Date of Award

August 2017

Degree Type


Degree Name

Master of Science



First Advisor

Lingfeng Wang

Committee Members

Chiu Tai Law, Chao Zhu


There are two parts in this thesis: the first part was conducted at UWM, and the second part was conducted at Johnson Controls using the knowledge and skills that I learned throughout my time in the Master’s Degree program.

The primary purpose of my time at UWM was to compare different types of buildings with two popular machine learning regression algorithms, artificial neural network (ANN), supported vector machine regression (SVR) algorithms, and lastly to provide the results of my research to better help building managers make more informed decisions in regard to electrical utilities. The major objective is to use algorithms and neural networks to detect the occupancy of a room using real-time data from accelerometers. This data could then be used to enable HVAC systems to be more efficient and intelligent.

My research at UWM consists of 6 chapters. The background and related research are shown first in chapter 1 and chapter 2. Chapter 3 focuses on analyzing different building types, which aims to provide an overlook in the feature of the data. The basic concepts of ANN and SVR are included in the Chapter 4. The last chapter is the summary of internship in Johnson Controls during the summer. The project goal, data analysis and results are presented with details. A brief occupancy detection review of the industry as well as the basic knowledge of Wavelet Transform and K-means++ algorithm are also mentioned in Chapter 7.

The result of my research at UWM shows that it is necessary to apply different models for different building types if high accuracy is required. Compared to SVR, ANN is more accurate in all the building types. However, the difference of the accuracy depends on the building features. In a hospital, SVR and ANN both show high accuracy, but in restaurants, they are both underperforming. Additionally, using vibration magnitude measured from accelerometers to detect occupancy has proved to be feasible during the first stage. However, more complicated cases and patterns need to be considered and higher resolution sensors will need to be tested in the future work.

Available for download on Saturday, August 31, 2019