EEG And EOG-based Multimodal Vigilance Estimation Using Deep Learning Method

Physiological signals-based vigilance estimation aims to solve the problem of frequent traffic accidents in Public Transportation Safety(PTS)due to the lower level of driver vigilance,how to ensure the accuracy and efficiency of predicting the level of alertness of drivers in the driving scene is the technical bottleneck in this field.This paper focuses on the optimization method of feature dimensionality reduction and reconstruction under different conditions of single-modal and multi-modal,how the learning method can has better approximation ability and feature learning at the same time,how to improve the training speed of the learning method and enable the external control equipment to identify the level of human vigilance more accurately,etc.A series of multi-mode deep network models based on the combination of physiological characteristics and Deep Learning(DL)are proposed.The comprehensive efficiency and prediction accuracy have reached the international frontier or leading level,The main contributions are as follows:(1)In order to solve the problem that the existing multi-layer network model is only used for classification applications,a single-mode autoencoder deep network model(DAESN)is proposed.Besides the general functions of multi-layer network model,the additional functions of feature reduction and signal reconstruction are realized.At the same time,compared with other existing dimensionality reduction methods,this model has better generalization performance.In particular,the current weight of the encoding layer is replaced by the previous decoding layer,and is very related to the input data,and effective features can also be extracted for pattern recognition.Therefore,the training time of this model is ten times or even dozens of times faster than other compared methods,and the driver's vigilance level can be quickly detected.Finally,the root mean square error(RMSE)and correlation coefficient(COR)index of the continuous output are analyzed.The experimental results show that the feasibility and superiority of this method are verified.(2)The neural network-based learning method is very sensitive to parameter combinations,and it is difficult to quickly select the optimal parameters.In response to this problem,a two-layer neural network model with subnetwork nodes(DNNSN)is proposed,the model can be randomly chosen at the beginning of the training parameters,and will not affect the generalization performance in the learning process.The performance of the model is not sensitive to the parameters,which solves the above-mentioned problems.At the same time,the single-layer architecture for the Extreme Learning Machine(ELM)may lack effectiveness when applied to natural signals.This model has good approximation capabilities while also having the ability to learn features,greatly enhance the practicability of the method.In addition,after setting the optimal learning rate at the entrance layer of the model,its performance is significantly improved after the model is reasonably converged,which verifies the effectiveness of the method.(3)To solve the problems of multiple steps and long training time in the iterative process of traditional deep learning network(DL),a multi-channel autoencoder multi-modal deep network model for vigilance estimation is proposed.The model includes a small number of subnet neurons,which have the abilities of feature selection,representation learning,feature reduction and signal reconstruction,and can effectively solve the problem that highdimensional input features affect the training speed.At the same time,the hidden layer of the model can be calculated with only four steps of replacement technology,without a tedious iterative process,which greatly reduces the time cost,thereby improving the efficiency of model learning.In addition,multiple batches of randomly generated data can be input as a multi-channel auto-encoder at the same time.Through the process of subspace feature dimensionality reduction,extraction and fusion,the running speed of the model can be improved by a factor of ten.Finally,this model uses analysis of variance(ANOVA)to evaluate the statistical significance of the final experimental results.Experimental research shows that the two pooling fusion strategies used in the multi-channel deep learning model have obtained good experimental results,surpassing Support Vector Regression(SVR)and Long Short-Term Memory(LSTM)and other state-of-the-art methods.The accuracy of the method is verified by the good recognition of EEG and ECG related activities related to the level of vigilance.Based on methods such as physiological feature extraction and deep learning,this paper proposed a variety of effective models,which are expected to provide a new feasible way to quickly and effectively estimate the driver's vigilance level,which has important practical significance for reducing the rate of traffic accidents.