A Fault-Tolerant Algorithm to Enhance Generalization of EMG-based Pattern Recognition for Lower Limb Movement

Real-time application of exoskeleton remains a challenge due to the limited stability of electromyogram (EMG) collected on lower limb. To enhance generalization of EMG based pattern recognition (PR), this work proposed a novel fault-tolerant algorithm based on conventional liner discriminant analysis (LDA). Based on two patterns of EMG collection (static data and dynamic data), the most accurate feature set was first selected in static data to guarantee basic performance of LDA. Detections of LDA formed a decision stream, and provided later analysis with information of movement and muscle activity of lower limb. A BP classifier was trained based on the decision stream and features set to recover misclassification of LDA when interference occurred. Both LDA and BP were trained on static data. Dynamic data were set to imitate instability of EMG in realistic application such as different force intensity and muscle fatigue. The proposed algorithm was tested on a EMG based PR system. 4 able-bodied subjects participated in the experiment. The enhancement in classification accuracy ranged from 7.52% to 13.94% over subjects compared with unoptimized system. After EMG channels reduced, this enhancement ranged from 9.01% to 27.35%, which significantly improved classification performance under limited condition. These results suggested that the proposed algorithm yield excellent performance in improving the generalization of EMG based PR.