ML-based Thermal Sensor Calibration by Bivariate Gaussian Mixture Model Estimation

This paper presents a machine-learning-based post signal processing to calibrate thermal sensors. The proposed calibration scheme is shown to be immune to the interference from the environment and fulfills the high-resolution requirements of human body temperature measurements. The sensing module comprises two resistive sensing circuits, one is for sensing the external temperature, and the other is for sensing the internal die temperature. By using these two thermal outputs, we trained two-dimensional multivariate Gaussian models for several temperature intervals. Higher accuracy can be obtained via the probability-based estimation. The simulation results show high accuracy even in a noisy environment. The proposed algorithm is implemented and fabricated in UMC 0.18m CMOS-MEMS technology. The sensor chip is tested by an embedded system (ARM V2M-MPS2). The measurement results show that the proposed method can effectively improve the accuracy from 1 degree Celsius to 0.1 degree Celsius.