Automated Calibration of Heavy-Duty Low NO _x Aftertreatment System Controls using Deep Learning and Global Optimization Methods

Heavy duty diesel engines provide a robust power plant for transportation applications for both on highway and off road applications. Control of criteria pollutants such as particulate matter and NO_x at tailpipe for these applications based on standards set by regulatory bodies such CARB and EPA is critical. SwRI has demonstrated capability to achieve 0.02 g/bhp-hr. tailpipe NO_x standard through the application of a model based controls in EPA and CARB funded projects. This control mechanism enables precise urea dosing for both steady state and transient conditions by leveraging estimated ammonia storage state in a dual dosing system using a set of chemical kinetics-based SCR observer models. This controller is highly nonlinear, with a significant amount of controller tuning with up to 55 calibratable parameters. In order to improve the accuracy and reduce the time required for calibration of this controller, this work proposes the deployment of a Deep Learning-based SCR plant model in conjunction with a Genetic Algorithm based optimization script in a closed loop with the low NO_x controller that can enable identification of near-optimal controller calibration in a simulation environment. This work describes the optimization framework and its validation with real-world experimental data. The calibration determined in this framework was able to achieve 0.02 g/bhp-hr for regulatory cycles consisting of CFTP, HFTP, RMC, and LLC. A quantitative and qualitative analysis on the results from this proposed framework is performed and is compared against existing expert driven manual calibration process.