基于迁移学习多层级融合的运动想象EEG辨识算法

EEG identification algorithm of motor imagination basedon multi-level fusion of transfer learning

为了准确获取运动想象脑电信号的全局特征和个体间的共性特征,进而提高其分类准确率和模型鲁棒性,提出一种参数共享迁移学习的融合卷积神经网络算法。首先把源域上训练完成的网络逐层迁移至目标网络以获取最佳迁移层。其次,在迁移层后分别连接不同数量的卷积-池化块构成4个不同深度的卷积网络,并将其并行融合后连接分类器得到分类结果。利用BCI竞赛IV Datasets 2a对提出方法进行实验分析。结果显示,使用100%和50%样本时所有受试者的平均辨识率分别为80.85%和78.9%,验证了提出方法在全局特征提取上的有效性小样本问题上的优势。

In order to accurately obtain the global characteristics of motor imaging EEG signals and the common characteristics between individuals,and then improve its classification accuracy and model robustness,a fusion convolutional neural network algorithm with parameter sharing transfer learning is proposed.First,the trained model on the source domain is migrated layer by layer to the target network to obtain the best migration layers.Secondly,after the migration layers,different numbers of convolution-pooling blocks are connected to form four convolutional networks with different depths,and they are merged in parallel and finally the classification results are obtained through the classifier.Use the BCI competition IV Datasets 2 a to conduct experimental analysis on the proposed method.The results show that the average recognition rate of all subjects when using 100%and 50%samples is 80.85%and 78.9%,respectively,which verifies the effectiveness of the proposed method on global feature extraction and the advantages of small sample problems.