Application of Machine Learning Techniques to Hybrid Engine Restart Prediction

Neural network-based artificial intelligence techniques to replace or augment traditional control systems are gaining in popularity. The present work reports on our evaluation of machine learning techniques to predict engine restarts for a P2 hybrid system. A typical P2 hybrid system operates by launching the vehicle electrically from a standstill and allows low speed electric driving. When the driver accelerator input exceeds an electric-drive power demand threshold, the internal combustion engine is started and connected to the driveline to provide added power. The non-zero time between the driver's power request and engine power delivery is referred to as the tip-in response delay. To minimize this delay, we explore the use of machine learning models to predict an engine restart, implying a driver request beyond the electric-drive threshold. We create deep and recurrent neural networks and train them with engine-on timing data, vehicle trajectory and power demand history. Once trained, the neural network flavors are remarkably accurate at predicting the required engine-on state, from which we can infer driver tip-in. A neural network can thus serve as a simplified embedded model to predict engine starts given a preview of the vehicle's probable trajectory. As a result, we can restart the engine ahead of time to minimize the torque hole during tip-in.