Development of neural network based algorithm of active ankle prosthesis using gait analysis data

Abstract

Amputation is the removal of a part or all of a limb due to disease, accident or trauma and it has a large incidence rate. For example, in the United States, an average of 500 people loses at least one limb every day, approximately 65% of which is comprised of lower limb amputations. Since energetically active prostheses are costly, amputees usually continue with their daily lives using a wheelchair or a passive prosthesis. The aim of this study is to determine the optimum sensor needs for an active ankle prosthesis and to develop an algorithm suitable for this sensor infrastructure. In the long run, design of a device that is both easy to use and nancially feasible is aimed at and the present work is central to that aim. In this context, three neural networks structures with di erent inputs were developed to facilitate ankle movement and their performances were evaluated. The results show that if a device in which only EMG signals are to be used as network inputs, a total of 5 signals should be collected from di erent muscles that are responsible for hip, knee and ankle movements. The results also show that, if the use of a smaller number of EMG sensors is preferred, it is necessary to incorporate also a force or torque feedback into the system. In such application, three EMG signals collected from tibialis anterior, soleus and gastrocnemius medialis muscles were shown to su ce. These ndings shed an important light to our understanding of the number and kind of sensor inputs necessary for an active ankle prosthesis requirements of which can be variable depending on the amputation level of the patient and the mechanical design exibility.|Keywords : Electromyography, EMG, Neural Network, Active Ankle Prosthesis.