A two-phase hybrid clustering framework exploring transitional activities in HAR.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) faces challenges with high-dimensional, noisy sensor data and limited labeled data in unsupervised learning.
• Traditional clustering models struggle with time-series sensor data, despite good performance on simulated datasets.

Purpose of the Study:

• To explore autoencoder (AE) architectures for dimensionality reduction and feature extraction from streaming HAR datasets.
• To develop an effective unsupervised clustering model for identifying human activity patterns from sensor data.

Main Methods:

• Investigated various autoencoder architectures including convolutional, LSTM, and hybrid CNN-LSTM layers for spatio-temporal feature extraction.
• Employed supervised learning to train the AE model and an unsupervised K-Means clustering model on extracted features.
• Utilized MobiAct and UCI HAR datasets for model evaluation.

Main Results:

• The hybrid convolutional AE+LSTM feature extractor combined with K-Means achieved state-of-the-art clustering accuracy (up to 0.99 NMI and ARI).
• Demonstrated over 50% improvement in clustering performance compared to previous methods.
• Presented cluster visualizations to explain transitional activity patterns.

Conclusions:

• The proposed integrated hybrid model effectively addresses challenges in unsupervised HAR using wearable sensor data.
• Achieved superior performance in identifying human activity patterns, outperforming existing approaches.
• The method offers a robust solution for real-world HAR applications with unlabeled sensor streams.