Knock Detection in Combustion Engine Time Series Using a Theory-Guided 1-D Convolutional Neural Network Approach

This article introduces a method for the detection of knock occurrences in an internal combustion engine (ICE) using 1-D convolutional neural network trained on in-cylinder pressure data. The model architecture is based expected frequency characteristics knocking combustion. All cycles were reduced to $60^{\circ }$ CA long windows with no further processing applied traces. networks exclusively traces from multiple conditions, labels provided by human experts. best-performing achieves accuracy above 92% all test sets tenfold cross-validation when distinguishing between and non-knocking cycles. In multiclass problem where each cycle was labeled number experts who rated it as knocking, 78% perfectly, while 90% classified at most one class ground truth. They thus considerably outperform broadly maximum amplitude oscillation (MAPO) method, well references reconstructed previous works. Our analysis indicates that learned physically meaningful features connected engine-characteristic resonances, verifying intended theory-guided data science approach. Deeper performance investigation shows remarkable generalization ability unseen operating points. addition, proved classify engines increased 89% after adapting their via training small algorithm takes below 1 ms individual cycles, effectively making suitable real-time control.