Innovative Dual-Decoupling CNN With Layer-Wise Temporal-Spatial Attention for Sensor-Based Human Activity...

Area of Science:

• Computer Science
• Biomedical Engineering
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for health monitoring, but traditional methods struggle with complex sensor data dependencies.
• Existing attention-based models often oversimplify temporal and spatial features, limiting high-dimensional interaction modeling.
• Inadequate feature extraction leads to suboptimal performance in health-related HAR applications like fall detection.

Purpose of the Study:

• To introduce a novel framework, CNN-TSFDU-LW, for enhanced Human Activity Recognition (HAR).
• To decouple temporal and spatial dependencies in sensor data for more precise feature extraction.
• To improve computational efficiency and memory usage in HAR models.

Main Methods:

• Developed the Temporal-Spatial Feature Decoupling Unit (TSFDU) for parallel processing of temporal and spatial features.
• Implemented layer-wise training with a local error function for independent CNN layer updates.
• Utilized Convolutional Neural Networks (CNNs) integrated with the TSFDU mechanism.

Main Results:

• Achieved accuracy improvements of 0.9% to 4.19% over state-of-the-art methods on benchmark datasets (UCI-HAR, PAMAP2, UNIMIB-SHAR, USC-HAD).
• Demonstrated high accuracy rates: 97.90% (UCI-HAR), 94.34% (PAMAP2), 78.90% (UNIMIB-SHAR), and 94.71% (USC-HAD).
• Reduced computational complexity and memory requirements without performance degradation.

Conclusions:

• The CNN-TSFDU-LW framework significantly advances sensor-based HAR by enhancing accuracy and efficiency.
• Decoupling temporal and spatial features leads to richer representations and better performance.
• The model shows strong potential for improving health monitoring systems through more effective HAR.