A genetic algorithm to optimize the adaptive Support Vector Regression model for forecasting the reliability of diesel engine systems

Abstract

This paper presents the use of the Support Vector Regression (SVR) technique to forecast the reliability of a system. Future predicted values of system reliability are highly sensitive to the choice of SVR parameters and the type of kernel SVR function. The dataset of a turbocharged diesel engine system was used as a case study. The Normalize Root Mean Square Error (NRMSE) measure was used to evaluate the SVR model in predicting the reliability of the system. Many experimental attempts were done using the optimal SVR parameters and the proper kernel function. Results showed that Order 5 of the polynomial kernel outperformed both Gaussian and linear kernel functions in predicting the future reliability values with minimal NRMSE. Experimentally, choosing the proper parameters for the SVR is a hard process, and there are no guarantees that the good parameters and the best kernel function are used. Therefore, artificial intelligence must be used. A genetic algorithm (GA) was used as an AI search optimization method to optimize both the SVR parameters and the type of the kernel function by generating a GA-SVR model. The GA successfully optimized the SVR model to ensure accurate predictions. The adaptive GASVR model was used to overcome such problems as small size of the dataset, varying lifetimes of the system components, and odd behavior of the system because of external environmental causes. Results confirmed the efficiency of the adaptive model to predict precisely the reliability of the system, even with a small dataset.