Conventional emission frameworks such as the U.S. EPA’s MOVES estimate on-road emissions using operating-mode (OpMode) bins defined by vehicle speed, acceleration, and vehicle-specific power (VSP). However, real-world telematics datasets rarely include engine speed or gear information, even though drivetrain operation strongly governs fuel consumption and CO 2 output. We propose a MOVES-compatible probabilistic framework that infers discrete gear states from GPS-based vehicle speed using a Markov-switching dynamic regression model. The inferred posteriors P ( s t = j ) act as latent variables that (i) stabilize OpMode assignment through gear-consistent filtering, (ii) correct drivetrain-induced bias in emission factors, and (iii) propagate drivetrain uncertainty into emission estimates. Across six gasoline vehicles (Class I–II, 5–10 gears) equipped with portable emission measurement systems (PEMS), conditioning on inferred gear states reduced within-OpMode CO 2 variance, lowered second-by-second MAE by 2–11%, and improved R 2 by 5–15% relative to standard MOVES binning. Requiring only GPS speed and fewer than 50 parameters per vehicle, the method remains fully interoperable with MOVES while substantially enhancing its physical interpretability and uncertainty awareness. This study demonstrates a scalable pathway toward drivetrain-informed, uncertainty-aware microscale CO 2 inventories. Code is available at: https://github.com/Chogaliu/probabilistic-inference-of-gear-and-engine-dynamics-from-GPS-speed .
Liu et al. (Thu,) studied this question.