aalto1 untyped-item.component.html
Predicting missing data based on multi-site wearable sensor data
Loading...
URL
Journal Title
Journal ISSN
Volume Title
School of Science |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Major/Subject
Mcode
Language
en
Pages
58
Series
Abstract
Human motion tracking and human activity recognition technology have significant applications in medical rehabilitation, virtual reality, and sports science. However, its deployment faces two main challenges: 1) sensor data collection can be disrupted by environmental interference, causing data loss or quality fluctuations; 2) traditional motion capture systems, which rely on dense sensor networks, negatively impact wearability and energy efficiency.
To address these issues, this study explores several deep-learning-based frameworks for sensor data generation. The frameworks aim to accurately predict missing motion data from other sites' sensor information, supporting lightweight motion capture systems.
The study reviews sensor data imputation and generation methods, comparing four deep-learning models: 1) convolutional autoencoder (AE) for feature reconstruction; 2) variational autoencoder (VAE) with probabilistic modelling; 3) convolutional neural network combined with long short-term memory (CNN-LSTM) for spatiotemporal feature fusion; 4) conditional generative adversarial network (cGAN) with adversarial training.
To assess the impact of sensor topology on prediction, two experimental scenarios are designed. The first scenario uses a global mode with all non-target site sensor data. The second one uses a local mode with only neighbouring site sensor data of the target site.
The experiments use a multi-metric evaluation system. Results show that in global sensing mode, CNN-LSTM, with hierarchical spatiotemporal feature extraction, performs best in loss control, temporal correlation, and signal fidelity. cGAN's accuracy drops due to adversarial training induced uncertainty. In local sensing mode, VAE shows more robustness, but all models' absolute performance declines significantly, indicating limited predictive capability with neighbouring node data.
The study finds: 1) Discriminative models (e.g., CNN-LSTM) excel in deterministic prediction with sufficient global data by explicitly modelling spatiotemporal dependencies; 2) Generative models (e.g., VAE) are more adaptable to local data (smaller datasets) due to their probabilistic representation; 3) Model performance is highly sensitive to sensor topology, requiring dynamic architecture selection based on sensor density. These findings offer theoretical support for missing-sensor prediction and lightweight motion capture systems. Future research should further explore the impact of sensor node spatial topology on generation.