Related Experiment Video
Updated: Jan 7, 2026
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
5.2K
ASSAFormer: a sensor data-based approach to human activity recognition.
Jinzhu Zeng1, Beiping Peng1, Changzhou Chen2
1School of Physical Education, Hunan University of Finance and Economics, Hunan, China.
Scientific Reports
|January 4, 2026
Summary
ASSAFormer enhances human activity recognition for health monitoring by integrating mode decomposition and an improved Transformer model. This method improves accuracy and generalization, overcoming noise and data variability challenges in wearable sensor data.
Area of Science:
- Biomedical Engineering
- Artificial Intelligence
- Wearable Technology
Background:
- Human Activity Recognition (HAR) is crucial for health monitoring and intelligent healthcare, enabling continuous tracking and behavioral state analysis.
- Existing HAR methods struggle with noise and data variability from wearable sensors, limiting accuracy and generalization across diverse scenarios and individuals.
- Challenges include maintaining stable early-warning performance due to motion variability and individual differences.
Purpose of the Study:
- To propose a novel human activity recognition method, ASSAFormer, to address the limitations of existing HAR techniques in health monitoring.
- To improve the accuracy and generalization ability of HAR systems by mitigating noise and data distribution variability.
- To enhance the reliability of early-warning systems in diverse motion scenarios and for individual differences.
Main Methods:
- ASSAFormer integrates Variational Mode Decomposition (VMD) for noise filtering and the Whale Optimization Algorithm (WOA) for optimizing VMD parameters.
- The core architecture features an improved Transformer incorporating Adaptive Sparse Self-Attention (ASSA) and Contrastive Normalization (ContraNorm).
- ASSA combines Sparse Self-Attention (SSA) for relevant information filtering and Dense Self-Attention (DSA) to retain potentially overlooked useful information, while ContraNorm addresses dimensional collapse.
Main Results:
- Comparative experiments demonstrated that ASSAFormer achieved superior performance on both the UCI and URFD datasets.
- Ablation studies confirmed the effectiveness of the individual improved modules within the ASSAFormer architecture.
- The method successfully addressed noise interference and improved representation dispersion in the feature space.
Conclusions:
- ASSAFormer presents a robust solution for human activity recognition in health monitoring, outperforming existing methods.
- The integration of VMD, WOA, ASSA, and ContraNorm effectively enhances HAR accuracy and generalization.
- The proposed method offers a reliable approach for intelligent healthcare applications requiring precise activity recognition.
