A Siamese Network Fusion Time-Frequency Domain Features for Mental Workload Level Identification

Abstract

Mental workload level can reflect subjects’ personal ability. In addition, continuous high level of mental workload can reduce subjects’ performance level, so it is necessary to detect subjects’ mental workload level in the time. In this paper, we propose a CNN model for time-frequency analysis based on Siamese networks (Siamese-EEGNet), in which the original Electroencephalogram (EEG) signal and the Power Spectral Density (PSD) of the signal are used as model inputs, and the features of the signal are extracted layer by layer through convolutional layers. Using P3 as a measure, the model is pretrained on a large volume data set using transfer learning, and successfully transfer to a smaller volume data set with mental workload level by fine-tuning the model parameters. SiameseEEGNet is able to consider both time domain and frequency domain information in the data, which is suitable for EEG structure characteristics. In practical completion, it can detect the mental workload level of subjects, measure their individual ability and improve their performance level.