Wavelet Dynamic Selection Network for Inertial Sensor Signal Enhancement

Yifeng Wang, Yi Zhao

Abstract

Inertial sensors, such as accelerometers and gyroscopes, are fundamental to modern navigation and motion tracking but are plagued by noise and drift. Traditional denoising methods, like wavelet transforms, often require manual selection of basis functions, while deep learning approaches can lack interpretability. This paper introduces the Wavelet Dynamic Selection Network (WDSNet), a framework that intelligently selects optimal wavelet bases through a data-driven category representation mechanism. By leveraging weakly-supervised guidance from downstream tasks, WDSNet achieves state-of-the-art signal enhancement, significantly reducing noise and improving trajectory reconstruction accuracy across diverse hardware platforms.

Introduction

Inertial sensors are ubiquitous in devices ranging from smartphones to drones. However, they are notoriously susceptible to noise—including quantization errors, bias instability, and thermal noise—which accumulates over time, leading to significant "drift" in applications like indoor navigation.

Wavelet transforms are a standard tool for signal processing due to their multi-resolution analysis capabilities. However, their effectiveness depends heavily on the choice of a "basis function." Traditionally, this selection is manual or empirical, and picking an inappropriate basis can degrade signal quality. While deep learning denoisers offer a potential solution, they often struggle to generalize or can inadvertently destroy the semantic meaning of the motion data. WDSNet addresses these challenges by automating wavelet selection within a robust neural framework.

Methodology

The WDSNet architecture treats wavelet selection as an intelligent, data-driven process, integrating classical signal processing with modern deep learning.

System Architecture

The framework utilizes a 1D-ResNet to extract features from raw inertial signals. These features are fed into a classifier that selects the most appropriate wavelet basis from a library of 16 candidates. The selected wavelet then performs threshold denoising on the original signal. This approach ensures that the output remains grounded in established signal processing theory while benefiting from the adaptive power of neural networks.

The WDSNet architecture integrates feature extraction, dynamic wavelet selection, and task-based guidance to enhance inertial signals.

Category Representation and Feature Supervision

To improve the network's internal learning, the authors implemented two novel mechanisms:

  1. Category Representation Mechanism (CRM): Standard networks often treat classification labels as abstract indices. CRM forces the network's internal weights to represent the actual features of the wavelet bases. It employs $\alpha$-order Renyi information entropy to ensure these representations are both informative and distinct (orthogonal).
  2. Feature Supervision Mechanism (FSM): FSM uses the CRM to provide direct supervision to the early layers of the network. By aligning data features with category features, the model mitigates the "gradient vanishing" problem and ensures that the feature extractor learns representations that are truly relevant to the denoising task.
The Feature Supervision Mechanism (FSM) aligns data features with category features derived from the CRM to improve training stability and feature relevance.

Weakly-Supervised Guidance

A critical innovation of WDSNet is its reliance on "guidance tasks" rather than perfectly clean ground-truth signals, which are often impossible to obtain in real-world scenarios. The network is trained to optimize for downstream performance in attitude (orientation) and displacement prediction. If a wavelet choice results in a more accurate trajectory, the network is rewarded, allowing it to learn a denoising strategy based on the utility of the signal.

Experimental Setup

The researchers curated a diverse dataset using 15 different smartphones from manufacturers including HUAWEI, Apple, Samsung, Xiaomi, and VIVO. These devices featured a variety of low-cost IMUs (e.g., LSM6DSM, ICM20690). To establish high-precision ground truth for the guidance tasks, phones were mounted on a ROKAE xMate ER3 Pro industrial robotic arm. The model was trained on data from a single device (HUAWEI Mate30 Pro) and tested on the remaining 14 to evaluate its generalization capabilities across different hardware.

Key Results

WDSNet demonstrated state-of-the-art (SOTA) performance across multiple metrics:

  • Noise Reduction: Allan variance analysis showed a 95% reduction in acceleration quantization noise compared to raw signals, outperforming both Kalman filters and standard CNNs.
  • Task Accuracy: The model reduced attitude estimation error by 64.35% and position estimation error by 73.62%.
  • Trajectory Reconstruction: WDSNet enabled the clear reconstruction of handwritten characters (0-9) from smartphone sensor data—a task that resulted in unreadable noise when using raw data.

Conclusion

WDSNet represents a significant step forward in inertial sensor enhancement by combining the mathematical rigor of wavelet transforms with the adaptive learning of deep ResNets. By focusing on the utility of the signal for downstream tasks and employing novel supervision mechanisms, the framework provides a robust, interpretable, and highly effective solution for denoising low-cost sensor data. Future work will explore expanding the wavelet library and optimizing the architecture for real-time deployment on low-power microcontrollers.

Summary updated

BibTeX
@inproceedings{wang2024wavelet,
  author = {Yifeng Wang and Yi Zhao},
  title = {{Wavelet Dynamic Selection Network for Inertial Sensor Signal Enhancement}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i14.29496},
}
Figure