Personalized models for human activity recognition with wearable sensors: deep neural networks and signal processing

Human Activity Recognition (HAR) has been attracting research attention because of its importance in applications such as health monitoring, assisted living, and active living. In recent years, deep learning, specifically Convolutional Neural Networks (CNNs), have been achieving great results due to...

Full description

Saved in:
Bibliographic Details
Published in:Applied intelligence (Dordrecht, Netherlands) Netherlands), 2023-03, Vol.53 (5), p.6041-6061
Main Authors: Gholamiangonabadi, Davoud, Grolinger, Katarina
Format: Article
Language:English
Subjects:
Artificial Intelligence
Artificial neural networks
Computer Science
Customization
Data integration
Feature extraction
Human activity recognition
Machine learning
Machines
Manufacturing
Mechanical Engineering
Neural networks
Performance degradation
Processes
Signal processing
Wavelet transforms
Wearable computers
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Human Activity Recognition (HAR) has been attracting research attention because of its importance in applications such as health monitoring, assisted living, and active living. In recent years, deep learning, specifically Convolutional Neural Networks (CNNs), have been achieving great results due to their ability to extract features and model complex actions. These generic models work great for the subjects on which they were trained, but their performance degrades substantially for new subjects. Consequently, this paper proposes a personalized HAR model based on CNN and signal decomposition. First, features are extracted from multi-modal sensor data with signal processing techniques, including Stationary Wavelet Transform, Empirical Mode Decomposition (EMD), and Ensemble EMD. Next, CNN carries out the information fusion and the final classification. Personalization is achieved by using a few seconds of the target subject data to select the version of the trained CNN best suited for the target subject. Results show that EMD with cubic spline achieves better accuracy than other signal processing techniques. Moreover, the proposed approach, irrelevant of the type of signal processing, outperforms the state-of-the-art CNN approaches with time-domain features.
ISSN:0924-669X
1573-7497
DOI:10.1007/s10489-022-03832-6