Answers are grounded in this paper and cite their sources.
Chemoresistive gas sensors, and metal-oxide (MOX) sensors in particular, are among the most widely used gas sensing technologies, valued for their low cost, high sensitivity, and suitability for embedded and portable applications. Their long-term deployment, however, is strongly limited by measurement drift: a progressive, non-deterministic change in sensor response over time that degrades models calibrated on earlier data and ultimately forces costly recalibration. This paper reviews the principal strategies proposed to compensate for drift in chemoresistive sensor arrays. Because most of this literature, including its main benchmark, concerns MOX arrays, these receive particular emphasis, and evidence from other chemoresistive technologies, such as conducting polymers, is discussed where it informs the conclusions. Rather than cataloging the field exhaustively, a task addressed by a recent comprehensive survey, it takes a critical perspective, organizing compensation methods into three families, component correction, domain adaptation, and continuous learning, and examining the assumptions on which each rests. Fundamental concepts, drift types, and physical causes are first clarified to reduce terminological ambiguity across disciplines. The reviewed methods are then compared through a set of cross-cutting distinctions that are usually left implicit: whether compensation operates online or offline, at the level of the individual sensor or the array as a whole, and at what cost in complexity relative to the resource-constrained platforms on which these sensors are deployed. Particular attention is given to the near-universal default toward multivariate, array-level correction, which is argued to be a design choice rather than a physical necessity, since drift comprises both a shared, environmentally driven component and a heterogeneous, per-sensor one. The evaluation practices and benchmark datasets that shape the field are also examined critically, including their reliance on classification accuracy as a proxy and on a single dominant dataset. The review closes with practical guidance on selecting compensation strategies according to drift behavior, data availability, and computational constraints, and argues that the univariate and deployment-oriented design space remains comparatively underexplored.
No takes yet. Share an insight, caveat, or question.
Miranda et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: