Abstract:

A neural-network-based intelligent knock prediction method was developed based on neural-network, by using a calibrated 1-dimension simulation model and engine test bench data, to predict knock in a spark-ignition (SI) engine with a compression ratio of 14. Two neural network architectures with multilayer perception were constructed and compared, including an end-to-end model and a two-stage pipeline model. The end-to-end model directly used control parameters for knock prediction. The pipeline model followed an explicit physical reasoning chain of “control parameters, combustion state, knock prediction”. Some validation tests were performed on an engine bench. The results show that the pipeline model achieves a prediction accuracy of 98.6%, which is superior to the end-to-end model's accuracy of 96.1%. The model achieves 100% prediction accuracy for 20 actual control parameter sets. Therefore, this model can predict knock events in real engines and will provide a solution for real-time engine control and performance optimization.

Key words: spark ignition (SI) engines, knock prediction, control parameters, combustion process parameters, neural networks, multilayer perceptron