Region-aware neural network modeling for high-fidelity interpolative reconstruction of engine in-cylinder pressure traces

Abstract

Accurate reconstruction of in-cylinder pressure traces at a held-out intermediate operating condition is important for combustion diagnostics and calibration-oriented analysis. Beyond scalar quantities such as peak pressure or IMEP, reconstruction of complete crank-angle-resolved pressure (P–CA) traces provides more detailed information on combustion evolution and enables the derivation of additional combustion-sensitive quantities. This work presents a region-aware neural network framework for high-fidelity reconstruction of full-cycle in-cylinder pressures. The model uses crank angle together with a scalar operating-condition variable as inputs and predicts pointwise in-cylinder pressure as the output, enabling assembly of continuous P–CA traces. To enhance local fidelity in combustion-sensitive regions, two alternative region-focused strategies are investigated: region-specific training, which segments the crank-angle domain into physically meaningful intervals based on pressure evolution and apparent heat-release characteristics, and combustion-zone-focused data augmentation, which increases representation of high-gradient pressure regions. The results show that all three methods accurately reconstruct the full-cycle pressure trace at the held-out intermediate condition, with R² > 0.99 and full-cycle RMSE values below 0.30 bar. For the reported model realizations, the region-aware strategies show lower errors for several combustion-sensitive local features; for example, the error in maximum pressure-rise rate at the held-out condition decreases from more than 30% for the baseline ANN to below 4% for the region-aware approaches. Comparable accuracy is retained for global and integral quantities such as peak cylinder pressure and IMEP. Overall, these findings demonstrate the potential of the proposed region-aware framework for high-fidelity interpolative pressure reconstruction and combustion-sensitive analysis within the investigated operating domain.