Hand prosthesis have helped amputees to partially live normal lives as healthy people
for ages since invented. However, due to the technical limitations such as the essential
time required for signal sending and receiving, the state-of-the-art hand prosthesis still
cannot fully restore real hand functions. As Artificial Intelligence (AI) technologies,
especially deep learning and artificial neural networks, nowadays show an impressive
performance when building smart machines, it is possible to use them to bring
improvements to conventional hand prosthesis. This advanced AI prosthesis can learn
real hand functions through self-learning and eventually, fully achieve all real hand
functions. Moreover, with the help of current biologically inspired neural network
models and spiking neural networks (SNN), the power consumption and reaction delays
of a prosthesis can be further minimized.
A novel approach which attempts to address the long-existing problems such as
frequent misclassification faced by current hand prosthesis is presented in this thesis.
Through converting the raw surface electromyograph (sEMG) signals from amputees
with different hand amputation levels and able-bodied people into heatmaps, with an
applied properly designed convolution neural network which extracts and learns the
features contained within the heatmaps, the mis-trigger disadvantage rapidly decreases.
The experimental results, from 8 hand gestures so far, indicate that this novel approach
is effective. Moreover, the relationship between the number of sensors, used to record
the sEMG signals, and the recognition accuracy is also examined and presented in this
thesis. According to the experimental results, the optimal or required number of sensors
when recording sEMG signals can be minimized for different hand gestures without
classification accuracy degradation.
This thesis also contains another novelty: a spiking sEMG signal maps-based hand
gesture classification algorithm. The algorithm employs a spiking neural network as the
classifier, as well as a common heatmap technique which efficiently converts the raw
sEMG signals into common heatmaps that can be used by the spiking neural network.
The classification results obtained by two different sEMG datasets denote high
robustness of the novel algorithm. Moreover, further evaluation of the experimental
results denotes that both the mis-trigger issue as well as power consumption are reduced
when applying this novel algorithm.
Additionally, to further utilize the potential of sEMG signals, another novel algorithm
which aims at frequency domain features is presented in this thesis. This algorithm
successfully extracts the frequency domain features by applying singular system analysis
(SSA) to the raw sEMG signals and converts them into frequency density maps (FDM).
With the help of a proposed SNN, the algorithm is proved to be efficient for different
hand gestures. In addition, further evaluation of this algorithm indicates that it
demonstrates a significant reduction in computational costs, training time, power
consumption whilst, at the same time, results in lower classification errors/mis-triggers
when compared to other state of the art hand gesture recognition methodologies.
| Date of Award | 22 Jan 2026 |
|---|
| Original language | English |
|---|
| Awarding Institution | - University Of Strathclyde
|
|---|
| Sponsors | University of Strathclyde |
|---|
| Supervisor | Lykourgos Petropoulakis (Supervisor), John Soraghan (Supervisor) & Jaime Zabalza (Supervisor) |
|---|