Augmentation-Free Self-Supervised Human Activity Recognition With Attention Mechanism and Adaptive Time Series Mixer

Area of Science:

• Machine Learning
• Signal Processing
• Biomedical Engineering

Background:

• Self-supervised learning (SSL) excels in human activity recognition (HAR) by leveraging unlabeled data.
• Existing SSL methods often rely heavily on data augmentation and can misinterpret noise.
• Limited dataset scale can hinder accuracy and practical application of HAR models.

Purpose of the Study:

• To propose a novel SSL objective for HAR that mitigates reliance on data augmentation and improves noise handling.
• To enhance the extraction of global and local features from inertial measurement unit (IMU) data.
• To improve the accuracy and generalizability of HAR models.

Main Methods:

• Developed a new SSL objective integrating an attention mechanism and an adaptive time series mixer.
• Focused on capturing global dependencies and local features within IMU data without data augmentation.
• Assigned lower weights to noise during the mask reconstruction process.

Main Results:

• The proposed model demonstrated significant improvements on both self-collected (CQJTU-FCE) and public datasets (UCI, Motion, HHAR).
• Achieved average accuracy, F1 score, and Cohen's kappa improvements of 6.54%, 8.55%, and 7.88% on the self-collected dataset.
• Showcased average enhancements of 10.63%, 11.77%, and 13.39% across metrics on public datasets, confirming generalizability.

Conclusions:

• The proposed SSL method offers a more efficient and reliable solution for HAR tasks.
• The model effectively extracts relevant features from IMU data while minimizing the impact of noise.
• Validated generalizability across diverse datasets, indicating strong real-world applicability for HAR.