脑电时间效应对情绪识别模型的影响及解决措施研究

Study on Time-Effect on EEG-Based Emotion Recognition Model and Countermeasures

近年来,越来越多的研究者投入到基于脑电的情绪识别研究中。然而在实际应用中,建立高精确度的情绪识别模型仍面临巨大的挑战,其中一个难点就是如何剔除或降低脑电信号的时间效应,进而提高情绪识别模型的时间鲁棒性。拟通过增加情绪模型中训练样本的天数,降低时间效应对识别模型的影响。利用视频诱发被试的正性、中性、负性3种情绪状态,共9名被试参与实验,每名被试需在1个月内进行5次数据采集,每次采集的时间间隔分别是1天、3天、1周和2周。采集被试60导联的脑电信号,并提取6个频段的功率谱特征。在模式识别阶段,分类器的训练样本分别来自N天的样本(N=1,2,3,4),剩余(5-N)天的数据则作为测试样本,得到不同训练天数下的分类正确率。结果表明:脑电时间效应的确会影响情绪识别的正确率,当训练集与测试集中的样本来源于不同的两天时,识别率显著下降(P<0.01);随着训练集样本天数的增加,正确率提高,正确率与训练样本的天数呈正相关;当训练集中样本来源于2~4天时,相比于1天的情况,平均正确率的提高率分别为6.45%(P=0.006)、10.48%(P=0.000)、14.40%(P=0.000),即增加训练集中样本的来源天数,能显著降低时间效应对分类效果的影响。结果证实,脑电时间效应能显著降低情绪识别模型的识别正确率,增加训练样本的天数可降低时间效应对识别模型的影响,并提高情绪识别模型的时间鲁棒性,从而为情绪模型从理论研究走向应用提供技术支持与研究思路。

There are numerous studies measuring the brain emotional status by analyzing EEGs under the emotional stimuli that have occurred,however,in practical application,an important but unresolved question is the extent to which the emotion model may generalize over time,since people could have a different expression of the same physiological signal on different days even when he experiences the same emotion. This paper attempted to add multiple days to the training set in purpose to weaken the impact of day-effect,and then to improve the generalization of the classifier. Eight subjects participated in this experiment,in which movie clips were presented to evoke the subjects＇ three emotional states of neutral,positive and negative. Moreover,EEG was recorded 5 times within one month for each subjects. Support vector machine（ SVM） was used to obtain the 3-class classification rates in all the collecting conditions including1-day collection,2-day collection,3-day collectionand 4-day collection. N-day collection represented the case in which data from N days were sent totrain the SVM and the remaining（ 5-N） days were used to form the testing set. Results showed that the accuracy was increased with the number of days in the training set for most of the subjects. Compared with 1-day collection,the increasing rates of the accuracies were 6. 45%（ P = 0. 006）,10. 48%（ P = 0. 000）,and14. 40%（ P = 0. 000） in 2-day,3-day and 4-day collections. These results suggested that adding data from more days to the training set could improve the performance and generalization of an emotion classifier. Though it is still a big challenge in EEG-based emotion recognition,these results provided a promising solution and take EEG-based model one step closer to being able to discriminate emotions in practical application.