Automotive engine power performance tuning under numerical and nominal data

Abstract

Modern automotive engines are controlled by an electronic control unit (ECU), and engine power performance is significantly affected by the selection of both ECU parameters and engine components. The engine performance tuning is usually done by a trial-and-error method. In the current literature, very little research has considered the selection of engine parts because engine parts are complicated objects that are usually represented as nominal data. These data are meaningless values in terms of computation. This paper presents a novel multiple-input/output least-squares support vector machine plus one-of-n remapping method for modelling engine power performance using both numerical (ECU parameters) and nominal data (candidate engine parts). The Quasi-Newton method, a genetic algorithm and particle swarm optimisation are then applied to the engine model to determine the optimal engine setup automatically. A simple binary code synthesis rule is also proposed to optimise the nominal variable. Both experimental and simulation results show that the proposed methodology can successfully yield an optimal engine setup.

Introduction

Modern automotive engines are four-stroke, electronic fuel-injection engines controlled by an electronic control unit (ECU). The engine power performance is significantly affected by the setting of control parameters in the ECU and the selection of engine parts. Engine power performance tuning thus requires an adjustment or modification of both ECU parameters and engine parts to yield optimal performance based on different user requirements. In practice, automotive engine power performance tuning involves a compromise between maximum engine torque, minimum fuel consumption and reduced emissions. The ECU parameters, such as fuel injection time and ignition timing (usually formulated in look-up tables), are classified into numerical variables. Herein, “engine parts” refers to optional or performance parts such as intake systems, ignition systems, exhaust systems, camshafts, headers and others. Typically, engine parts are complicated and nonadjustable objects, but they can be replaced with other high-performance parts. The engine parts are normally represented as models or brand names which are meaningless values (i.e., nominal values) in terms of computation, so each engine part is classified as a nominal variable. Of course, the ECU parameters must also match the selected engine parts. This kind of problem is common in aftermarket engine performance tuning and remains very challenging.

In the current literature, very little research has considered the selection of engine parts, but studies on ECU parameter setup (also called ECU calibration) are available. This situation exists because all prior studies have assumed that the engine parts were chosen perfectly by the manufacturer. In fact, the selection of engine parts is quite important for engine power performance, yet it is very difficult to arrive at an optimal configuration. Moreover, there has been no comprehensive discussion of integrating engine-part selection with ECU calibration, although these two factors are both very important in engine power optimisation problems.

Traditionally, ECU calibration is done empirically through testing on the dynamometer (dyno) because the automotive engine is an integrated system of sensors and signal processing, thermofluid, mechanical and computer control systems (Li, 2007, Robert Bosch GmbH, 2004). The relationship between the input and output parameters of an electronically controlled automotive engine forms a complex multivariable nonlinear function that is very difficult to determine. Hence, in current practice, the engineer first determines an engine setup based on experience, past data and simple mathematical equations, and the engine is then run on a dynamometer to test its actual performance. If engine performance is poor, the engineer adjusts the settings and repeats the procedure until the performance is satisfactory. As a result, vehicle manufacturers normally spend 12 months tuning an engine to optimise it for a new car model (Bell, 2006). Moreover, the performance function is also engine dependent; every engine must undergo a similar tuning procedure.

Considering the above limitations of traditional ECU calibration, several computer-aided ECU calibration methods have been proposed in recent years. The most common approach is simulation-based calibration (Rask and Sellnau, 2004, Tan et al., 2006), which divides the calibration process into two parts: modelling and optimisation. In terms of modelling, a mathematical engine model can be constructed either from empirical models (Errico et al., 2011, Qi et al., 2011, Saerens et al., 2009, Zhang et al., 2010) or the many proprietary modules found in commercial engine simulation programs such as Engine Analyser Pro, GT Power and WAVE's 1-D cycle simulation (Wu, Filipi, Prucka, Kramer, & Ohl, 2007). The empirical models and simulation software are derived by resorting to basic physical laws. These mathematical models can simulate engine performance and thus can be employed as objective functions. The optimal ECU parameters can then be determined based on the objective functions by using computer-aided optimisation methods such as gradient search methods (Wu et al., 2007) and genetic algorithm-based methods (Errico et al., 2011). Hence, the amount of time and resources required for engine development can be significantly reduced. However, the predictive accuracy of the mathematical engine models is poor due to the following facts:

