Cross-Domain Human Activity Recognition via Domain Adaptation and Fused Attention.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Wearable sensors are crucial for Human Activity Recognition (HAR) in health monitoring and sports.
• HAR models struggle with generalization due to limited labeled data for complex activities.

Purpose of the Study:

• To develop a novel network, Transfer Component Analysis-Bidirectional Long Short-Term Memory (TCA-BiLSTM), to enhance HAR generalization.
• To address the challenge of insufficient labeled data in sensor-based HAR.

Main Methods:

• Utilized Transfer Component Analysis (TCA) with Maximum Mean Difference (MMD) in Reproducing Kernel Hilbert Space (RKHS) for domain adaptation.
• Employed a two-layer Bidirectional Long Short-Term Memory (BiLSTM) network with a fused attention mechanism for activity classification.
• Aligned sensor data from different body parts to facilitate cross-domain HAR data mapping.

Main Results:

• TCA-BiLSTM demonstrated superior performance compared to state-of-the-art methods like DSAN and FNet.
• Achieved performance improvements of 6.1% over DSAN and 2.5% over FNet on the DSADS and PAMAP2 datasets.
• Effectively captured multi-granularity activity information post-TCA domain adaptation.

Conclusions:

• The proposed TCA-BiLSTM network significantly improves generalization in sensor-based HAR.
• The fused attention mechanism enhances the classification of unseen activities after domain adaptation.
• TCA-BiLSTM offers a promising solution for robust HAR in real-world applications with limited labeled data.