Indicated Torque Reconstruction from Instantaneous Engine Speed in a Six-cylinder SI Engine Using Support Vector Machines

Abstract

This paper proposes the use of support vector machines to reconstruct the indicated torque from crankshaft velocity in a six-cylinder spark ignition engine. Real-time knowledge of indicated torque is typically important for engine diagnostics, and recently, an engine idle speed controller capable of reducing the effects of cyclic combustion variability through the use of indicated torque information was proposed. While measurements of in-cylinder pressure can be used to determine indicated torque, these sensors are generally deemed prohibitively expensive for implementation in a production engine. Estimation methods, particularly traditional model based estimation, are typically computationally expensive and require independent data throughout the cylinder expansion stroke. Overlap of the expansion strokes in a six-cylinder engine complicates the problem and limits the ability of traditional model based approaches in fully reconstructing the torque production process. Intelligent estimation techniques can identify the key characteristics of the instantaneous engine speed as a function of crank angle for the combustion stroke, thereby allowing an optimal subset of the available information to be utilised. A support vector machine approach is used in this paper due to the inherent optimisation of the training procedure, resulting in an optimised prediction matrix. Torque estimation is then obtained through simple mathematical evaluations, allowing fast reconstruction of indicated torque. A comparison between reconstruction performances at different operating points is presented as an evidence of the capabilities of the support vector machines. This paper demonstrates that indicated torque reconstruction with support vector machines in a sixcylinder engine is not only possible, but it can be done with sufficient accuracy for the purpose of advanced engine control applications. INTRODUCTION Combustion in an engine cylinder, even at a fixed operating point, is subject to cycle-by-cycle variations, due to factors such as internal exhaust gas recirculation, uneven air-fuel mixture distribution, and variations in mixture density. During low-load idle conditions, air density and fuel mass fed into the cylinder are at their lowest, thereby increasing the likelihood of uneven distribution of the mixture throughout the cylinder. Consequently, this causes variability in combustion process, and subsequently, indicated torque. As variations in combustion or torque production directly influence engine speed, fluctuations of the crankshaft speed are also increased. These speed fluctuations may lead to unnecessary control action at idle [1], or requirements such as higher idle set points to counteract engine stall, and reduce vehicle noise vibration and harshness quality (NVH). Obviously, increasing idle set points also increases emissions and fuel consumption, and should be avoided. Knowledge of indicated torque generated during the combustion process is desirable for several key applications, such as idle speed control and engine diagnostics. Recently, an engine idle speed controller capable of reducing the effects of cyclic combustion variability was proposed [1]. This relies on information about indicated torque in order to predict and compensate for cyclic variations. While there exist incylinder pressure sensors, and there is a direct correlation between in-cylinder pressure and indicated torque, these sensors are generally deemed too expensive and impractical for implementation in a production line engine. Torque sensors, while available, generally have complicated construction, and their realworld production line performance and reliability are yet to be established. Consequently, there is considerable research into methods that will allow the indicated torque to be inferred from available measurements. Downloaded from SAE International by University of Melbourne, Tuesday, January 07, 2014 06:17:04 PM