EEG-based Mental Workload Estimation using Encoder-Decoder Networks with Multilevel Feature Fusion

In this paper, we propose a model that combines the multilevel feature fusion algorithm and encoder-decoder structure for evaluation of mental workload using electroencephalogram (EEG) signals. The encoder-decoder structure was used to reduce additive noise and subject variations of EEG data. The encoder is structured by incorporating a 3D convolutional neural network (3DCNN) and multilevel feature fusion concept, which extracts unified key features from combining the low-level and high-level features. The decoder consists of simple 3DCNN layers to recover the input EEG image from the latent vector. The proposed model can achieve higher performance by mitigating feature variations. We evaluate our network with EEG data obtained through the Sternberg task to estimate mental workload, which has 91.6% accuracy and outperforms the conventional algorithm.