Machine Learning-Enhanced Laser Absorption Spectroscopy for Harsh-Environment Combustion Diagnosis

Laser absorption spectroscopy (LAS) has been widely adopted as a diagnostic tool for reactive flow-field monitoring in industrial combustion applications. Despite various advancements in LAS signal processing schemes, these harsh environments inevitably impose noise and interference on the LAS measurement data, thus increasing inaccuracy and uncertainty in combustion analysis. This article proposes a machine learning (ML)-enhanced LAS methodology that significantly mitigates noise-induced distortions in absorption spectra, yielding a more accurate representation of the original spectral sequence through continuous measurements. The proposed method is a novel architecture that integrates a denoising autoencoder (DAE) with a long short-term memory (LSTM) network for enhanced LAS signal analysis. Developed entirely using in situ experimental data, this approach ensures strong portability for industrial combustion diagnostics, where multisource measurement noise is difficult to model or quantify. To validate the proposed method, we conducted a combustion experiment on an auxiliary power unit (APU), a full-scale commercial gas turbine aeroengine, focusing on exhaust temperature measurements using our advanced LAS technique. The experimental results demonstrate the efficacy of the proposed method in recovering high-fidelity absorption spectra from the noise-contaminated data, enabling more convenient, accurate, and stable APU exhaust temperature measurements with a standard deviation below $7.9~^{\circ }$ C. This indicates its significant potential in industrial combustion diagnostics, offering a reliable tool for precise analysis and assessment in harsh environments.