Subject independent slip detection using hip exoskeleton sensors
Author(s)
Peterson, Reese
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Slip and fall incidents are some of the most common causes of injuries for workers as
well as the elderly. Due to this, significant research has been conducted towards preventing
slips and assisting in recovery from slips. A major focus within that research has been
on the detection of slips. A majority of the research into slip detection has been focused
on either detecting the occurrence of a fall due to a slip or just detecting slips accurately,
with minimal measurement of the time it takes to detect the slips. Research into detection
algorithms that have reported detection speeds and accuracies are sparse, with most only
focused on early slips. Thus, there is a space to pursue alternate methods of detection, as
well as develop detection methods for both early and late slips.
This thesis is a study on the use of machine learning to detect early and late slips while
using only sensors on a hip exoskeleton. The main hypothesis for this thesis was that by
using machine learning, it is possible to improve detection time to within 250 ms of slip
onset while maintaining a detection accuracy of greater than 90%. To complete the study,
an exoskeleton for walking assistance and a protocol for simulated slips on a treadmill to
collect slip data while wearing said exoskeleton. Using the data from the slip protocol,
three machine learning models were developed and optimized for slip detection: LDA,
XGBoost, and CNN. From the results obtained through the optimization of the models, we
were able to successfully improve detection time for both early and late slips through the
use of XGBoost, with a detection time of 157.23 and 237.26 ms respectively. We were also
able to keep accuracy above 90% for both early and late slips, with respective accuracies
of 93.25% and 92.3%.
Sponsor
Date
2022-05-03
Extent
Resource Type
Text
Resource Subtype
Thesis