Abstract: Aiming at channel redundancy caused by the use of manyelectrodes in EMG recognition, an optimal channel combination selection method based on thegradient boosting decision tree (GBDT) model is proposed. First, five features commonly used for each channel's EMG signal are manually extracted, and then the GBDT model is used to generate implicit new features.Second, the original features and new features under all channel combinations are combined, and another GBDT model is trained to predict the movement recognition rate of each combination.Finally, the optimal channel combination is selected for on-line control experiment. The experimental results show that the optimal channel combination has a high off-line recognition rate and on-line control accuracy and can achieve accurate and real-time control of the robot hand.When using the EMG recognition system to conduct on-line control experiments multiple times, selecting the optimal channel combination can reduce the number of electrodes to be attached, reduce the information redundancy and interference caused by redundant channels, and improve the practicability and robustness of the system.