Using a Semi-supervised Learning Model for Recognition of Human Daily Activities from Wearable Sensor Data

Abstract

The application of Machine Learning (ML) and Artificial Intelligence (AI) is growing, and also becoming more important as the aging population increases. Smart support systems for distinguish Activities of Daily Living (ADL) can help the elders live more independently and safely. Many machine learning methods have been proposed for Human Activity Recognition (HAR), including complex networks containing convolutional, recurrent, and attentional layers. This study explores the application of ML techniques in ADL classification, leveraging wearable devices' time-series data capturing various parameters such as acceleration. The acceleration data obtained from sensors is so huge that it is difficult and expensive to accurately label every sample collected, so this study applies the Semi-supervised Learning model to unlabeled samples. Long Short-Term Memory (LSTM) has always been used for time series data such as acceleration, and recently, the Transformer model has emerged in many applications such as Natural Language Processing (NLP) or creating ChatGPT. In this study we proposed ADL classification method using the Self-Attention Transformer block and the Recurrent LSTM block and evaluated their results. After comparison, the model built with LSTM block gives better results than the model built with Transformer block.