Multi-component gas sensing via spectral feature engineering

Abstract

We present a straightforward yet powerful spectral feature engineering technique designed to improve multi-species detection in complex mixtures. By applying convolutions of first derivatives with the composite spectra of target species before feeding the data into a convolutional neural network (CNN) model, this method significantly enhances the detection of weak absorbers and overlapping spectral features. To validate the approach, we developed a laser-based sensor that integrates wavelength tuning with a 1-D CNN model. The system utilizes a distributed feedback inter-band cascade laser operating near $3.34\mu m$, enabling selective and simultaneous measurement of $C_1-C_3$ hydrocarbons. Experiments were conducted at ambient conditions with a temporal resolution of 10 ms, while (intentionally) keeping the signal-to-noise ratio at relatively low levels. Gaseous mixtures contained methane, ethane, propane and propyne ranging in mole fraction values of 0%–1%, and ethylene mole fraction below 200 ppm. Ethylene was deliberately kept at very low levels to demonstrate the effectiveness of the feature engineering technique in detecting a weak absorbing species. The proposed method reduced the mean squared error by ten times compared to standard models. This demonstrates its potential for accurate detection in challenging environments.