Lower Limb Activity Recognition using sEMG Signals via Weighted Random Forest

Identifying the motion intention of the wearer and integrating it into the lower limb rehabilitation device is a rehabilitation interaction method that can effectively improve the rehabilitation effect. Existing surface electromyography (sEMG) based intention identification methods have the shortcomings of limited generalization ability in practical applications. In this paper, we proposed a weighted random forest (WRF) algorithm for the intention prediction. The performance of the ensemble model is improved by weighted ensemble of each base prediction decision tree. Detailed comparative studies between proposed method and conventional methods have been carried out through the lower limb activity identification data at different time periods. Compared with Linear Discriminant Analysis (LDA), Gradient Boosting (GB), AdaBoost, K-Nearest Neighbor (KNN), Random Forest (RF), the generalization ability of WRF is significantly superior considering the best accuracy and the best precision achieved by WRF. Compared to standard RF, the WRF improved classification accuracy performance by 1.19% and the classification precision accuracy performance by 1.1% on average. Conclusively, the proposed method can provide satisfactory generalization performance.