TL;DR
HEROS-GAN is a generative adversarial framework that doubles the dynamic measurement range of cheap accelerometers (from 8g to 16g) while cutting sensor noise by nearly two orders of magnitude. Because perfectly synchronized high- and low-cost accelerometer data cannot be recorded simultaneously under violent motion, the model uses Optimal Transport and a novel Laplace energy regularizer to learn high-fidelity signal restoration from completely unpaired data.
What Problem Does This Paper Address?
Low-cost MEMS accelerometers (costing $0.20 to $0.50) are ubiquitous in consumer electronics, wearables, and IoT devices. However, they suffer from two major limitations:
- Severe noise, which prevents fine vibration monitoring and accurate motion tracking.
- Narrow measurement ranges (typically $\pm 2g$ or $\pm 8g$), which leads to sensor saturation and signal clipping during high-impact events like falls, automotive crashes, or industrial machinery operations.
While high-end industrial sensors ($\pm 16g$ and higher) resolve these issues, they cost tens to thousands of dollars. Upgrading low-cost signals via standard supervised deep learning is bottlenecked by data collection: it is practically impossible to collect frame-by-frame, perfectly aligned paired data between low- and high-cost sensors during high-speed, violent motion.
Methods
HEROS-GAN builds on CycleGAN to map unpaired low-cost sensor signals to high-cost equivalents, adding two core mathematical innovations:
- Optimal Transport Supervision (OTS): Instead of relying on naive point-to-point loss functions (like L1/L2), OTS applies optimal transport theory (solved via the Sinkhorn algorithm) to match the feature distributions of unpaired low- and high-cost signals. This allows the model to extract structural supervision from unpaired trajectories without enforcing unrealistic point-wise alignment.
- Modulated Laplace Energy (MLE): Standard GANs tend to over-smooth signals, failing to reconstruct the sharp peaks lost to sensor saturation. MLE measures the second-order derivative of latent feature representations using a discrete Laplacian operator ($\nabla^2$). It then modulates this energy using the feature kurtosis ($\kappa$):
- Low-kurtosis (smoother) features are injected with more energy to restore high-frequency signal details and overcome clipping barriers.
- High-kurtosis (volatile) features receive less energy to prevent the network from amplifying noise or creating unstable artifacts.
Datasets
- LASED (Low-cost Accelerometer Signal Enhancement Dataset): The authors introduced and open-sourced this benchmark. It consists of tens of thousands of samples collected across 10 different smartphones equipped with sub-$0.50 internal accelerometers, assisted by an 8-camera optical motion capture system (Nokov Mars2H).
- Train/Test Split: To rigorously evaluate real-world generalization across hardware, data from only one smartphone (HONOR Magic 4) was used for training, while the remaining nine phones (including devices from Huawei, Oppo, Realme, Xiaomi, Lenovo, Vivo, and Apple) were reserved solely for zero-shot testing.
Key Insights and Ideas
- Feature alignment over point matching: When sensors physically disagree or desynchronize under high-dynamic force, forcing strict element-wise loss destroys performance. Optimal transport discovers the underlying latent manifold similarities across independent trials.
- Controlled feature volatility breaks saturation limits: Rather than treating saturation as a purely missing-data interpolation task, injecting adaptively scaled Laplacian energy allows the model to expand compressed, flat-topped clipped signals into realistic dynamic peaks.
- Validation through physics: High signal-to-noise ratios mean little if reconstructed acceleration violates physical laws. The paper measures physical plausibility by integrating acceleration over a static-dynamic-static cycle; genuine movement must integrate back to zero velocity.
Key Results
HEROS-GAN consistently outperformed state-of-the-art model-driven and supervised deep-learning baselines across all metrics:
- Over-Range Recovery (CSRE at $\tau = 6g$): Reduced Clipped Signal Reconstruction Error from 1.981 (raw) to 0.329 (an 83.4% error reduction), substantially outperforming traditional polynomial fitting (1.779) and supervised GRU-LSTM models (1.742).
- Physical Consistency (ZVRE): Dropped the Zero-Velocity Residual Error from 350.78 m/s to 21.05 m/s (a 94.0% reduction), confirming that the generated dynamics obey real-world physical transitions.
- Static Signal Noise Reduction (Allan Variance):
- Quantization Noise (QN): reduced by 95.3% (1.205 $\to$ 0.057).
- Velocity Random Walk (VRW): reduced by 96.9% (1.917 $\to$ 0.060).
- Bias Instability (BI): reduced by 97.9% (3.119 $\to$ 0.065).
Limitations and Future Work
- Heuristic Restoration: Reconstructing clipped signals relies on learned data distributions and contextual priors rather than direct physical measurement; extreme, out-of-distribution motion shocks may still lead to inaccurate peak amplitudes.
- Cross-Domain Potential: The authors note that both OTS and the kurtosis-modulated Laplace regularizer are domain-agnostic and suggest applying them to multi-modal translation, medical imaging (e.g., MRI-to-CT alignment), and robotic state estimation under heterogeneous sensing.
Summary updated