A prediction model for exhaust gas regeneration (EGR) clogging using offline and online machine learning

Abstract

The exhaust gas regeneration (EGR) also called exhaust gas re-circulation system in an engine of construction machine (CM) often gets clogged due to various ways of driving the machines. Currently, there does not exist any model that can predict clogging for maintenance planning. Hence, clogging is only recognized when it has occurred, and often causes the CM to drop out. Engines still operated despite clogging causes frequent cold engine running, and excessive exhaustion of nitrogen, which leads to loss of the engine’s performance and reduces their lives. We propose an approach that builds on virtual key sensors. Virtual key sensors are usually used to replace real sensors. However, we propose to compare the virtual and the real sensor outcomes. If differences between the estimated and the real value emerge, we assume changes of the systems because of, e.g., clogging or leakage in pipes. EGR pressure is identified as an important sensor to estimate clogging. A virtual sensor of EGR pressure is built from other real sensors based on a polynomial regression model [1]. The error between the real and the virtual EGR pressure sensors varies between 5-10% depending on the driver’s behaviors. The model discriminates the ideal ways of working and abnormalities. Moreover, we suggest to adapt the weights of the regression model to other engine types of the same engine family based on online stochastic gradient descent algorithm. Since the deployed regression and adaptation algorithms are computationally inexpensive, the approach could be applied using existing CM micro controllers.