Scale and Direction Guided GAN for Inertial Sensor Signal Enhancement

Yifeng Wang, Yi Zhao

Abstract

Inertial sensors, serving as attitude and motion sensing components, are extensively used in various portable devices spanning consumer electronics, sports health, aerospace, etc. However, the severe intrinsic errors of inertial sensors greatly restrict their capability to implement advanced functions, such as motion tracking and semantic recognition. Although generative models hold significant potential for signal enhancement, unsupervised or weakly-supervised generative methods may not achieve ideal generation results due to the absence of guidance from paired data. To address this, we propose a scale and direction-guided generative adversarial network (SDG-GAN), which provides dual guidance mechanisms for GAN with unpaired data across two practical application scenarios. In the unsupervised scenario where only unpaired signals of varying quality are available, our scale-guided GAN (SG-GAN) forces the generator to learn high-quality signal characteristics at different scales simultaneously via the proposed self-supervised zoom constraint, thereby facilitating multi-scale interactive learning. In the weakly-supervised scenario, where additional experimental equipment can provide some motion information, our direction-guided GAN (DG-GAN) introduces auxiliary tasks to supervise signal generation while avoiding interference from auxiliary tasks on the main generation task. Extensive experiments demonstrate that both the unsupervised SG-GAN and the weakly-supervised DG-GAN significantly outperform all comparison methods, including fully-supervised approaches. The combined SDG-GAN achieves remarkable results, enabling unimaginable tasks based on the original inertial signal, such as 3D motion tracking.

Introduction

Inertial Measurement Units (IMUs), comprising accelerometers and gyroscopes, are ubiquitous in modern mobile devices. However, consumer-grade sensors suffer from significant noise, drift, and bias, limiting their utility in high-precision applications like 3D motion tracking. Traditional deep learning denoising methods typically require paired data—simultaneous recordings from low-quality and high-precision sensors—which is prohibitively difficult to collect at scale. While unsupervised GANs offer a solution, they often fail to capture fine-grained signal details or suffer from task interference when auxiliary supervision is added.

Methodology: The SDG-GAN Framework

The proposed Scale and Direction Guided GAN (SDG-GAN) addresses these challenges through two specialized guidance mechanisms designed for unpaired data scenarios.

Scale-Guided GAN (SG-GAN)

In fully unsupervised settings, SG-GAN utilizes a self-supervised zoom constraint to enhance signal quality. As shown in Figure 1, the architecture processes signals across three scales: 1.0x (intrinsic), 1.2x, and 2.0x zoom views. By enforcing consistency between high-resolution generated signals and their downsampled counterparts, the model learns to preserve both global structural integrity and local high-frequency details, preventing common GAN artifacts like blurring.

The SG-GAN architecture utilizes multi-scale zoom views and cycle consistency to guide unsupervised signal enhancement.

Direction-Guided GAN (DG-GAN)

For weakly-supervised scenarios where some motion information is available, DG-GAN incorporates an Auxiliary Task Library including Attitude Estimation (AE), Displacement Prediction (DP), and Semantic Recognition (SR). To prevent these tasks from diverging from the primary denoising goal, a Learning Direction Guidance mechanism monitors feature consistency. If an auxiliary task's gradient conflicts with the main generation task, its influence is dynamically suppressed, as illustrated in Figure 2.

The DG-GAN architecture incorporates auxiliary tasks with a feature consistency constraint to provide direction guidance without task interference.

Experimental Results

The framework was evaluated using a diverse dataset collected from 15 different smartphone models. Ground truth for motion capture was provided by a ROKAE xMate ER3 Pro robotic arm and a Nokov Mars2H optical system.

Quantitative Performance

  • Noise Reduction: Allan variance analysis revealed that SDG-GAN reduces noise indicators, such as Angle Random Walk and Bias Instability, by 95% to 98% compared to raw signals.
  • Trajectory Reconstruction: In 3D motion tracking tasks—the "holy grail" of IMU processing—the model achieved a 93.4% reduction in Final System State Error (FSSE), enabling precise trajectory reconstruction from a single noisy smartphone sensor.
  • Generalization: Despite being trained on a single device model (HUAWEI Mate30 Pro), the system demonstrated robust performance across 14 unseen smartphone brands, highlighting its ability to handle diverse hardware characteristics.

Conclusion

SDG-GAN represents a significant advancement in inertial sensor signal processing. By leveraging multi-scale zoom constraints and dynamic direction guidance, the model overcomes the limitations of unpaired data in both unsupervised and weakly-supervised scenarios. Future work will focus on optimizing the framework for real-time mobile deployment, potentially extending its application to audio enhancement and medical signal processing.

Summary updated

Abstract

Inertial sensors, serving as attitude and motion sensing components, are extensively used in various portable devices spanning consumer electronics, sports health, aerospace, etc. However, the severe intrinsic errors of inertial sensors greatly restrict their capability to implement advanced functions, such as motion tracking and semantic recognition. Although generative models hold significant potential for signal enhancement, unsupervised or weakly-supervised generative methods may not achieve ideal generation results due to the absence of guidance from paired data. To address this, we propose a scale and direction-guided generative adversarial network (SDG-GAN), which provides dual guidance mechanisms for GAN with unpaired data across two practical application scenarios. In the unsupervised scenario where only unpaired signals of varying quality are available, our scale-guided GAN (SG-GAN) forces the generator to learn high-quality signal characteristics at different scales simultaneously via the proposed self-supervised zoom constraint, thereby facilitating multi-scale interactive learning. In the weakly-supervised scenario, where additional experimental equipment can provide some motion information, our direction-guided GAN (DG-GAN) introduces auxiliary tasks to supervise signal generation while avoiding interference from auxiliary tasks on the main generation task. Extensive experiments demonstrate that both the unsupervised SG-GAN and the weakly-supervised DG-GAN significantly outperform all comparison methods, including fully-supervised approaches. The combined SDG-GAN achieves remarkable results, enabling unimaginable tasks based on the original inertial signal, such as 3D motion tracking.
BibTeX
@inproceedings{wang2024scale,
  author = {Yifeng Wang and Yi Zhao},
  title = {{Scale and Direction Guided GAN for Inertial Sensor Signal Enhancement}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=mJJF7Ea1NZ},
}

Loading related content...

Figure