Wavelet Dynamic Selection Network for Inertial Sensor Signal Enhancement

Yifeng Wang, Yi Zhao

Introduction

Inertial Measurement Units (IMUs), comprising accelerometers and gyroscopes, are fundamental to robotics, wearable technology, and autonomous navigation. However, these sensors are inherently prone to noise—such as quantization errors and bias instability—which accumulates over time, causing significant "drift" in trajectory estimation. Traditional denoising methods, like Wavelet Transforms, require manual selection of basis functions, while modern deep learning approaches often lack interpretability and require extensive labeled datasets.

To address these challenges, researchers Yifeng Wang and Yi Zhao presented the Wavelet Dynamic Selection Network (WDSNet) at AAAI-24. WDSNet is a hybrid framework that intelligently and dynamically selects the optimal wavelet basis for real-time signal denoising using weak supervision from downstream tasks.

The WDSNet Architecture

WDSNet treats wavelet selection as a dynamic classification problem rather than a static engineering choice. The architecture integrates signal processing principles with deep learning to create an interpretable, high-performance denoiser.

Overview of the WDSNet architecture, showing the flow from raw inertial signals to enhanced signals guided by downstream tasks.

The framework follows a multi-stage process:

  1. Feature Extraction: A 1D-ResNet extracts deep features from the raw 6-axis IMU signal.
  2. Dynamic Selection: These features are passed to a classifier that selects the most appropriate wavelet basis from a library of 16 candidates (e.g., Haar, Daubechies).
  3. Wavelet Thresholding: The original signal is processed using the selected wavelet to produce an "enhanced" inertial signal.
  4. Weakly-Supervised Guidance: The model is trained by minimizing loss on downstream Guidance Tasks, specifically attitude (rotation) and displacement (position) prediction. This allows the network to learn optimal denoising strategies without needing ground-truth "clean" signals.

Innovative Mechanisms: CRM and FSM

A core contribution of WDSNet is the Category Representation Mechanism (CRM). Unlike standard networks that use arbitrary "one-hot" labels, CRM learns high-dimensional category vectors for each wavelet type. This forces the network to capture the physical characteristics of different wavelets within its weights.

Complementing this is the Feature Supervision Mechanism (FSM). FSM uses these learned category vectors to directly supervise the feature extractor. By aligning data features with category-specific properties early in the network, WDSNet alleviates gradient vanishing and ensures that internal representations are physically meaningful.

Experimental Evaluation and Results

The researchers evaluated WDSNet using a diverse dataset collected from 15 different smartphones. To establish high-precision ground truth for motion, the devices were mounted on a ROKAE xMate ER3 Pro mechanical arm. WDSNet demonstrated superior performance in both static and dynamic scenarios:

  • Static Performance: Significant reduction in quantization noise and bias instability compared to Kalman filters and standard CNNs.
  • Dynamic Accuracy: WDSNet improved attitude estimation by 64.35% and trajectory reconstruction by 86.95% over raw sensor data.
  • Visual Reconstruction: In handwriting tasks, where users "write" in the air, WDSNet successfully reconstructed legible digits from signals that were otherwise unintelligible.
Handwriting reconstruction of the digit '4'. The WDSNet-enhanced signal (black) closely matches the ground truth, whereas the raw signal is unreadable.

Conclusion

WDSNet represents a significant step forward in inertial sensing by combining the mathematical rigor of wavelet transforms with the adaptive power of deep learning. By leveraging weak supervision and novel representation mechanisms, it provides a robust, interpretable, and highly accurate solution for the persistent problem of IMU noise. Future work will look toward expanding the wavelet library and optimizing the framework for low-power embedded hardware.

Summary updated

Abstract

As attitude and motion sensing components, inertial sensors are widely used in various portable devices, covering consumer electronics, sports health, aerospace, etc. But the severe intrinsic errors of inertial sensors heavily restrain their function implementation, especially the advanced functionality, including motion trajectory recovery and motion semantic recognition, which attracts considerable attention. As a mainstream signal processing method, wavelet is hailed as the mathematical microscope of signal due to the plentiful and diverse wavelet basis functions. However, complicated noise types and application scenarios of inertial sensors make selecting wavelet basis perplexing. To this end, we propose a wavelet dynamic selection network (WDSNet), which intelligently selects the appropriate wavelet basis for variable inertial signals. In addition, existing deep learning architectures excel at extracting features from input data but neglect to learn the characteristics of target categories, which is essential to enhance the category awareness capability, thereby improving the selection of wavelet basis. Therefore, we propose a category representation mechanism (CRM), which enables the network to extract and represent category features without increasing trainable parameters. Furthermore, CRM transforms the common fully connected network into category representations, which provide closer supervision to the feature extractor than the far and trivial one-hot classification labels. We call this process of imposing interpretability on a network and using it to supervise the feature extractor the feature supervision mechanism, and its effectiveness is demonstrated experimentally and theoretically in this paper. The enhanced inertial signal can perform impracticable tasks with regard to the original signal, such as trajectory reconstruction. Both quantitative and visual results show that WDSNet outperforms the existing methods. Remarkably, WDSNet, as a weakly-supervised method, achieves the state-of-the-art performance of all the compared fully-supervised methods.
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