基于迁移学习的表面肌电信号手势识别

Surface electromyographic signal hand gesture recognition based on transfer learning

为了解决基于深度学习方法的表面肌电(sEMG)信号手势识别模型训练中需要大量数据的问题,该文将迁移学习方法与卷积神经网络(CNN)相结合,设计了一种微调模型参数的训练方法。设计实验对16名被试者采集了8种手势的sEMG信号。利用已有被试者的sEMG信号作为源域数据对CNN进行训练,模型在源域数据上的分类准确率为85.3%~98.11%。取其他被试者的sEMG信号作为目标域数据,数据量大小为源域数据的10%。在已训练好的模型基础上固定前2层卷积层,微调其余层级的参数,完成模型在不同个体间的迁移。迁移后的模型在目标域上的分类准确率为84.14%~97.4%。

In order to solve the problem that a large amount of data is needed in the training of hand gesture recognition model of surface electromyography(sEMG)signal based on depth learning method,a training method for fine-tuning the model parameters is designed combining the transfer learning method with the convolutional neural network(CNN).The sEMG signals of 8 gestures of 16 subjects are collected in this experiment.The sEMG signals of the existing subjects are used as the source domain data to train the CNN,and the classification accuracy of the model for the source domain data is 85.3%~98.11%.Taking the sEMG signals of other subjects as the target domain data with the size of 10%of the source domain data,the parameters of 2 front convolution layer is fixed,and the parameters of the later layers of the model are fine-tuned on the basis of the trained model,so that the model can be transferred among different individuals.The classification accuracy of the transferred model in the target domain is 84.14%~97.4%.