Abstract
We propose methods which train deep learning models (DNN, Deep Neural Net) to separate knocking sounds from engine radiation noise measured by a microphone. These DNNs contribute to the automation of ignition timing calibration for the gasoline engine by evaluating the intensity of knocking. The previous method has two problems. First, separation performance deteriorated for engines not included in the training data. Second, this method required in-cylinder pressure for training. In this paper, we propose a method that can separate the knocking sounds of engines not included in training data and an unsupervised method that does not require in-cylinder pressure.
