Abstract:
In Brazil the number of lower limb ablations performed in the public health system
is prevalent among surgeries of this type, with transfemoral amputations having the
highest incidence. With the loss of a considerable part of a lower limb, body’s balance and
stability are drastically affected. The attempt to compensate the loss affects patients’ body
biomechanics and the deprivation of mobility directly influences its quality of life. Over
the years, efforts to enlarge transfemural amputee patients’ quality of life have grown and
numerous researches have been carried out, benefiting prosthetic development that tries
to compensate the limb loss or at least to mitigate the influences on body biomechanics.
Limb prostheses can be passive, semi-active and active, being the last mentioned the
most appropriate to provide an experience very close to the natural one to its users.
Magneto-rheological fluids are a class of intelligent materials, whose main characteristic is
to change from a liquid to a semi-solid state through the application of a magnetic field.
Its advantages make its use very promising in several implementations, such as in the
field of prosthetic devices. This study presents the development of a control system for a
magneto-rheological knee that aims to be subsequently applied in an active transfemoral
prosthesis. The system uses neural networks and predictive control techniques, that brings
the skills of learning, forecasting, self-organization, the ability to predict and guarantee
the behaviour of the system within the prediction horizon. Together these two techniques
enable a refinement of the control system. In order to assure its correct functioning, torque
and knee positioning controls are performed, which acting as sub-systems are responsible for
ensuring that the engine is able to provide sufficient torque to perform the movements and
the leg could be properly positioned during every gait cycle. The two sub-systems acting
simultaneously and steadily can guarantee that all system requirements are met. Different
network arrangements are tested using Long Short-Term Memory, Gated Recurrent Unit
and simple recurrent layers. The results showed that the magneto-rheological knee can
be controlled using the techniques presented, which demonstrates its application in real
prostheses is feasible and promising.