Machine learning based compensation system for high-precision temperature measurement

Abstract

To address the systematic errors introduced by the coupling of excitation current drift, reference voltage drift, and ADC gain/offset in PT1000 platinum resistance measurement systems, this paper proposes an error-structure-aware two-stage compensation framework. A three-dimensional physical feature vector is derived from a multiplicative-additive error decomposition model, capturing the dominant error channels in a physically interpretable manner. In the first stage, a gain correction polynomial removes the amplitude-dependent multiplicative error; in the second stage, a lightweight MLP refines the remaining nonlinear residuals. The model contains only 2,833 parameters, occupies 11 KB of Flash, and achieves an inference latency of 0.8 ms on the STM32F103 platform. Experiments conducted across 36 generalization test points show that the proposed method reduces MAE from 1702.79 m℃ to 13.43 m℃ (126.8×), outperforming polynomial (50.30 m℃) and SVR (17.92 m℃) baselines, with 80.6% of test points meeting the 20 m℃ target and a maximum error of 28.99 m℃—well within the 50 m℃ requirement for battery thermal characterization calorimetry.