Abstract
Keywords: Learning systems, exhaust emissions, Support Vector Machines
Diesel engines are decisive in lowering fuel consumption. This contributes to their increasing market and leads to challenging developments in order to fulfill emission regulations by providing, for example, the lowest possible CO2 emission while ensuring good drivability.
As a result, diesel engines have been equipped with sophisticated electronically controlled actuators based on lookup tables. These tables must be set up in order to optimize the trade off between the different emissions and fuel consumption. The main goal of engine mapping is to minimize fuel consumption while verifying emission regulations under vehicle tests such as FTP or EUDC cycles.
Actually, engine mapping is realized for some specific speed and load set points that are representative of the test cycle. This is the steady state approach where exhaust emissions local models are set up for each set point. These models are used to determine optimal control parameters satisfying our goal.
This steady state approach presents some limitations: first, considering that models do not exist for every operating point, corresponding control parameters are actually determined by smoothing the optimal parameters computed for several (Speed, Load) set points that are representative of the test cycle. The second limitation is the ability to estimate emissions over the whole cycle. This requires a global model that can be used to "drive" the (Speed, Load) cycle points, through fuel and emissions maps generated from the global model.
The aim of this paper is to present a new method for determining the global model. This model relies on estimated local models and offers good generalization performance over the whole (Speed, Load) domain. This model can also be used to minimize the number of operating points measured on the test bench.
We propose a two stage modeling technique. First, local models are generated for every predetermined set point; second, they are combined by using a mixture statistic with the help of Support Vector Machines. The global model accuracy is estimated by a leaving one out strategy.
This paper is organized as follows: in the next sections, we first present the industrial application and the actually used methodology for Design of Experiments and emission modeling. Second, we present the proposed Support Vector Machines mixture of local models. We then apply the proposed methodology on a toy problem and finally on the exhaust emissions modeling problem. The last section gives some concluding remarks.