Task Transfer Learning for Prediction of Transient Nitrogen Oxides, Soot, and Total Hydrocarbon Emissions of a Diesel Engine

According to previous studies on internal combustion engine applications using deep learning, deep learning model should be individually optimized and trained to predict different phenomena. This study introduces task transfer learning to predict transient nitrogen oxides (NO
$_{\mathrm {x}}$
), soot, and total hydrocarbon (THC) emissions, which are the major emissions from diesel engines. Using the concept of task transfer learning, when there is a pretrained model relevant to the target task, the model can be transferred to predict another phenomenon by training only the last two layers with hyperparameters of the pretrained model. This concept omits the need for optimizing and training separate models that can save computational time and cost. The results of task transfer learning were evaluated using Worldwide Harmonized Light Vehicles Test Procedure (WLTP) cycle data, which are representative transient cycles of the internal combustion engine, and all possible transfer cases with NOx, soot, and THC emissions were investigated. The R2 values of pretrained NOx, soot, and THC models were 0.9780, 0.9215, and 0.9390, respectively. The R2 gaps between the pretrained and transferred models were within 0.012, with a value of 0.0015 for the NOx emission, 0.011 for the soot emission, and 0.0115 for the THC emission. The relative mean absolute errors (MAEs) to the maximum emission values were approximately 0.57-0.82% for NOx emissions, 0.69-2.02% for soot emissions, and 1.52-2.42% for THC emissions. These accuracy results were comparable to the accuracy of the emission measurement device, which was better than that of the sensors for practical use in vehicles. The results indicated that task transfer learning was valid for predicting emissions of an internal combustion engine, and it achieved efficient organization of prediction models using a pretrained model.