Reconstructing In-Cylinder Pressure from Head Vibrations with Signal-to-Signal Deep Learning Architectures

Considering that piston internal combustion engines will remain essential converters of chemical energy into mechanical energy for an extended period, providing optimal diagnostic tools for their operation is imperative. Mechanical vibrations generated during machine operation constitute one of the most valuable sources of information about their technical condition. Their primary advantage lies in conveying diagnostic data with minimal time delay. This article presents a novel approach to vibroacoustic diagnostics of the combustion process in internal combustion piston engines. It leverages vibration signals carrying information about the pressure in the engine cylinder during fuel–air mixture combustion. In the proposed method, cylinder pressure information is reconstructed from vibration signals recorded on the cylinder head of the internal combustion engine. This method of signal-to-signal processing uses deep artificial neural network (ANN) models for signal reconstruction, providing an extensive exploration of the abilities of the presented models in the reconstruction of the pressure measurements. Furthermore, a novel two-network model, utilizing a U-net architecture with a dedicated smoothing network (SmN), allows for producing signals with minimal noise and outperforms other commonly used signal-to-signal architectures explored in this paper. To test the proposed methods, the study was limited to a single-cylinder engine, which presents certain constraints. However, this initial approach may serve as an inspiration for researchers to extend its application to multi-cylinder engines.