基于自注意力的胶囊网络脑工作负荷分类

Mental Workload Classification Based on Self-Attention Capsule Network

脑工作量是为执行任务而支付的脑力资源的数量,评价工作任务引起的脑工作负荷水平对于优化操作员的负荷分担至关重要.本文提出一种基于自注意力的胶囊神经网络脑工作负荷检测方法:使用脑电图频谱特征和空域特征的融合来完成工作负荷评估.模型主要由卷积层、自注意力层、主胶囊层和脑工作负荷胶囊层构成,对7名受试者执行的脑工作负荷任务中的3个工作负荷水平分别提取了频带能量和功能连接性特征,并进行了融合,利用卷积层捕获融合特征的深度特征,随后使用自注意力机制进一步挖掘融合特征中大脑区域间传导的规律以及不同区域的频带能量变化(即其空间维度信息的转变过程),从而学习到不同状态下脑电特征中隐含的异常信息模式,同时利用胶囊网络在提取局部特征的同时关注全局结构信息.本文方法对于脑工作负荷评估的分类准确率为87.72%,且对高等级的脑工作负荷识别率更高.

Brain workload is the amount of mental resources given to perform a task,and evaluating the level of brain workload caused by work tasks is crucial to optimizing the operator's load sharing.In this paper,a method of brain workload detection based on self-attention capsule neural network is proposed,which uses the fusion of spectral and spatial features of Electroencephalogram(EEG)to evaluate the workload.The model is mainly composed of a convolutional layer,a self-attention layer,the main capsule layer and the brain workload capsule layer.The frequency band energy and functional connectivity features were extracted respectively from the 3 work load levels of the brain workload tasks performed by 7 subjects,and fusion was carried out.The depth features of the fusion features were captured by the convolutional layer,and then the self-attention mechanism was used to further explore the conduction rules between brain regions and the frequency band energy changes of different regions in the fusion features,that is,the transformation process of its spatial dimension information,to learn the abnormal information patterns hidden in EEG features under different states,and at the same time,the capsule network is used to extract local features and pay attention to global structural information.The classification accuracy of this method for brain workload assessment was 87.72%,and the recognition rate of high-level brain workload was higher.