IMU User Transparent Tiny Neural Self-Calibration

This paper introduces a methodology to minimize measurement drift in inertial Micro-Electro-Mechanical Systems due to aging and stress. The approach initially stored high-precision sensor responses to specific stimuli, which served as ground truth. During any re-calibration process, responses to the same stimuli were compared to the ground truth, revealing time-varying errors. Tiny neural networks models were trained to approximate these errors, to achieve compensation of uncalibrated measurements. A 1D Convolutional - Dense model achieved the average mean compensation of 1.5g with only 1331 parameters. Deployment was feasible on the intelligent signal processing inertial sensor, with an execution time of 828.1ms.