Time-Series Foundation Models Show Promise, Limitations for E-Nose Data

The emerging field of "time-series foundation models" (TSFMs) has shown great promise in generalizing across various time-series tasks and domains, from healthcare to manufacturing. However, their applicability to gas-sensing data, specifically from electronic noses (E-Noses), has remained largely unexplored. This research systematically evaluates the effectiveness of prominent TSFMs, including Chronos-2 and MOMENT, in processing E-Nose data. The study investigated whether the embeddings generated by these TSFMs provide robust representations for critical tasks such as gas identification and concentration prediction. The findings indicate that while TSFMs hold potential, achieving satisfactory performance on E-Nose data necessitates fine-tuning the models for the specific application. Furthermore, the research reveals that combining TSFM embeddings with representations learned by specialized predictive models can lead to even greater performance improvements. This suggests that while current TSFMs offer valuable foundational capabilities, they may still require domain-specific adaptation and integration with more tailored models to fully unlock their potential in complex gas-sensing applications.