Fine-tuning Transfer Learning for Knock Intensity Modeling of an Engine Fuelled with High Octane Number Gasoline

Despite the potential of high octane number (ON) gasoline in knocking prohibition and fuel efficiency improvement, its application to the existing engines requires recalibration to release its full advantages, where the optimal control parameters, constrained by knock intensity, are re-modification. To reduce experimental efforts for obtaining knock intensity, this paper proposes a knock intensity modeling approach of fine-tuning transfer learning based on multilayer perceptron for the engine fueled with high-ON gasoline. The multilayer perceptron model is pre-trained with the experimental data of the engine fueled with the base gasoline (92#), which have been obtained during the previous engine development and thus consume no more experimental efforts. Then, the model is fine-tuned with the experimental data of the high-ON gasoline (98#) by freezing the first two hidden layers and updating the weights and biases of the last four hidden layers. By a comprehensive study, the results demonstrate that the developed approach can achieve competitive prediction accuracy whilst significantly reducing experimental efforts for 30% of 98# data.