LAUSR

Under the ideal condition, traditional myoelectric pattern recognition systems can achieve superior performance in recognizing motion intents. However, in practical applications, electrode shift is inevitable during the rewearing of the electrodes. The data variation caused by the shift can lead to dramatic performance degradation. In the literature, deep learning, especially convolutional neural networks (CNNs), proved effective in solving this issue. However, the aliasing effect inherent in the downsampling layer makes the common CNN shift vulnerable. The small input translations caused by the electrode shift can change the output dramatically, leading to incorrect classification. In this article, we propose a shift-robust CNN (SR-CNN) that replaces the downsampling layer with the ensemble of an anti-aliasing filter and adaptive polyphase sampling (APS) module. The anti-aliasing filter first counters the high-frequency components that can lead to unreasonable pulses during downsampling. The APS module then subsamples feature maps from several potential components adaptively, reserving the critical and stable information when the shift occurs. Experiments on two high-density surface electromyography (HD-sEMG) datasets show that the proposed SR-CNN method outperforms the state-of-the-art baselines. This study provides a promising solution to realize robust myoelectric control against the electrode shift.