The Transfer Learning Method for Reducing the Influence of Electrode Shift

In systems based on pattern recognition using surface electromyography (sEMG) signals, there exists a problem of electrode shift during the process of donning and doffing. Directly using models trained on the source domain may result in limited prediction ability and a decrease in recognition accuracy. In this paper, a Domain-Adversarial Neural Networks(DANN) network is constructed based on CNN RES and the last layer of the feature extraction network and label predictor is fine-tuned. The experimental dataset used in the study was SeNic-main, proposed by the Shenyang Institute of Automation, Chinese Academy of Sciences. The experimental results were compared with the baseline network CNN_RES(CNN_ResNet), and eight tests were conducted on data with offsets of 10 degrees and -10 degrees. The results showed that the average testing accuracy was increased by 15.47% and 13.96% respectively, compared to CNN_RES.