HEROS-GAN: Honed-Energy Regularized and Optimal Supervised GAN for Enhancing Accuracy and Range of Low-Cost Accelerometers

Yifeng Wang, Yi Zhao

Introduction

Accelerometers are fundamental components in modern technology, integrated into everything from consumer smartphones to sophisticated industrial robotics. However, a significant performance-to-price gap persists in the market. While consumer-grade sensors are inexpensive (approximately $0.50), they suffer from high noise levels and "clipping"—a phenomenon where the sensor saturates and flatlines if acceleration exceeds a narrow range, typically ±8g. In contrast, professional-grade sensors capable of handling ±16g with high fidelity can cost hundreds or thousands of dollars.

Bridging this gap through software-based enhancement presents a significant challenge: data pairing. Training standard deep learning models requires frame-by-frame synchronized data from both low-cost and high-end sensors. In real-world scenarios involving vigorous movement or impacts, achieving millisecond-level synchronization is nearly impossible. This paper introduces HEROS-GAN, a novel framework designed to "upcycle" low-cost accelerometer data into high-fidelity signals using unpaired data.

Methodology

HEROS-GAN is built upon a CycleGAN backbone, specifically modified to handle the unique characteristics of inertial signals. The architecture introduces two primary innovations to overcome the limitations of unpaired training data and signal clipping.

1. Optimal Transport Supervision (OTS)

To address the lack of perfectly paired signal frames, the researchers utilized Optimal Transport theory. The OTS module employs the Sinkhorn algorithm to align the feature distributions between low-cost and high-cost sensors. Rather than attempting to match specific data points in time, OTS finds the most mathematically efficient mapping between the feature spaces of the two sensor types. This allows the model to maximize supervisory information and learn the underlying transformation from noisy, clipped data to clean, high-range signals without requiring rigid synchronization.

2. Modulated Laplace Energy (MLE)

The "Honed-Energy" component of the model refers to the Modulated Laplace Energy regularization. This technique uses the Laplacian operator—which measures second-order derivatives—to evaluate the "volatility" or high-frequency detail of a signal.

To prevent the model from injecting arbitrary noise, the MLE is guided by Kurtosis, a statistical measure of the "tailedness" of a distribution. When the model detects a clipped peak (where the signal is unnaturally smooth), the MLE pushes the generator to inject the necessary "energy" to reconstruct the missing physical details. This adaptive approach ensures that the reconstructed signal remains physically plausible.

The LASED Dataset

To validate HEROS-GAN, the authors developed the Low-cost Accelerometer Signal Enhancement Dataset (LASED). The dataset was collected using:

  • Hardware: Ten different smartphone models (including Honor, Huawei, Oppo, and Xiaomi) to test cross-device generalization.
  • Ground Truth: A high-precision Nokov Mars2H optical motion capture system consisting of eight cameras.
  • Protocol: The training was performed on data from a single smartphone type, while the remaining nine types were reserved for testing. This rigorous protocol ensures that the model learns the fundamental physics of sensor degradation rather than over-fitting to a specific hardware chip.

Results and Discussion

HEROS-GAN demonstrates substantial performance gains, often outperforming previous state-of-the-art (SOTA) methods by an order of magnitude across several key metrics:

  • Range Extension: The model successfully reconstructed signals from sensors with a ±8g hardware limit to match a ±16g reference, effectively doubling the sensor's functional range.
  • Noise Reduction: Signal noise, including quantization noise and bias instability, was reduced by two orders of magnitude compared to the raw sensor output.
  • Accuracy in Reconstruction: In the Clipped Signal Reconstruction Error (CSRE) metric at 6g, HEROS-GAN achieved an error of only 0.329g. For comparison, traditional model-driven and fully supervised deep learning methods typically yield errors between 1.5g and 1.8g.
  • Physical Consistency: The Zero-Velocity Residual Error (ZVRE), which measures the "drift" accumulated after a movement, was reduced by 94%. This improvement is critical for applications in dead-reckoning and navigation where cumulative error is a primary concern.

Conclusion and Future Work

HEROS-GAN proves that sophisticated generative modeling can compensate for hardware limitations, turning $0.50 sensors into high-performance instruments. By prioritizing distribution alignment over point-to-point alignment and using energy-based regularization, the framework overcomes the traditional barriers of unpaired data and signal saturation.

Future research will focus on reducing the computational overhead of the GAN architecture to allow for real-time implementation on mobile digital signal processors (DSPs). Additionally, the authors aim to explore sensor fusion—integrating gyroscope data to provide a holistic enhancement of the entire Inertial Measurement Unit (IMU)—and expanding the dataset to include the extreme dynamics found in automotive crash testing.

Summary updated

BibTeX
@inproceedings{wang2025heros,
  author = {Yifeng Wang and Yi Zhao},
  title = {{HEROS-GAN: Honed-Energy Regularized and Optimal Supervised GAN for Enhancing Accuracy and Range of Low-Cost Accelerometers}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=42bKDBP8F5},
}
Figure