EEG情感识别中基于集成深度学习模型的多分析域特征融合

提出一种基于集成深度学习模型的情感状态检测方法.首先从脑电信号的时域、频域和时频域中提取4种表征情绪状态显著信息的初始特征;然后使用胶质细胞链改进的深度信念网络分别提取这些特征的高层抽象表示;最后利用判别式受限玻尔兹曼机对高层抽象特征进行融合,进行情感状态预测.在DEAP数据集上进行的实验显示,胶质链能够挖掘和利用EEG不同通道之间的相关性信息,而集成深度学习模型能够有效集成EEG信号在时域、频域和时频域蕴含的情感状态相关的显著性信息.

Multi-analysis domain feature fusion of EEG emotion recognition based on integrated deep learning model

An emotion recognition method based on the integrated deep learning model is presented. Firstly, four kinds of raw features are extracted from the time domain, frequency domain and time-frequency domain of electroence- phalogram(EEG) signals. Then, the high-level representations of these features are extracted respectively by using deep belief networks with glia chains. Finally, the high-level representations are fused by a discriminative restricted Boltzmann machine to implement emotion recognition task. Experiments are conducted on the DEAP dataset. The results show glia chains can mine inter-channel correlation information and the complementary models can integrate the four kinds of raw features effectively.