Graduation Year

2021

Document Type

Dissertation

Degree

Ph.D.

Degree Name

Doctor of Philosophy (Ph.D.)

Degree Granting Department

Mathematics and Statistics

Major Professor

Kandethody M. Ramachandran, Ph.D.

Committee Member

Chris P. Tsokos, Ph.D.

Committee Member

Sherwin Kouchekian, Ph.D.

Committee Member

Lu Lu, Ph.D.

Committee Member

Yicheng Tu, Ph.D.

Keywords

Adaptive Learning, HAR, Machine Learning, Online Analysis, Smart Device

Abstract

Wearable human activity recognition (HAR) is a widely application system for our daily life. It hasbeen built in many devices, such as smartphone, smartwatch, activity tracker, and health monitor. Many researchers try to develop a system which requires less memory space and power, but has fast and accurate classification results. Moreover, the objective of adjusting the classifier by the system self is also a study direction. In the present study, we introduced the machine learning methods to both smartphone data and smartwatch data and an adjusted model with the continuous generating data. Further, we also proposed a new HAR system which could adjust it by customer’s personal input.

First, we present a comparison with several popular machine learning methods using smartphone data and try to find the most effective model for identifying activities, such as Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Linear Discriminant Analysis (LDA), Decision Tree (DT), and so forth. Two datasets with different transition methods from smartphone are used. Also, we used grid search, multi-fold cross validation, and dimensional reduction method to improve the performance. Meanwhile, a two-layer method for activity identification is proposed. This method is more flexible of choosing classifiers for activities.

Then, in order to avoid the fixed built in HAR system, we used an online method called Very Fast Decision Tree (VFDT) to mimic the real scenario, since we do not have data collected in a streaming basis. There are two main improvements from the existing models: 1) we train the model online and use the training data for training and adjusting then delete the previous data; 2) after building VFDT, the model can be adjusted to identify new activities by adding only small amount of labeled observations.

Finally, we proposed a personalized HAR system with interactive function. With this system, customers could build their personal HAR system by inputting their data. The system includes two steps, unsupervised method and supervised method. The unsupervised step is used for identifying if the new input data has any new activities. K-cluster is applied. The supervised step is used for identifying the specific activities and update the activity classes if there are new activities. Quadratic discriminant analysis (QDA) is applied.

Share

COinS