An image understanding based model with ion current signals for predicting combustion information

Abstract

With the gradually stringent emission regulation, it is in needs for new methods to optimize conventional combustion engines in terms of exhaust emission, working performance, and abnormal combustion. In response, cheaper, more reliable, and more responsive engine control schemes based on the ion current detection method appear, and a challenge in the method is to retrieve combustion information involving max-pressure and knock condition from ion current signals. To cope with the challenge, we develop an image understanding based FusionNet model that transforms ion current signals to spectrograms and takes the spectrograms to predict max-pressure and knock condition simultaneously. As a result, FusionNet can predict the crank angle and the numerical value of the max-pressure of samples in the test set, with an average of the Mean Squared Error valuing 6.802 and 7.142 respectively. Moreover, FusionNet can predict pressure oscillation related to the knock condition with an average of the Cosine Similarity valuing 0.00932, and apply the detection result of the oscillation to predict the knock condition of samples in the test set, with an average of the F1-Score valuing 0.92684.