Stochastic Modeling for NDIR Gas Sensor Drift
Definition, Decomposition, Dynamics and Calibration
Time: Fri 2026-11-06 14.00
Location: F3 (Flodis), Lindstedtsvägen 26 & 28, Stockholm
Video link: https://kth-se.zoom.us/j/62661699362
Language: English
Subject area: Electrical Engineering
Doctoral student: Cheng Yang , Teknisk informationsvetenskap
Opponent: Associate Professor Raj Thilak Rajan, Delft University of Technology, Delft, The Netherlands
Supervisor: Professor Tobias J. Oechtering, Teknisk informationsvetenskap; Professor Saikat Chatterjee, Teknisk informationsvetenskap
QC 20261010
Abstract
Sensor drift is a major obstacle to reliable long-term measurements of low-cost sensors. It is commonly treated as a calibration problem, although its definition, causes, and temporal behavior are often left implicit. Focusing on non-dispersive infrared (NDIR) CO2 sensors, we develop a systematic approach to defining, understanding, characterizing, and addressing drift.
We first describe sensor measurements and measurement errors as stochastic processes. Within this framework, drift occurs when the statistical properties of the measurement error process vary with time. We then introduce an additive drift decomposition model (ADDM) to separate two principal sources of drift: reversible effects caused by environmental conditions and irreversible intrinsic changes in the sensing instrument. Experiments using long-term field data demonstrate the feasibility of estimating the individual components and support the additive formulation.
After environmental correction, most sensors exhibit mean-reverting instrumental errors rather than a common deterministic trend. We model these dynamics using an Ornstein–Uhlenbeck process, whose parameters characterize the sensor-specific equilibrium error, mean-reversion rate, and stochastic variation. These characteristics provide theoretical guidance for designing sensor-specific one-point calibration strategies.
Finally, we address calibration without a co-located reference instrument by extracting pseudo-reference information from three external sources: same-type sensors, distributed reference stations, and local sensor networks. First, nonlinear environmental dependencies estimated from tested sensors are transferred to other sensors of the same type for environmental compensation. Second, periodic minima are interpreted as order-statistic estimators of atmospheric background concentrations, and observations from distributed reference stations are fused using a spatio-temporal generalized additive model to construct dynamic baselines for one-point calibration. Third, normalized pairwise differences are used to identify rendezvous events within a local sensor network; adaptive piecewise constant approximation and a two-state explicit-duration hidden semi-Markov model are proposed to detect rendezvous suitable for the downstream network calibration. These methods are evaluated using long-term field data and dedicated indoor deployment experiments.