Abstract: Abstract: To address the issues of excessive reliance on manual feature extraction in cognitive load recognition models, the neglect of spatial information in ElectroEncephaloGram (EEG) signals, and the inability to effectively learn graph structure data, a cognitive load EEG recognition model integrating Variational graph autoencoder and Local-Global Graph Network (VLGGNet) was proposed. The model VLGGNet consisted of two parts: a temporal learning module and a graph learning module. Firstly, the temporal learning module was employed to capture the dynamic frequency representation of EEG signals using multi-scale temporal convolution, and features were fused through spatial and channel reconstruction convolution and 1×1 convolutional kernel cascade modules. Then, the graph learning module was employed to define the EEG data as a local-global graph, where the local graph feature extraction layer aggregated the node attributes into a low-dimensional vector, and the global graph feature extraction layer reconstructed the graph structure using a variational graph autoencoder. Finally, the lightweight graph convolution operations were performed on the global graph and node feature vectors, and the prediction results were output by the fully connected layer. Through nested cross-validation, the average accuracy and F1 score of VLGGNet are improved by 4.07 and 3.86 percentage points, respectively, compared with the sub-optimal Local-Global Graph Network (LGGNet) on the Mental Arithmetic Task (MAT) dataset. Compared with the best-performing multi-scale Temporal-Spatial convolutional neural network (TSception) on the Simultaneous Task EEG Workload (STEW) dataset, the average accuracy of VLGGNet is the same as that of TSception, and the average F1 score is reduced by 0.01 percentage points. VLGGNet improves the cognitive load classification performance and verifies that the prefrontal and frontal regions are closely related to cognitive load status.

Key words: cognitive load, ElectroEncephaloGram (EEG), signals, multi-scale temporal convolution, variational graph autoencoder, Local-Global Graph Network (LGGNet)

摘要： 针对认知负荷识别模型存在过于依赖手动提取特征、忽视脑电信号的空间信息以及无法有效学习图结构数据的问题，提出一种融合变分图自编码器与局部-全局图网络（VLGGNet）的认知负荷脑电识别模型。该模型由时间学习模块和图形学习模块两部分组成。首先，时间学习模块采用多尺度时间卷积捕捉脑电图（EEG）信号的动态频率表示，并通过空间与通道重建卷积和 1×1 卷积核级联模块融合特征。其次，图形学习模块将 EEG 数据定义为局部-全局图，局部图特征提取层将节点属性聚合到一个低维向量，全局图特征提取层通过变分图自编码器重构图结构。最后，对全局图和节点特征向量执行轻量化图卷积操作，由全连接层输出预测结果。通过嵌套交叉验证，在心算任务（MAT）数据集上，相较于次优的局部-全局图网络（LGGNet），VLGGNet 的平均准确率和 F1 分数分别提升了 4.07 和 3.86 个百分点。在同时任务 EEG 工作量（STEW）数据集上，相较于表现最好的多尺度时空卷积神经网络（TSception），VLGGNet 的平均准确率与 TSception 相同，平均 F1 分数降低了 0.01 个百分点。VLGGNet 提高了认知负荷分类性能，同时验证了前额叶和额叶区域与认知负荷状态密切相关。

关键词: 认知负荷, 脑电信号, 多尺度时间卷积, 变分图自编码器, 局部-全局图网络