脑电时间效应对情绪识别模型的影响及解决措施研究

发布时间：2019-04-03 08:27

【摘要】：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(P0.01);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。
[Abstract]:In recent years, more and more researchers are involved in EEG-based emotion recognition. However, in practical applications, the establishment of high-precision emotion recognition model is still facing a great challenge, one of the difficulties is how to eliminate or reduce the time effect of EEG signals, so as to improve the time robustness of emotion recognition models. By increasing the number of days of training samples in the emotional model, the effect of time effect on the recognition model is reduced. The positive, neutral and negative emotional states of the subjects were induced by video. Nine subjects were involved in the experiment. Each participant was required to collect data five times within one month, and the interval of each data acquisition was 1 day, 3 days, 1 week and 2 weeks respectively. The 60-lead EEG signals were collected and the power spectrum features of 6 bands were extracted. In the stage of pattern recognition, the training samples of the classifier are from the samples of N days (N ~ (1), (2), (3), (4), and the data of the remaining (5 ~ N) days are used as the test samples, and the correct rate of classification is obtained under different training days. The results showed that EEG time effect did affect the correct rate of emotion recognition. When the samples in training set and test set came from different samples for two days, the recognition rate decreased significantly (P0.01). With the increase of the number of days in the training set, the correct rate increases, and the correct rate is positively correlated with the number of days in the training sample. When the training samples were from 2 to 4 days, the increase rates of the average correct rate were 6.45% (P = 0.006), 10.48% (P = 0.000), 14.40% (P = 0.000), respectively, compared with the case of 1 day. That is to say, increasing the source days of training samples can significantly reduce the impact of time effect on the classification effect. The results show that EEG time effect can significantly reduce the recognition accuracy of emotion recognition model, increase the number of days of training samples can reduce the impact of time effect on the recognition model, and improve the time robustness of emotion recognition model. So as to provide technical support and research ideas for emotional model from theoretical research to application.
【作者单位】：天津大学精密仪器与光电子工程学院生物医学工程系;
【基金】：国家自然科学基金重大研究计划(91520205);国家自然科学基金(81571762,31500865)
【分类号】：R318

【相似文献】