EMG based motion intention detection for control of a robot-assisted training platform

Abstract

In order to improve the effectiveness of therapeutic training as well as to reduce the labour-intensive job of physiotherapist, robot-assisted training platforms have been developed as tools to assist the affected subjects in performing prescribed training tasks. It has been suggested that active participation from the subject in training session has a positive impact on the therapeutic training outcomes. As a response, past researchers have developed various assist-as-needed (AAN) control strategies for different robotic platforms to assist the subject in the case he was unable to complete the training task. Although the implementation of the AAN control strategy could promote the required assistance when needed, the implementation of the control strategy alone was found to be insufficient. In particular, the main problem with the current AAN control approaches was due to the absence of an important parameter that is a reliable triggering source from the subject in the form of intention to move the impaired limb. The ‘intention’ is paramount in improving the efficacy of interaction between the subject and the robotic-assisted training platform, as well as to ensure the subject safety. Therefore, the proposed study was attempted to address the issue, by developing an algorithm to predict the intention from electromyography (EMG) signal as well as to develop the adaptive assist-as-needed (AAAN) control strategy. It is worth noting that the study is novel as it evaluates the intention signal from the EMG signals within the range of 40 milliseconds to 100 milliseconds. The signal was acquired from a group of muscles (biceps) located at the upper arm when subjected to a flexion range of motion around the elbow joint along the sagittal plane. From the collected signal, time-domain analyses were implemented to extract the salient features of the signal. These features were then used to develop motion intention model in the form of k-Nearest Neighbour (k-NN) classifier. By leveraging the output of the classifier, a dedicated hybrid automata (HA) control framework was designed by integrating suitable impedance control scheme for different mode of motions namely passive, active and semi-active motions. A total of 30 able-bodied subjects have been recruited for the experiment to collect the signals upon attaining ethical clearance. It was demonstrated from the investigation that the coarse k-NN classifier was able to provide good classification of both the motion intention and pre-intention, preamble to the actual motion with the average training and test classification accuracy of 82.1% and 79.6%, respectively. From the hardware implementation, it was demonstrated that the proposed control strategy was able to provide the required assistance and resistance torques according to the different class of motion abilities that were categorized earlier. Based on the findings, it was evident that the proposed control strategy could provide intuitive and natural assist-as-needed torque input to the system based on one’s motion intention and ability.