• To predict engine performance, the simulation software and mathematical models require the engineer to provide engine-specific parameters such as combustion efficiency. In practice, it is too demanding to obtain or estimate such parameters, particularly for mechanics who do not work for the manufacturer.

• Excessive numbers of assumptions have been made in these models. This renders such engine models overly simple in comparison with real engine systems.

Several recent studies have described the use of neural networks (NN) to process collected experimental data and generate engine models (Celik and Arcaklioglu, 2005, Dickinson and Shenton, 2009, Garcia-Nieto et al., 2008, Togun and Baysec, 2010). In this way, many computer-aided optimisation methods can then be applied to search for the optimal engine setting automatically. However, there are two main drawbacks to the neural network modelling approach:

• The architecture, including the number of hidden neurons, must be determined a priori or modified during training by heuristics, which results in a sub-optimal network structure.

• During the training process, neural networks can easily become stuck in local minima. Various ways of avoiding local minima (e.g., early stopping and weight decay) have been employed. However, these methods greatly affect the generalisation of the estimated function (Haykin, 1999).

Vong, Wong, & Li (2006) showed that the emerging technique of least-squares support vector machines (LS-SVM) (Cristianini and Shawe-Taylor, 2000, Vapnik, 1998) combined the advantages of neural networks (particularly their ability to handle a large amount of highly nonlinear data) and nonlinear regression (with its high capacity for generalisation) to achieve better accuracy than traditional NNs for automotive engine torque estimation. Subsequently, we proposed the application of multi-input/multi-output least-squares support vector machines (MIMO LS-SVM) and a genetic algorithm (GA) to determine the optimal setpoints for a proportional-integral-derivative idle-air valve controller and some engine control variables for automotive engine idle-speed control (Wong, Tam, Li, & Wong, 2008), which is a very difficult multi-objective optimisation problem. The integrated approach can overcome all of the limitations of the existing ECU calibration methods mentioned above, but it is not suitable for engine-part selection because engine parts are usually represented as nominal data that are meaningless values in terms of computation.

In view of the limitations of the current engine performance tuning approaches, an advanced data preprocessing technique (namely, one-of-n remapping) is introduced to handle the nominal data in this work. By integrating this data preprocessing technique and the MIMO LS-SVM modelling framework we proposed in our previous study (Wong et al., 2008), engine power performance can be simulated with different combinations of engine parts and ECU parameters. After building the engine performance model, an advanced global optimisation method, particle swarm optimisation (PSO), is then applied to the MIMO LS-SVM model to determine the optimal engine parts and ECU setup automatically. Additionally, the traditional optimisation methods for ECU calibration such as the Quasi-Newton method (QNM) (Wu et al., 2007) and GA (Errico et al., 2011, Wong et al., 2008) are also applied for comparison purposes. The QNM and GA are representative of gradient-based methods (i.e., iterative methods) and population-based methods (i.e., heuristic algorithms), respectively. In the current literature, there has been no comparative study applying these two kinds of optimisation methods to MIMO LS-SVM models. This work is the first in the literature to conduct this comparative study.

The objective of this study was to develop a comprehensive and reliable engine modelling methodology, so that by integrating any proper computer-aided optimisation method with an engine model, both the optimal engine parts and ECU settings can easily be determined. The modelled car engine then only requires a dyno test for verification after obtaining a satisfactory configuration from the integrated approach. Hence, a number of unnecessary dyno tests for the trial setup can be eliminated, considerably reducing testing time and cost.