Abstract
Spectroscopic methods are advantageous for gas detection with applications ranging from safety to operational efficiency. Despite the potential of laser-based sensors, real-world challenges, such as noise, interference and unseen conditions, hinder the accurate identification of species. The use of conventional machine learning (ML) models is constrained by extensive data requirements and their limited adaptability to new conditions. Although augmentation-based strategies have proven to improve the robustness of machine learning models, they still do not offer a complete defense. To address these challenges, this study focuses on three primary goals: first, to detect pressure-induced spectral broadening using simple yet effective augmentations; second, to bypass the need for extensive data sets by deploying a one-shot learning approach that can identify up to 12 volatile organic compounds (VOCs); and third, to provide a provable certification for the one-shot learning model predictions via randomized smoothing. To assess the effectiveness of our proposed augmentations and randomized smoothing, we perform a comparative study with four distinct models: VOC-net, VOC-lite, VOC-plus, and VOC-certifire. Remarkably, the one-shot learning model proposed herein, VOC-certifire, delivers predictions that match the baseline model VOC-net. The VOC-certifire predictions not only exhibit robustness and reliability but are also certified within a predefined norm radius. Such a certification is particularly useful for gas detection, where the robustness, precision and consistency are key to well-informed decision-making.