Gait detection algorithm applying for transtibial prosthesis

New generations of powered transtibial prostheses have been developed through-
out the last decade to restore the functionalities of the biological leg and the ap-
pearance of the amputation. However, there are still limitations in technology as
amputees wearing prostheses walk at slower speeds and lose more metabolism
than those with the biological limb. Prosthetic design efforts have mainly been
focused on mechanical device, while gait phase detection algorithms are con-
strained. Especially gait phase detection is considered a great important param-
eter for creating the artificial limb to define where the prosthesis is in the gait
cycle during walking. Thus, this thesis aims to develop the gait phase predic-
tion embedded into the prosthetic control strategy for improving the precision
and stability of the prosthesis so that the user’s safety is enhanced. Therefore, a
novel gait phase prediction algorithm is proposed in this thesis that could predict
a full gait cycle discretised within one percent. This algorithm is called an Expo-
nentially Delayed Fully Connected Neural Network (ED-FNN). Currently, most
studies propose methods to detect from two to eight phases of the step. However,
for real-world applications, more than this may be required for controlling active
prosthetic devices. As a result, a more densely sampled gait phase is needed to
obtain more important gait phase information and give more control possibilities.
The performance of the proposed model ED-FNN is evaluated in terms of Accu-
racy, Mean Square Error (MSE). The obtained results of MSE demonstrate that
the proposed method is a robust and reliable solution for gait phase detection. Be-
sides, the prosthesis has yet to walk like a human because it is difficult to adjust
the movements to respond to the transitions to different environments that they
encounter during daily activities such as going up and down stairs, slopes, stop-
ping walking, and flat walking. To resolve this issue, the prosthesis is provided
with a locomotion mode recognition algorithm which helps the prosthesis switch
smoothly between different locomotion modes. Though existing studies have em-
ployed several classical machine learning methods for locomotion mode recogni-
tion, these methods are less effective for data with complex decision boundaries
and result in misclassifications of motion recognition. Deep learning is the lat-
est advancement of machine learning which potentially resolves these limitations
because deep learning is more sophisticated, and it can learn and make intelli-
gent decisions on its own. This thesis, therefore, introduces several deep learning
models such as Recurrent Neural Network (RNN), Long Short-Term Memory
(LSTM) neural network, and Convolutional Neural Network (CNN) for locomo-
tion mode recognition, including level ground walking (LW), standing (ST), and
stair ascent/stair descent (SA/SD) and evaluates their performance of them. The
results show that CNN and LSTM models outperform other models, and these
models are promising for applying locomotion mode recognition in real-time for
robotic prostheses. The proposed gait phase detection method was applied for the
AMP-foot, which is an active prosthesis in the Robotics & Multibody Mechanics
research group. In order to collect the necessary data to train the neural network
deep learning models, the thesis proposes an electronic system design for the mea-
surement unit, which mainly consists of a microcontroller board, one IMU sensor
attached to the lower shank, and two force sensors mounted under the foot. This
unit will measure the acceleration and velocity of the subjects when they walk in
different scenarios. Although the experiments are conducted in an offline setting,
due to the forecasting capabilities of the deep learning approaches, this system
provides an opportunity to eliminate detection delays for real-time applications
and the capability to reproduce the dynamic functions of the limbs of amputees.