In-cylinder pressure reconstruction from engine block vibrations via a branched convolutional neural network

We introduce a novel approach to reconstructing the in-cylinder pressure trace from vibration signals recorded with common knock sensors. The proposed methodology is purely data-driven and employs a convolutional neural network that has two distinct branches. Each branch is allowed to learn individual aspects of the mapping process, with boundary conditions within the model architecture set to incentivize the individual branches to learn low-frequency and high-frequency contents of the pressure trace. The reconstruction achieves calculated Pearson coefficients and coefficients of determination above 0.99 for all investigated datasets and a Mean Absolute Error of under 2.7 bar across all processed cycles. Furthermore, peak firing pressure and peak pressure position were extracted from the reconstructed cycles. Hereby, the method achieves Mean Absolute Error values of under 4.3 bar for peak firing pressure and under 1°crank angle for peak pressure position across all processed datasets, despite them not explicitly being targets of the underlying task. Deeper investigation of the results shows that combustion anomalies such as knocking do not negatively influence model fit. Moreover, model limitations were identified for high-pressure cycles and cycles exemplifying rather slow combustion.