Submit

Login Login

Register Register

Home / Journals / CMC / Online First / doi:10.32604/cmc.2026.086339
Journal Menu
Special Issues
Table of Content

All issues

Online First

2026

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

2005

2004

Open Access

ARTICLE

Attention-Driven Multi-Kernel Edge–Cloud Learning for Wearable IMU-Based Human Activity Recognition in IoT Health Monitoring

Fatimah Alhayan1, Muhammad Hanzla2, Hadeel Alsolai1, Bayan Alabdullah1, Ahmad Jalal2,3,*, Hui Liu4,*
1 Department of Information Systems, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia
2 Department of Computer Science, Air University, E-9, Islamabad, Pakistan
3 Department of Computer Science and Engineering, College of Informatics, Korea University, Seoul, Republic of Korea
4 Guodian Nanjing Automation Co., Ltd., Nanjing, China
* Corresponding Author: Ahmad Jalal. Email: email; Hui Liu. Email: email

Computers, Materials & Continua https://doi.org/10.32604/cmc.2026.086339

Received 28 May 2026; Accepted 19 August 2026; Published online 14 September 2026

Abstract

Human activity recognition (HAR) from wearable inertial sensors underpins IoT-based smart-health monitoring, yet practical deployments are constrained by noisy Inertial Measurement Unit (IMU) signals, inter-subject variability, temporal misalignment across sensor streams, and the limited compute available on edge nodes. Existing pipelines typically optimize recognition accuracy in isolation from these deployment constraints. This paper addresses both aspects through a distributed edge–cloud HAR framework. At the edge, IMU signals are denoised using Tukey filtering and segmented using Planck-Taper windowing before compact features are transmitted, reducing uplink overhead. At the cloud, an attention-based deep multiple-kernel learning (A-DMKL) adaptively combines multi-resolution representations (Stockwell Transform (S-Transform), Gramian Angular Summation Field (GASF), Synchrosqueezed Wavelet Transform (SSWT), Hankel), which are optimized using the Coyote Optimization Algorithm (COA) and classified through Soft Dynamic Time Warping Soft-DTW temporal alignment, a Hierarchical Hidden Semi-Markov Model, and an InceptionTime ensemble. Under strictly subject-independent leave-one-subject-out evaluation on 24 and 8 participants, with all fitting steps confined to the training folds, the framework attains 87.22% and 91.33% accuracy on the Hang-Time HAR and VIMPT2019 datasets, respectively. To assess deployment robustness, the proposed pipeline is analyzed under simulated network service profiles, Message Queuing Telemetry Transport (MQTT) packet loss, and inter-stream synchronization error, with all simulation assumptions explicitly stated. The framework targets communication-efficient human activity recognition for IoT healthcare applications. Clinical validation and deployment on Advanced RISC Machine (ARM-class) embedded hardware remains beyond the scope of the present study and are identified as future work.

Keywords

Human activity recognition; wearable inertial sensors; multiple kernel learning; attention mechanism; time–frequency analysis; edge–cloud computing; Internet of Things (IoT); smart health monitoring

  • 76

    View

  • 15

    Download

  • 0

    Like

Related articles
Copyright© 2026 Tech Science Press
© 1997-2026 TSP (Henderson, USA) unless otherwise stated

Share Link