Learning finger movement patterns for a transradial prosthesis using artificial neural networks

Abstract

This article presents an approach to achieving flexibility in a transradial prosthesis that allows the grasping of objects of different shapes through a hardware-implemented control architecture, enabling users to perform various activities of daily living. The proposed generalized hardware architecture utilizes an artificial neural network, facilitating the adjustment and interconnection between neurons, as well as providing adequate resolution to adapt the behavior to diverse finger movement patterns. To this end, distance sensors were incorporated into the prosthesis fingers to obtain information about the distance to objects. Servomotors were also used to manipulate the position of the fingers based on the data obtained from the sensors. A central composite design was used to train the network to identify finger movement patterns, generating appropriate combinations of independent variables (sensor data) and their association with their respective responses (motor movements). The main result of this proposal is that the assumption of the values assigned to the patterns is matched by the prosthesis through the gripping and holding of cylindrical, spherical and rectangular objects with an accuracy of 97.8%, a mean square error of 1.7042° and a response time of 0.5 seconds.