Electronic Thesis and Dissertation Repository

Thesis Format

Monograph

Degree

Master of Engineering Science

Program

Electrical and Computer Engineering

Collaborative Specialization

Artificial Intelligence

Supervisor

Grolinger, Katarina

2nd Supervisor

Trejos, Ana Luisa

Co-Supervisor

Abstract

When humans repeat the same motion, the tendons, muscles, and nerves can be damaged, causing repetitive stress injuries (RSI). Symptoms usually begin slowly and become more intense and constant over time. If the motions that lead to RSI are recognized early, these injuries can be prevented. A preventative approach could be implemented in factories to warn workers about possible injuries. By detecting the movements that can cause RSI, the worker can be alerted to stop carrying out those movements. For this purpose, machine learning models can detect human motion with the human activity recognition (HAR) model. HAR models typically require data from each participant before being trained; therefore, they cannot easily be adapted to new participants. This problem arises because humans move differently. To solve this problem, a model can be personalized to a particular individual to help detect their movements more easily. The model training procedure to create a personalized model consists of two phases: create a generic model, and then personalize the generic model with transfer learning. In this thesis, CNN, transformer, and Trans-CNN were selected for the model training procedure. To assess the model training procedure, the WISDM 2019 dataset was selected. Both the generic model and personalized model were evaluated with three different methods: all movement, only upper body movement, and lower body movement. In each of the evaluations, the same following trends were seen: personalization increased all of the performance metrics for all three models; the generic Trans-CNN model significantly outperformed the other two generic models for all four performance metrics; and there were no statistical differences between the personalized model. When evaluating only lower body movement data, the generic model performed substantially higher than when evaluating with only upper body movement data and slightly higher metrics when all movement data are used. A personalized model, however, performed almost identically across all evaluations, no matter the kind of data used (all movement, upper body movement, or lower body movement). This study demonstrates that using HAR models can potentially detect motions that cause RSI, which could result in significant financial benefits for society.

Summary for Lay Audience

Repeating the same motion can lead to muscle and nerve injuries. Normally, these injuries develop slowly and the pain becomes more constant over time, but if the repetitive movements are recognized early and stopped, injuries can be easily prevented. Most algorithms for detecting human motion require data upfront since people move differently. To solve this problem, an algorithm can be personalized to a particular person to help detect their movements more easily. This thesis proposes a method to detect upper and lower body movements with a personalized algorithm. This method involves creating a general algorithm that generally understands how humans move, then personalizing that algorithm to each person. The algorithms were tested in the following ways: only upper body movement data, lower body movement data, and all movement data. The same trends were seen in all the tests: the personalized algorithm performed better in detecting human motion than the general algorithm. The personalized algorithm performed almost the same with different kinds of data. In contrast, when the same tests were carried out on the general algorithm, the results differed with different data. The personalized algorithm shows promise in accurately detecting repetitive movements to prevent injuries.

Download

Share

COinS