Reference-Free Multi-Species Gas Detection via Unsupervised Learning

Conventional chemometric models rely heavily on reference spectra for accurate species quantification. However, the availability of reference spectra is not always guaranteed, deeming these models unreliable and impractical in many scenarios. To address this limitation, we introduce an AutoencoderBased Blind Source Separation (AE-BSS) model for multi-species detection. The AE-BSS model can infer both species concentrations and their reference spectra based solely on composite mixture spectra, eliminating the need for individual reference spectra. Initially tested with simulated mixtures of methane, ethylene, ethane, propane, benzene, toluene, o-xylene, and isoprene, the AE-BSS model demonstrated accurate quantification without requiring their reference spectra. To validate these findings experimentally, we employed a setup using a single infrared laser to train the AE-BSS model. The results showed that the model accurately predicted both the reference spectra and species concentrations. Our findings confirm the effectiveness and robustness of the AE-BSS in real-world scenarios, significantly enhancing gas sensing applications. Future work will focus on extending the application of the model to more complex environments, including high-pressure and high-temperature conditions.