Non-parametric and transfer learning algorithms for autonomous wearable computing

Yuchao Ma

With the rapid integration of wireless sensing technologies and computational algorithms, various motion analysis applications have come to realization for use with wearable sensor devices. These technologies have attracted significant attention in recent years, due to their great potential in healthcare and wellness domains. This dissertation focuses on (1) enhancing the utilization of wearable sensing systems to advance personalized health examination, and (2) developing computational autonomy solutions to support accurate motion analysis for use in uncontrolled environments. Gait analysis is a typical application in motion analysis, which measures one's mobility for health assessment. We design and develop a shoe-integrated sensing system for continuous data collection and quantitative gait analysis. We analyze sensor data collected from patients with Rett Syndrome and glaucoma during a series of standard gait tests in two clinical studies. We present signal processing algorithms to obtain patients' biophysical information, and develop machine learning models using feature representation of the sensor data, to automatically identify pathological gait patterns. To scale up gait analysis applications to less controlled environments, we propose a platform-independent framework for gait cycle detection, a major task in gait analysis. We utilize physical properties of human gait to enable autonomous parameter learning and model reconfiguration as sensor platform properties such as bit resolution, sampling frequency, signal dynamic range, and sensor orientation change. To further support sensor-enabled motion analysis in dynamic settings, we propose two autonomous algorithms to improve the performance of machine learning models for activity recognition. We design an asynchronous transfer learning algorithm, to reduce the performance degradation of activity recognition models caused by considerable differences among various sensing platforms, as well as variations in movement patterns across users. In addition, we propose a novel non-parametric semi-supervised learning framework, to overcome the dependency on large amounts of labeled data for machine learning model training. The proposed framework employs a graph-based approach for sample selection and label inference, and a silhouette-based filtering strategy to finalize the obtained labels to construct a highly reliable activity recognition model.

Non-parametric and transfer learning algorithms for autonomous wearable computing

Files and links (1)

Abstract

Metrics

Details