Deep neural networks for simultaneous BTEX sensing at high temperatures

In the study of chemical reactions, it is desirable to have a diagnostic strategy that can detect multiple species simultaneously with high sensitivity, selectivity, and fast time response. Laser-based selective detection of benzene, toluene, ethylbenzene, and xylenes (BTEX) has been challenging due to the similarly broad absorbance spectra of these species. Here, a mid-infrared laser sensor is presented for selective and simultaneous BTEX detection in high-temperature shock tube experiments using deep neural networks (DNN). A shock tube was coupled with a non-intrusive mid-infrared laser source, scanned over 3038.6–3039.8 cm⁻¹, and an off-axis cavity enhanced absorption spectroscopy (OA-CEAS) setup of ∼ 100 gain to enable trace detection. Absorption cross-sections of BTEX species were measured at temperatures of 1000–1250 K and pressures near 1 atm. A DNN model with five hidden layers of 256, 128, 64, 32, and 16 nodes was implemented to split the composite measured spectra into the contributing spectra of each species. Several BTEX mixtures with varying mole fractions (0–600 ppm) of each species were prepared manometrically and shock-heated to 1000–1250 K and 1 atm, and the composite measured absorbance were split into contributions from each BTEX species using the developed DNN model, and thus make selective determinations of BTEX species. Predicted and manometric mole fractions were in good agreement with an absolute relative error of ∼ 11%. We obtained a minimum detection limit of 0.73–1.38 ppm of the target species at 1180 K. To the best of our knowledge, this work reports the first successful implementation of multispecies detection with a single narrow wavelength-tuning laser in a shock tube with laser absorption spectroscopy.