Deep Wavelet Convolutional Neural Networks for Multimodal Human Activity Recognition Using Wearable Inertial Sensors

Area of Science:

• Wearable systems and sensor technology.
• Machine learning for human activity recognition (HAR).
• Signal processing and deep learning.

Background:

• Wearable inertial sensors (accelerometers, gyroscopes) are compact, accurate, and multimodal.
• Multimodal HAR methods leverage sensor data but often lack domain knowledge or fail to capture time-frequency signal dependencies.
• Existing approaches struggle with complex multimodal sensor signal analysis.

Purpose of the Study:

• To propose a novel deep wavelet convolutional neural network (DWCNN) for enhanced multimodal human activity recognition.
• To effectively learn features from the time-frequency domain of multimodal sensor signals.
• To improve the accuracy and feature representation for wearable inertial sensor-based HAR tasks.

Main Methods:

• Developed Deep Wavelet Convolutional Neural Networks (DWCNN) integrating Continuous Wavelet Transform (CWT) with Deep Convolutional Neural Networks (DCNN).
• Proposed an algorithm to estimate the wavelet scale parameter for optimized CWT performance.
• Utilized DCNN with residual and attention blocks for feature extraction and multimodal signal fusion.

Main Results:

• Extensive experiments conducted on five benchmark HAR datasets (WISDM, UCI-HAR, Heterogeneous, PAMAP2, UniMiB SHAR).
• The proposed DWCNN model demonstrated superior performance compared to existing HAR methods.
• Achieved enhanced feature representation by capturing time-frequency dependencies in multimodal sensor signals.

Conclusions:

• DWCNN offers a significant advancement in multimodal human activity recognition using wearable inertial sensors.
• The integration of CWT and DCNN effectively addresses limitations of previous HAR approaches.
• The method shows strong potential for real-world applications requiring accurate activity recognition from sensor data.