Active learning for IoT sensor-based healthcare monitoring

Abstract

This dissertation presents a comprehensive study on resource-efficient human activity recognition (HAR) with a novel active learning strategy for supporting IoT-based healthcare monitoring. The study first introduces a novel deep learning-based active learning method designed to address the challenges associated with data labeling in supervised model training for HAR. With an extension of a single autoencoder to a multiclass scheme, the proposed method exploits the unique pattern characteristics inherent to each activity class for preferentially identifying the most informative samples for labeling. In addition, the proposed method introduces a margin scoring criteria based on cosine similarity by comparing pattern characteristics of each activity class to strategically select samples for constituting an informative training dataset, thereby enhancing the effectiveness of the process of supervised learning and the overall performance of the human activity recognition model. The method demonstrates superior classification performance of the HAR model, outperforming existing selecting methods. Additionally, the proposed methodology shows exceptional performance in terms of efficiency, leading to a substantial reduction in labeling costs. This is achieved by strategically selecting fewer, but more informative samples, thereby maintaining the same level of performance with reduced labeled samples. This study contributes to the field of IoT-based healthcare monitoring, fostering advancements in human activity recognition. By efficiently curating informative training datasets, the proposed method increases the effectiveness of the learning process of HAR. Furthermore, this study presents a resource-efficient human activity recognition in IoT environments, addressing the challenge of resource requirements in the deployment of IoT-embedded devices. The proposed method embraces a structured framework that integrates several essential components, including feature extraction in different levels of hierarchy within time series data, as well as an advanced technique to leverage temporal relationships between each hierarchically extracted feature. Specifically, the proposed method efficiently exploits temporal causality in feature spaces, thereby the number of recurrent operations is significantly reduced. The proposed method demonstrates superior classification performance over existing methods, particularly with substantial reductions in computational complexity and model complexity across diverse datasets. In addition, the proposed method shows a significant reduction in inference time and energy consumption in scenarios of model deployment at IoT-embedded devices, which results in advantages such as exceptional resource efficiency of the proposed method in a resource-constrained environment. This research significantly contributes to the development of an efficient HAR model, ensuring the integrity and reliability of IoT-based monitoring applications.