Gesture Recognition of EMG Signals Based on Migration Learning

For the gesture recognition system of electromyographic signals(EMGs), the disadvantage of low generalization ability of recognition algorithms limits its use among different users due to factors such as differences in human structure, differences in the types of muscle fibers, and differences in the contact between sensors and muscles. In this paper, a transfer learning strategy is proposed to design and train a long-short-term recurrent neural network based on a trained model as the feature extractor of the target network, selecting different states of gestures to form the source task, and designing and training the long-short-term recurrent neural network as the myoelectric gesture signal source network model. Two types of migration learning strategies are designed, namely, migration in pre-training mode and migration strategy in fixed-value mode. Gesture recognition experiments were conducted on six different target gesture datasets under the two migration and non-migration strategies, respectively. The results show that the recognition effect of the migration learning strategy using the pre-training approach is better than that of the migration strategy using the fixed-value approach. For the improved Long Short Term Memory (LSTM) gesture recognition algorithm using the migration strategy approach, the proposed migration strategy reduces the training time by 16.45 times and achieves an accuracy of 80.2% when the gesture data training set is only 2% of the total dataset, which is 22% higher than that of the gesture recognition method without migration learning strategy.