Abstract:
Mental workload can be viewed as the utilization rate of brain resources in a task, which reflects the mental state of humans in their work. The accurate recognition of mental workload can help the system adjust the operator’s work tasks, and can be applied to a series of complex human-computer interaction systems such as remote surgery, driving monitoring, equipment maintenance, and so on. In many physiological signals, near infrared spectroscopy (NIRS) is insensitive to electrical noise and has good resistance to motion artifacts. Moreover, the spatial resolution of NIRS is higher than that of electroencephalogram (EEG). Hence, NIRS can better reveal the relationship between local brain areas and mental workload. The proposed mental workload recognition model based on multi-channel NIRS is motivated from two aspects: (1) The traditional feature extraction process of physiological signals largely depends on prior knowledge; (2) Most deep learning models do not consider the relationship between channels. The connectivity matrices based on mutual information, phase-locked value, and Pearson correlation coefficient are calculated to reflect the internal relationship between channels. Then, the original NIRS signals and the connectivity matrix form a graph structure, which is sent to the graph convolution neural network for training and testing. The experimental results show that the graph convolution neural network has good ability to extract abstract features. The recognition accuracy of mental workload can be further improved with the consideration of the correlation coefficient between channels. Moreover, the visualization results of connectivity matrix show that the frontal lobe of the brain is sensitive to changes in mental workload.