Diesel engine emissions prediction using parallel neural networks

Emission legislation has forced the pace of development of engine management functions. Legislation that will be applied to diesel engines during the period 2010-2020 continue to put great emphasis on both nitrogen oxides NO_x and particulate matter (PM). With the increasing effort to reduce emissions and maintain fuel economy manufacturers are focussing on engine control. Engine control requires data acquisition and acquisition requires sensors, but hardware in the form of sensors adds further cost to the production. As a result, so called virtual sensors are introduced. These are estimators that predict the required data, which is costly to measure or simply incapable of measurement. In this paper a parallel neural network structure is built. It consists of three non-linear autoregressive exogenous input (NLARX) neural network models used to predict the smoke emissions of a diesel engine operated in a non-road-transient cycle. Existing resources from Matlab toolboxes are used in order to monitor both the cost and computational expenses of analysis. The data is re-ordered into training and validation sets and processed. To overcome the weakness of the neural network approach in respect of high frequency signals, the data is divided into layers to split up the frequencies and cut high amplitudes. Three horizontal layers of the signal are processed in parallel through independent NLARX-models and their performances are added to give an overall result.