A Large-Scale Empirical Study of Deep Learning for MEMS Gyroscope Bias Estimation

Abstract

MEMS gyroscope bias drift is a dominant error source in sensor calibration and inertial navigation. No prior work directly compares deep learning architectures for gyroscope bias estimation on common benchmarks. We report 2,811 experiments comparing MLP, LSTM, TCN, and Transformer models on two public visual-inertial datasets (EuRoC, TUM-VI) with leave-one-sequence-out cross-validation and 17 random seeds. Variance decomposition reveals that dataset choice explains approximately 97% of performance variation; model architecture accounts for less than 1%. On clean-label EuRoC, TCN achieves the highest attitude-drift improvement (96.19%, Friedman p<0.001); on noisy-label TUM-VI, LSTM reaches 21.57%, but only one of six pairwise comparisons is significant. Physics-informed SO(3) regularization fails to improve performance across 390 confirmatory experiments (N=17). Removing a common label-denoising step changes EuRoC results by +1.2-3.5 pp and reverses the TUM-VI ranking. A within-dataset label corruption experiment confirms that label quality drives performance far more than architecture selection (approximately 40:1). Label quality and evaluation protocol matter more than architecture choice here.