Research map / Label-efficient and robust learning / Recognition and long-tail
Wavelet Dynamic Selection Network for Inertial Sensor Signal Enhancement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes WDSNet, a weakly-supervised wavelet basis selection network with a category representation mechanism (CRM) and feature supervision mechanism (FSM) that denoises low-cost smartphone IMU signals.
Problem
Low-cost inertial sensor signals contain severe intrinsic errors, and the choice of wavelet basis for denoising is subjective, which limits downstream tasks such as trajectory reconstruction and motion semantic recognition.
Approach
A 1D-ResNet classifier whose FC layer columns represent candidate wavelets selects the wavelet for thresholding; CRM adds sparsity and Rényi-entropy encoding regularizers so the FC weights act as category representations that directly supervise the feature extractor, with attitude and displacement prediction as weak guidance tasks instead of clean-signal labels.
Key results
- On Allan variance, WDSNet reaches 0.06 QN for acceleration and 0.09 QN for angular velocity, versus 0.20 and 0.48 for the best baseline kNN.
- On downstream tasks, WDSNet lowers attitude error to 3.79 deg (-64.35% vs raw) and position error to 0.28 m, and raises semantic recognition accuracy to 87.01%.
- Ablation shows CRM is needed for scaling to 16 candidate wavelets: without CRM, performance degrades from 10 to 16 wavelets, while with CRM 16 wavelets gives the best results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CRM with 16 candidate wavelets (ablation) | WDSNet · Trajectory reconstruction | FSSE: 0.28 -> 0.10 | Helped |
| Add CRM with 10 candidate wavelets (ablation) | WDSNet · Trajectory reconstruction | FSSE: 0.24 -> 0.20 | Helped |
| Increase candidate wavelets from 5 to 10 without CRM (ablation) | WDSNet · Trajectory reconstruction | FSSE: 0.28 -> 0.24 | Helped |
| Increase candidate wavelets from 10 to 16 without CRM (ablation) | WDSNet · Trajectory reconstruction | FSSE: 0.24 -> 0.28 | Hurt |
| Increase candidate wavelets from 5 to 16 with CRM (ablation) | WDSNet · Trajectory reconstruction | FSSE: 0.25 -> 0.10 | Helped |
| Use FSM to supervise features directly with category representations (ablation) | WDSNet · Wavelet-category feature visualization | feature organization | Helped |
| Use WDSNet instead of kNN | inertial signal enhancement · Allan variance, acceleration | QN: 0.20 -> 0.06 | Helped |
| Use WDSNet instead of the strongest baseline | inertial signal enhancement · Attitude estimation | error (deg): 4.43 -> 3.79 | Helped |
| Use WDSNet instead of the strongest baseline | inertial signal enhancement · Position estimation | error (m): 0.38 -> 0.28 | Helped |
| Use WDSNet instead of the strongest baseline | inertial signal enhancement · Semantic recognition | accuracy: 86.22% -> 87.01% | Helped |
| Use WDSNet instead of the strongest baseline | inertial signal enhancement · Trajectory reconstruction | FSSE: 0.18 -> 0.10 | Helped |
Surprising findings
- Without CRM, adding more candidate wavelets (5 to 16) does not monotonically improve signal enhancement and can degrade it.
- Signal enhancement offers limited help for motion semantic recognition, since handwriting variability dominates the challenge rather than signal noise.
Limitations
- Evaluated on a single self-collected dataset of smartphone IMUs with a single training phone model.
- Relies on a set of 16 candidate wavelet bases chosen by the authors; generalization beyond them is not studied.
- Guidance tasks require attitude and position ground truth from a robotic arm, which is still a form of supervision.
Experimental setup
Datasets: self-collected smartphone IMU dataset
Benchmarks: Allan variance analysis (QN, ARW, VRW, BI) attitude estimation error position estimation error motion semantic recognition accuracy FSSE trajectory reconstruction
Metrics: Allan variance indicators attitude error (deg) position error (m) recognition accuracy Fréchet spline sliding error
Baselines: Kalman filter Savitzky-Golay filter EMD-Kalman filter CNN GRU-LSTM Optimized GRU-LSTM kNN
Builds on: 1D-ResNet wavelet thresholding Allan variance
Closest papers by meaning
- iMoT: Inertial Motion Transformer for Inertial Navigation (AAAI 2025 · similarity 0.81)
- HEROS-GAN: Honed-Energy Regularized and Optimal Supervised GAN for Enhancing Accuracy and Range of Low-Cost Accelerometers (AAAI 2025 · similarity 0.80)
- Scalable Hierarchical Deep Neural Network for Time Series Analysis in Wearable Sensor-based Human Activity Recognition (AAAI 2025 · similarity 0.80)
- Inertia-1: An Open Exploration of Wearable Motion Foundation Models (NeurIPS 2026 · similarity 0.80)
- Spatial-Related Sensors Matters: 3D Human Motion Reconstruction Assisted with Textual Semantics (AAAI 2024 · similarity 0.80)
- SEA-GWNN: Simple and Effective Adaptive Graph Wavelet Neural Network (AAAI 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.