MoE-Gyro: Self-Supervised Over-Range Reconstruction and Denoising for MEMS Gyroscopes

TL;DR

MoE-Gyro addresses the long-standing trade-off between measurement range and noise in MEMS gyroscopes by proposing a self-supervised Mixture-of-Experts framework with two specialized heads: Over-Range Reconstruction Expert (ORE) and Denoise Expert (DE). The architecture leverages a shared Masked Autoencoder backbone, a lightweight gating mechanism, Gaussian-Decay Attention, a physics-informed energy loss, dual-branch masking, and FFT-guided augmentation to jointly reconstruct clipped peaks and suppress noise without ground-truth labels. The authors also release ISEBench, a unified open-source benchmark for IMU signal enhancement, and demonstrate substantial enhancements in measurable range (from ±$450^ ext{o}/ ext{s}$ to ±$1500^ ext{o}/ ext{s}$) and significant reductions in Bias Instability (BI), while maintaining real-time applicability on edge hardware after pruning. The results establish a practical, scalable path to improve MEMS inertial sensing in diverse applications, from consumer electronics to robotics and autonomous systems. Overall, MoE-Gyro delivers a robust, self-supervised solution that extends usable range and suppresses noise more effectively than prior supervised or self-supervised baselines, with broad implications for downstream sensor fusion tasks.

Abstract

MEMS gyroscopes play a critical role in inertial navigation and motion control applications but typically suffer from a fundamental trade-off between measurement range and noise performance. Existing hardware-based solutions aimed at mitigating this issue introduce additional complexity, cost, and scalability challenges. Deep-learning methods primarily focus on noise reduction and typically require precisely aligned ground-truth signals, making them difficult to deploy in practical scenarios and leaving the fundamental trade-off unresolved. To address these challenges, we introduce Mixture of Experts for MEMS Gyroscopes (MoE-Gyro), a novel self-supervised framework specifically designed for simultaneous over-range signal reconstruction and noise suppression. MoE-Gyro employs two experts: an Over-Range Reconstruction Expert (ORE), featuring a Gaussian-Decay Attention mechanism for reconstructing saturated segments; and a Denoise Expert (DE), utilizing dual-branch complementary masking combined with FFT-guided augmentation for robust noise reduction. A lightweight gating module dynamically routes input segments to the appropriate expert. Furthermore, existing evaluation lack a comprehensive standard for assessing multi-dimensional signal enhancement. To bridge this gap, we introduce IMU Signal Enhancement Benchmark (ISEBench), an open-source benchmarking platform comprising the GyroPeak-100 dataset and a unified evaluation of IMU signal enhancement methods. We evaluate MoE-Gyro using our proposed ISEBench, demonstrating that our framework significantly extends the measurable range from 450 deg/s to 1500 deg/s, reduces Bias Instability by 98.4%, and achieves state-of-the-art performance, effectively addressing the long-standing trade-off in inertial sensing.

Figures (14)

• Figure 1: Pipeline of MoE‑Gyro framework. During inference, a low‑quality signal stream is segmented, routed by a gate to suitable expert, and the enhanced outputs are concatenated to form the final signal. During training, both experts are optimised in a fully self‑supervised manner on a shared MAE backbone, each equipped with task‑specific masking, attention, and loss mechanisms.
• Figure 2: Comparison of reconstruction P_MSE. We compare MoE-Gyro with two representative baselines,a drop of more than 75 % (dashed reference) marks high-quality recovery.
• Figure 2: Cross-Device Generalization results
• Figure 3: Allan‑variance comparison. The red curve corresponds to MoE-Gyro and shows the best Allan-variance performance—raising the device from consumer to nearly strategic grade.
• Figure 4: Range‑extension visualisation. At an actual angular rate of –1731.8°/s, our method reconstructs the signal clipped at the 450°/s sensor limit to –1453.7°/s.