Research map / Label-efficient and robust learning / Self-supervised and contrastive
MoE-Gyro: Self-Supervised Over-Range Reconstruction and Denoising for MEMS Gyroscopes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MoE-Gyro, a self-supervised mixture-of-experts over MAE backbones that reconstructs clipped MEMS gyroscope peaks and denoises signals, and introduces the ISEBench benchmark with the GyroPeak-100 dataset.
Problem
MEMS gyroscopes face a trade-off between measurement range and noise performance; existing deep learning enhancement needs time-aligned ground truth and does not jointly handle over-range reconstruction and denoising.
Approach
A gated two-expert MAE design: an over-range expert with learnable Gaussian-decay attention, correlation and physics-informed energy losses, and a denoise expert with dual-branch complementary cross-masking and FFT-guided noise-injection augmentation, all trained without aligned ground truth.
Key results
- Extends measurable range from ±450°/s to ±1500°/s; on ISEBench reduces P_MSE to 0.325 and PSNR rises to 8.29 dB versus 7.7 for HEROS-GAN.
- Reduces Bias Instability by 98.4% and QN by 98.0% relative to raw signal, the best among compared methods.
- Achieves best average rank (1.0) across ISEBench metrics; SNR 24.19 dB vs 19.23 for next best (LSTM_GRU).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MoE-Gyro instead of HEROS_GAN | Peak reconstruction · ISEBench, τ = 450°/s | PSNR: 7.7 -> 8.29 | Helped |
| use MoE-Gyro instead of HEROS_GAN | Peak reconstruction · ISEBench, τ = 450°/s | P_MSE: 0.354 -> 0.325 | Helped |
| use MoE-Gyro instead of HEROS_GAN | Peak reconstruction · ISEBench, τ = 450°/s | Corr: 0.89 -> 0.92 | Helped |
| use MoE-Gyro instead of LSTM_GRU | Weak-signal denoising · ISEBench, weak signal | SNR: 19.23 -> 24.19 | Helped |
| use MoE-Gyro instead of EMD | Denoising · ISEBench, Allan variance | BI reduction: -96.8% -> -98.4% | Helped |
| add PINN (ablation) | MoE-Gyro over-range reconstruction · Running Swing | PSNR: 8.48 -> 8.33 | Hurt |
| add PINN (ablation) | MoE-Gyro over-range reconstruction · Jumping | PSNR: 7.39 -> 7.70 | Helped |
| add PINN (ablation) | MoE-Gyro over-range reconstruction · Wrist Twist | PSNR: 8.01 -> 8.18 | Helped |
| replace the smoothness regularizer with PINN (ablation) | ORE · Unseen over-range data | PSNR: +0.27 dB | Helped |
| replace the smoothness regularizer with PINN (ablation) | ORE · Unseen over-range data | P_MSE: 8% | Helped |
| use decoder-only GD-Attn placement (ablation) | ORE · Over-range reconstruction | PSNR | Helped |
| use a 50% complementary mask ratio (ablation) | Denoise Expert · Denoising | Allan metrics | Helped |
Surprising findings
- Pruned 1.85M-parameter model shows only minor performance drop, and a learned MLP gate improved SNR but reduced PSNR relative to the heuristic gate.
- PINN loss slightly lowered PSNR on the Running Swing task while improving stability on other tasks.
Limitations
- Relatively large architecture may challenge resource-constrained embedded deployment.
- Denoising training targets still contain residual baseline noise, capping performance.
- Rule-based gate is heuristic; learned MLP gate was overly conservative in some cases.
Experimental setup
Datasets: GyroPeak-100 Visual-Inertial dataset Autonomous Platform Inertial dataset
Benchmarks: ISEBench
Metrics: PSNR P_MSE Corr SNR QN ARW BI average rank
Baselines: Matlab 2023 extrapolation EMD Savitzky-Golay filter CNN LSTM-GRU KNN HEROS-GAN IMUDB
Builds on: masked autoencoder LIMU-BERT IMUDB HEROS-GAN
Closest papers by meaning
- EqNIO: Subequivariant Neural Inertial Odometry (ICLR 2025 · similarity 0.83)
- Gyro-based Neural Single Image Deblurring (CVPR 2025 · similarity 0.83)
- iMoT: Inertial Motion Transformer for Inertial Navigation (AAAI 2025 · similarity 0.80)
- HEROS-GAN: Honed-Energy Regularized and Optimal Supervised GAN for Enhancing Accuracy and Range of Low-Cost Accelerometers (AAAI 2025 · similarity 0.79)
- Normalization Layer Per-Example Gradients are Sufficient to Predict Gradient Noise Scale in Transformers (NeurIPS 2024 · similarity 0.79)
- Accurate and Steady Inertial Pose Estimation through Sequence Structure Learning and Modulation (NeurIPS 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.