Evaluation of Activation Function Capability for Intent Recognition and Development of a Computerized Prosthetic Knee

Intent recognition is a basic requirement for computerized control of the prosthetic knee. Many scholars have used an ANN (Artificial Neural Network) and applied to a computerized prosthesis with good results. Determining an appropriate activation function in artificial neural networks is an essential issue. The main objective of this paper was to investigate the appropriate ANN activation function for intent recognition via accelerometer and gyroscope sensor data to develop a computerized prosthesis. The Feed-Forward Artificial Neural Networks (FFANN) with back-propagation learning method was used to recognize activity patterns. Efficiency of two activation functions were compared to choose an appropriate ANN activation function. Results indicate that log sigmoid function (LOGSIG) performs better than a tangent sigmoid function (TANSIG).