Abstract:
Electrical potentials are responsible for activating and coordinating skeletal muscle movement
in the human body. These potentials, called myoelectric activity, can be registered
and analysed using non-invasive recording electrodes in a technique known as surface electromyography
(sEMG). Key information about the muscles can then be extracted from
EMG signals using signal processing techniques and used in various applications ranging
from electrophysiology studies to robotics. More recently, high-resolution multi-electrode
arrays have expanded to the field of EMG, a technique named high-density electromyography
(HD-EMG). This technique retains the beneficial aspects of sEMG while increasing
the resolution of the collected data, which can provide an improved understanding of the
spatial distribution and electrophysiological principles of myoelectric activity. These new
insights into the mechanisms of voluntary muscle contraction can also be used to improve
myoelectric human-machine interfaces, which allow users to have an embodied interaction
with robotic devices using intuitive controls. This thesis focuses on the development of
improved methods for recording and analysing high-density EMG signals from the hand and
forearm to apply them to design novel machine-human interfaces for robotic control.
A variety of flexible high-density electrodes were designed and fabricated to efficiently
and accurately record myoelectric activity from the hand and forearm. Among these, a set
of novel electrode arrays was developed to specifically target the intrinsic hand muscles.
These consisted of 60 individual electrodes with a 4 mm inter-electrode distance (IED)
divided in two arrays. The individual electrodes were arranged in groups targeting ten
intrinsic muscles on the anterior and posterior sides of the hand. The hand arrays were
paired with a set of tessellating rectangular arrays of 32 electrodes with a 7 mm IED, which
can cover an area of up to 50 × 170 mm around the arm or forearm to record 256 EMG
signals simultaneously. All arrays were constructed using a flexible substrate with a thickness
less than 0.11 mm, which allowed users to retain full mobility during experimental tasks.
Signal processing methods for HD-EMG were developed, including noise reduction with
finite impulse response (FIR) and infinite impulse response (IIR) filters; data analysis with
signal segmentation, feature extraction, data reduction and interpolation techniques. For
visualisation, the distribution of myoelectric activity was imaged by transforming HD-EMG
data into 2D spatiotemporal maps. These maps allowed for a graphical representation of
muscle activity during voluntary movement. The applied signal processing methods were
then paired with machine-learning (ML) algorithms to develop a framework for classifying
HD-EMG signals using feature extraction and pattern recognition. This framework relied
on the extraction of features from EMG signals which could then be used to train models
capable of predicting the intended user motion based on myoelectric information. Five
time-domain (TD) features, zero crossings (ZC), slope sign changes (SSC), waveform length
(WL), and Willinson amplitude (WAMP), frequency-domain (FD) features such as median
frequency (MDF), and temporal-spatial domain (TSD) features were calculated and used
in four types of ML classifiers, support-vector machines (SVM), linear discriminant analysis
(LDA), neural networks (NN) and k-nearest neighbour (KNN) algorithms.
HD-EMG data was collected from 26 subjects while performing several different motions.
Studies were performed to determine the optimal configuration of the classification
framework according to different tasks. The influence of segmentation parameters, channel
count and feature selection was evaluated to determine their weight over the accuracy of
pattern-based classifiers.
Among the different parameters, channel count was found to have a significant impact.
Classification accuracy ranged from 74±4% when using only 5 bipolar channels, to 92±2%
with 41 channels recording EMG signals from hand muscles to classify grasps and finger
motions. In general, combinations of hand and wrist motions were decoded with an accuracy
ranging from 75 − 98% using HD-EMG data collected from intrinsic and extrinsic muscles
across the different experiments using the developed classification framework. The improved
spatiotemporal resolution of the arrays was also used to visually track the propagation of
motor-unit action potentials (MUAPs) to estimate the conduction of muscle fibres, which
yielded 4.7±0.3 m/s and were in range with previously reported literature values using
invasive EMG techniques.
Based on the analysis of the different factors that influenced the accuracy of HD-EMG
classification, the electrode arrays and classification framework were applied to two applications
for robotic control. The first was a virtual environment constructed using the
Webots robotic simulation platform, where a digital twin of an anthropomorphic robot was
controlled using HD-EMG signals to interact with objects within a fully rendered 3-D environment
by executing different hand grasps and motions. The second was the application of
HD-EMG electrode arrays towards the development of a control scheme for dexterous object
manipulation using robotic hands based on regression-based motion decoding. Regression
models were trained using HD-EMG signals and motion-tracking as inputs to decode the
position and trajectory of an object. The models achieved an accuracy of 63±7% and a
correlation of 81±5% between the real and predicted objects. These models were then
applied to control an under actuated anthropomorphic robot hand using HD-EMG inputs.
The work in this thesis presents an overview of the design, optimization and incorporation
of HD-EMG techniques into myoelectric control tasks. The proposed new technical
approaches can also contribute to electrophysiological research and improve the quantity
and quality of decoded motions in classification-based EMG interfaces.