A strain gauge based locomotion mode recognition method using convolutional neural network

Abstract

Locomotion mode recognition can contribute to precise control of active lower-limb prostheses in different environments. In this paper, we propose a novel locomotion mode recognition method based on convolutional neural network and strain gauge signals. The strain gauge only provides one-dimensional signals and is also used in the control strategy of the robotic prosthesis. The convolutional neural network takes the raw noisy signals as inputs. Three transtibial amputee subjects were recruited in the experiments, and three locomotion modes were recognized. The overall three-class locomotion mode recognition accuracy is 92.06±1.34% in the hold-out test and 92.53±1.61% in the 5-fold cross-validation. The results show that the strain gauge contains information of locomotion modes, and the convolutional neural network has the capacity of extracting features from raw signals.