MaskCAE: Masked Convolutional AutoEncoder via Sensor Data Reconstruction for Self-Supervised Human Activity...

Area of Science:

• Ubiquitous Computing
• Machine Learning
• Signal Processing

Background:

• Self-supervised Human Activity Recognition (HAR) is crucial for wearable devices.
• Manual labeling of sensor data is challenging and time-consuming.
• Current self-supervised methods face issues with data augmentation, model limitations, and computational demands.

Purpose of the Study:

• To propose a novel self-supervised approach for HAR that addresses existing challenges.
• To introduce a computationally efficient model for HAR on wearable devices.
• To fill the gap in masked sensor data modeling for HAR.

Main Methods:

• Developed Masked Convolutional AutoEncoder (MaskCAE), a denoising autoencoder approach.
• Utilized a fully convolutional network for reconstructing masked time-series sensor data.
• Evaluated performance across self-supervised, fully supervised, and semi-supervised settings.

Main Results:

• MaskCAE significantly outperforms state-of-the-art algorithms in HAR tasks.
• The model achieves high performance without requiring data augmentation.
• Demonstrated effective capture of temporal semantics in sensor data through visualization.

Conclusions:

• MaskCAE offers a powerful and efficient self-supervised solution for HAR.
• The approach shows great potential for modeling abstracted sensor data.
• The model's effectiveness is validated on an embedded platform.