Durham School of Architectural Engineering and Construction
First Advisor
Fadi Alsaleem
Date of this Version
11-2023
Citation
A thesis presented to the faculty of the Graduate College at the University of Nebraska in partial fulfillment of requirements for the degree of Master of Science
Major: Architectural Engineering
Under the supervision of Professor Fadi Alsaleem
Lincoln, Nebraska, November 2023
Abstract
Human activity detection is crucial to personalize the control of the building environment. For example, understanding certain human activities (e.g., walking, sitting, etc.) for an occupant in a building helps provide the proper thermal comfort control. However, these applications require large-scale neural networks that are challenging to implement and train.
In this thesis, we implemented a continuous-time recurrent neural network implementation (CTRNN) network that can solve real-time human activity detection with a smaller-size network. The CTRNN uses differential equations with a time constant to describe the neuron equations. It was implemented and trained for the first time using TensorFlow. Specifically, the forward path of the CTRNN was implemented using a new recurrent cell in TensorFlow, and the training was performed while utilizing the auto-differential function to implement the backpropagation through-time algorithm.
The CTRNN showed a high ability to capture a complex pattern in the temporal data of the acceleration measurements of human activities. More importantly, despite the smaller size, we show that the CTRNN outperforms the standard recurrent neural network (RNN) in accuracy, convergence, and stability. The research also investigates the impact of training the time-constant parameters of CTRNN for the first time to improve its performance.
Advisor: Fadi Alsaleem
Comments
Copyright 2023, Abdallah Al Zubi