Transfer learning based prediction of knock intensity in a hybrid dedicated engine using higher-octane gasoline for thermal efficiency improvement

Higher-octane gasoline has the potential to improve engine thermal efficiency via suppressing knock but requires re-calibration. Re-calibration consumes enormous experiment efforts to characterize knock intensity, and the available efforts are increasingly limited by the tight development schedule due to the fierce competition among automotive companies. To fully utilize the efficiency-improving potential of higher-octane gasoline with limited experimental efforts, this paper proposes a new modelling approach for knock intensity prediction, termed expertise-guided adaptive transfer learning. Different from conventional data-driven modelling, which utilizes a linear sampling strategy to acquire training samples for a non-transfer neural network, this approach utilizes an expertise-guided sampling strategy to acquire representative training samples for a domain adaptive neural network. Two gasoline fuels with different octane numbers were tested in a hybrid dedicated engine, where the knock intensity under swept spark advances was measured. A transfer learning model was established through the proposed modelling approach to predict knock intensity, with the constraint of which the engine control parameters were optimized. A significant domain discrepancy of the dataset was found, which made transfer learning based on fine-tuning produce negative transfer, while the transfer learning based on the domain adaptative neural network produced positive transfer. The proposed methodology reduced the prediction error of knock intensity by 50.5 % compared with the conventional methodology. Increasing the research octane number from 93.1 to 98.0 increased the engine efficiency by 2.0 %, from 36.9 % to 38.9 %, where 0.3 % benefited from the lower prediction error of knock intensity by the proposed model.