Research Of Human Mental State Recognition Methods Based On Deep Learning

People in a fast-paced environment gradually reveal more and more mental unhealthy problems,which are potential threats to both psychology and physiology,so it is urgent to use technology to identify and intervene in human mental states in time.Physiological signals are the most direct and objective expressions of human body in different mental states,and it is more reliable to build a human mental state recognition model based on physiological signals.This thesis is based on the current research to further refine the mental state and study the four classification recognition method of human mental state based on deep learning.The main work is as follows.(1)A multi-level dual-channel fusion model for human mental state recognition is designed to fuse deep mental state features extracted from CNN and LSTM dual-channel and shallow time-frequency domain features in a multi-level feature fusion,which makes up for the defects of various types of features and obtains more representational mental state features,thus improving recognition performance.In addition,this thesis also designs two methods of parallel splicing and composite network extraction in deep mental state feature extraction.The model solves the problem of poor recognition performance of a single model and improves the accuracy of mental state recognition compared with traditional machine learning methods or CNN or LSTM networks alone.(2)Mental state recognition models for SE-CNN channel attention and CBAM-CNN hybrid attention were designed.They obtain the time-frequency maps from ECG signals by S-transformation,and then input them to the attention module to calculate their feature weights,which avoids the problem of not equal importance when fusing individual feature vectors without overly increasing the complexity of the network,and obtains more influential feature vectors.The results demonstrate the effectiveness of attention mechanism for extracting physiological information and mental state recognition.(3)According to the internationally recognized method of mental state experiment acquisition,this thesis sets up a simulated flight driving experiment and acquires ECG signals under different mental states.It is used as the datastructure for model training and testing together with the public dataset to verify the generalization ability of the model designed in this thesis.In addition,seven human mental state recognition models based on five methods of traditional machine learning and two methods based on basic neural networks are built as a performance comparison of the models designed in this thesis.