TY - JOUR
T1 - Systematic hyperparameter selection in Machine Learning-based engine control to minimize calibration effort
AU - Garg, P.
AU - Silvas, Emilia
AU - Willems, Frank P.T.
PY - 2023/11
Y1 - 2023/11
N2 - For automotive powertrain control systems, the calibration effort is exploding due to growing system complexity and increasingly strict legal requirements for greenhouse gas and real-world pollutant emissions. These powertrain systems are characterized by their highly dynamic operation, so transient performance is key. Currently applied control methods require tuning of an increasing number of look-up tables and of parameters in the applied models. Especially for transient control this state-of-the-art calibration process is unsystematic and requires a large development effort. Also, embedding models in a controller can set challenging requirements to production control hardware. In this work, we assess the potential of Machine Learning to dramatically reduce the calibration effort in transient air path control development. This is not only done for the existing benchmark controller, but also for a new preview controller. In order to efficiently realize preview, a strategy is proposed where the existing reference signal is shifted in time. These reference signals are then modeled as a function of engine torque demand using a Long Short-Term Memory (LSTM) neural network, which can capture the dynamic input–output relationship. A multi-objective optimization problem is defined to systematically select hyperparameters that optimize the trade-off between model accuracy, system performance, calibration effort and computational requirements. This problem is solved using an exhaustive search approach. The control system performance is validated over a transient driving cycle. For the LSTM-based controllers, the proposed calibration approach achieves a significant reduction of 71% in the control calibration effort compared to the benchmark process. The expert effort and turbocharger experiments used in calibrating transient compensation maps in physics-based feedforward controller are replaced by little simulation time and parametrization effort in ML-based controller, which requires significantly less expert effort and system knowledge compared to benchmark process. The best trade-off between multi-objective cost terms is achieved with one layer and 32 cells LSTM neural network for both non-preview and preview control. For non-preview control, a comparable control system performance is achieved with the LSTM-based controller, while 5% reduction in cumulative NOx emissions and similar fuel consumption is achieved with preview controller.
AB - For automotive powertrain control systems, the calibration effort is exploding due to growing system complexity and increasingly strict legal requirements for greenhouse gas and real-world pollutant emissions. These powertrain systems are characterized by their highly dynamic operation, so transient performance is key. Currently applied control methods require tuning of an increasing number of look-up tables and of parameters in the applied models. Especially for transient control this state-of-the-art calibration process is unsystematic and requires a large development effort. Also, embedding models in a controller can set challenging requirements to production control hardware. In this work, we assess the potential of Machine Learning to dramatically reduce the calibration effort in transient air path control development. This is not only done for the existing benchmark controller, but also for a new preview controller. In order to efficiently realize preview, a strategy is proposed where the existing reference signal is shifted in time. These reference signals are then modeled as a function of engine torque demand using a Long Short-Term Memory (LSTM) neural network, which can capture the dynamic input–output relationship. A multi-objective optimization problem is defined to systematically select hyperparameters that optimize the trade-off between model accuracy, system performance, calibration effort and computational requirements. This problem is solved using an exhaustive search approach. The control system performance is validated over a transient driving cycle. For the LSTM-based controllers, the proposed calibration approach achieves a significant reduction of 71% in the control calibration effort compared to the benchmark process. The expert effort and turbocharger experiments used in calibrating transient compensation maps in physics-based feedforward controller are replaced by little simulation time and parametrization effort in ML-based controller, which requires significantly less expert effort and system knowledge compared to benchmark process. The best trade-off between multi-objective cost terms is achieved with one layer and 32 cells LSTM neural network for both non-preview and preview control. For non-preview control, a comparable control system performance is achieved with the LSTM-based controller, while 5% reduction in cumulative NOx emissions and similar fuel consumption is achieved with preview controller.
KW - Engine control calibration
KW - Long short-term memory neural network
KW - Machine Learning
KW - Preview control
UR - http://www.scopus.com/inward/record.url?scp=85170279409&partnerID=8YFLogxK
U2 - 10.1016/j.conengprac.2023.105666
DO - 10.1016/j.conengprac.2023.105666
M3 - Article
SN - 0967-0661
VL - 140
JO - Control Engineering Practice
JF - Control Engineering Practice
M1 - 105666
ER -