evaluated and compared. The comparison of ANN and SVM indicates that with limited experimental data, SVM can find the optimal global solution and has excellent predictive accuracy and generalization capability, while ANN may converge to local minima and face the overfitting problem. Eventually, this study suggests that SVM is well-suited for application to diesel engine response predictions and will reduce the experimental cost significantly.