Abstract:Surface electromyography (sEMG) occurs before the action. When the action is active, its willingness can be predicted in advance. However, traditional classification methods usually face problems that realtime and accuracy are difficult to be compatible. To make the EMG signal betterapplied tothe machineand equipment, this paperproposesa fast action classification methodfortheLMSrandom forest EMGsignal. It canclassify andidentifyknee bend, hip bend,knee bend,kneebendandknee stretch.Compared with thetraditionalclassification algorithm, this study only needs to collect the data before 120 ms for classification. LMS is used to filter and assign corresponding weight to the original data. Its weight represents the importance of data features. In this way, the classification of traditional surface EMG signals can be improved. The lack of realtime performance provides a solution for the integration of human and exoskeleton devices. Compared with the traditional support vector machine, back propagation neural network and other algorithms, experimental results show that the proposed algorithm takes less time and the speed is 78 times that of the linear fence method. It has high accuracy and stability, and the recognition accuracy is 973%.