MEDA-CBLSTM: Data Acquisition, Processing and Emotion Recognition Based on Multi-Layer EDA

Using wearable devices to extract physiological information to identify and classify human emotion is a common research problem in human-computer interaction. From a computational point of view, emotion is difficult to quantify, estimate, and understand. However, indications of human cognitive and emotional related processes affect a variety of derived changes in many human psychological signals. These signals can help to build a quantitative model of emotion. We design and build a set of wearable hardware to acquisit, extract and process Multi-layer Electrodermal activity(MEDA) physiological signal data,and bulid neural network models based on the MEDA data. We build the dataset useing videos that can trigger happiness, sadness, anger, etc. For the annotated and sorted data set, we build a multilayer neural network model to perform feature extraction, mapping, and recognition from MEDA to emotions. Ablation experiments and results comparison are carried out for each model. Experiments show the hybrid model: end-to-end convolutional neural network with bidirectional long short-term memory network(MEDA-CBLSTM) achieves the best result. The model trained by the train set of testers achieves an average accuracy of 92.31% in their corresponding test set, but the accuracy reduces when it is tested on the other test sets. Therefore, we get the conclusion:the EDA emotion data has the characteristics of highly subject dependence.