Predição da Emissão de Particulado Utilizando Algoritmo de Aprendizagem de Máquina

Abstract

One of the requirements of diesel engines certification process is that the engine do not exceed the specific particulate emissions limit in cycle ( _ ). To calculate the _ , a manual process is required. The total amount of mass impregnated in the particulate filter is obtained by weighting the filter on a balance after each test. Therefore, it is not possible to obtain it without human intervention. In order to allow test rigs to operate in automatic mode, without an operator conducting the tests, an automatic way of calculating _ is required. Thus, the aim of this work is to develop a _ prediction model that does not require human intervention. For this, a machine learning approach, based on the random forests algorithm, is used. Data collected from 2012 to 2019 from three test cells of 11 and 13 liters diesel engines of an automotive company, summing up 2500 valid test results, are used as input. This data are employed to train the algorithm and build a prediction model. The prediction model is then validated using another 72 validation tests results. The accuracy of the final model considering a confidence interval of 95% is ±3,00 mg/kWh for the European Transient Cycle, and ±1,96 mg/kWh for the European Stationary Cycle.