Search Results

Many excellent papers have been written about the subject of estimating engine-out NOx on diesel engines based on real-time available data. The claimed accuracy of these models is typically around 6-10% on validation data sets with known inputs. This reported accuracy typically ignores input uncertainties, thus arriving at an optimistic estimate of the model accuracy in a real-time application. In our paper we analyze the effect of input uncertainty on the accuracy of engine-out NOx estimates via a numerical Monte Carlo simulation and show that this effect can be significant. Even though our model is based on an in-cylinder pressure sensor, this sensor is limited in its capability to reduce the effect of other measured inputs on the model.

The advancement of Machine-learning (ML) methods enables data-driven creation of Reduced Order Models (ROMs) for automotive components and systems. For example, Gaussian Process Regression (GPR) has emerged as a powerful tool in recent years for building a static ROM as an alternative to a conventional parametric model or a multi-dimensional look-up table. GPR provides a mathematical framework for probabilistically representing complex non-linear behavior. Today, GPR is available in various programing tools and commercial CAE packages. However, the application of GPR is system dependent and often requires careful design considerations such as selection of input features and specification of kernel functions. Hence there is a need for GPR design optimization driven by application requirements. For example, a moving window size for training must be tuned to balance performance and computational efficiency for tracking changing system behavior.