Using Deep Learning to Predict the Engine Operating Point in Real-Time 2021-01-0186
The engine operating point (EOP), which is determined by the engine speed and torque, is an important part of a vehicle's powertrain performance and it impacts FC, available propulsion power, and emissions. Predicting instantaneous EOP in real-time subject to dynamic driver behaviour and environmental conditions is a challenging problem, and in existing literature, engine performance is predicted based on internal powertrain parameters. However, a driver cannot directly influence these internal parameters in real-time and can only accommodate changes in driving behaviour and cabin temperature. It would be beneficial to develop a direct relationship between the vehicle-level parameters that a driver could influence in real-time, and the instantaneous EOP. Such a relationship can be exploited to dynamically optimize engine performance. In this work, we investigate deep learning methods to correlate vehicle-level vectors, including parameters which define driving behaviour, with instantaneous EOP. We analyze the performance of these methods using field data obtained from a 2019 Cadillac XT6 test vehicle driven on highways that contain both straight and sharp curvature segments. We investigate two deep learning methods that are particularly suitable for modelling time-series data: Nonlinear Autoregressive Exogenous (NARX) and Long Short-Term Memory (LSTM) models. The models are trained by using a 5-minute contiguous snippet of data. We compare the predicted versus measured instantaneous EOP over many 10-second continuous snippets chosen randomly outside the training segment. The NARX model outperforms the LSTM model in three traditional ways of comparing signals, namely, signal-to-noise ratio (SNR) values, first-order derivatives (FOD), and root-mean-squared error (RMSE).