Date of Award

6-18-2021

Publication Type

Doctoral Thesis

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

Keywords

Artificial neural network, Extreme learning machine, Gasoline engine model, Torque prediction

Supervisor

Mohammed Khalid

Rights

info:eu-repo/semantics/embargoedAccess

Abstract

Traditionally the internal combustion engines and their subsystems are modeled purely based on their physical/mathematical principles. Such modeling techniques usually require deep prior knowledge of the internal combustion engine, which is often too difficult for many non-engine experts. In addition, the modeling process is usually very complicated and time-consuming. In some cases, the models may not be useful for many real-world applications due to oversimplified modeling assumptions. In recent years, with the rise of artificial intelligence technologies, the neural network based internal combustion engine modeling techniques have gained increasing popularity. In contrast to the traditional internal combustion engine modeling approaches, the neural network based methods can create the models directly from the system data instead of from the complicated physical/mathematical equations. This type of approach is easier to handle and often has fewer parameters to tune. This dissertation presents an extreme learning machine based neural network modeling technique for gasoline engine torque prediction. The technique utilizes a single-hidden layer feedforward neural-network structure that has the potential to approximate any continuous function with high accuracy. To verify the robustness of this technique, over 3300 data points collected from a real-world gasoline engine were used to train and test the model. The data points spanned from 1000 rpm to 4500 rpm engine speed, idle to full engine load, which mirrored the full map of normal engine operating conditions. The experimental results demonstrate that the created model predicts the gasoline engine torque with high accuracy. Furthermore, this research proposed a weight factor approach to further improve the model accuracy in the desired data regions without modifying the input data set. The model evaluation showed that the weight factor approach could reduce the overall prediction errors in the desired regions significantly. This feature is particularly useful in tuning the performance of the model when the significance of the individual data points varies, or when the distribution of the data points is imbalanced. Moreover, an innovative form of extreme learning (referred to as progressive extreme learning machine) was proposed and evaluated. It was capable of gradually improving the estimation accuracy with recursions. The new algorithm maintained the random weights generation feature of the traditional extreme learning machine and upheld the training speed advantage over many other competing algorithms. The experimental evaluation results show that progressive extreme learning machine has higher accuracy and superior generalization than many other extreme learning machine based algorithms. Furthermore, its performance was also compared with some nonlinear machine learning algorithms using the publicly available data sets. The experimental evaluation results showed that the progressive learning machine outperformed the support vector regression and had comparable performance with Levenberg-Marquardt Algorithm.

Download

Share

COinS