Modeling diesel oxidation catalyst exhaust gas temperatures using long short-term memory recurrent neural networks

Nowadays sensors are essential in powertrain engineering and exhaust aftertreatment systems due to the increasing need for high performance and fewer emissions. However, cost increases proportionally with the number of sensors in the vehicle. Mostly in diesel oxidation catalysts, the upstream and downstream temperature sensors are attached to the vehicle until their models get calibrated then removed in the end-user version. In some other cases, the vehicle is delivered to the consumer containing those sensors. The modeling process is both extravagant and time-consuming as it requires engine and chassis dynamometers. On the other hand, accurate sensors are expensive and require extra plausibility diagnoses and coherence monitoring. The importance of those models come from the truth that the oxidation catalyst efficiency is calculated from the difference between the downstream and upstream temperature sensor measurements in the time of diesel regeneration. The greater the difference is, the more efficient the oxidation catalyst becomes. Concretely, efficiency determines the aging level of the catalyst and yet, the amount of CO emission. As another reason for temperature models, the emission regulations obligate the monitoring of CO. Also, the other exhaust after-treatment system components use those models as inputs to calculate their efficiencies. In the past few years, deep structured learning became applicable for many sectors. After the development of the recurrent neural networks, language modeling, machine translation, image captioning, handwriting generation, and question answering became the most common applications. Later, the long short-term memory networks have proposed a solution to the gradient vanishing and exploding problem which included time series prediction and complex nonlinear modeling to the applications of recurrent neural networks. The purpose of this study is to develop a model for the diesel oxidation catalyst upstream and downstream temperatures by using long-short term memory networks. Measurements from engine sensors and actuators position feedback were recorded from a vehicle and used as training and validation data. After enough training, the model was utilized to evaluate and predict the modeled oxidation catalyst upstream and downstream temperatures. The training data used was supplied by AVL as well as the post-processing environment.