Research On High-speed Railway Driver Fatigue Detection Technology Based On Deep Learning And Speech Features

In recent years,the pressure of high-speed railway transportation has been increasing,and the requirements for safety supervision of train operation have also been raised.Railway running is a complex system engineering.In the "man-machine-ring" system of railway operation,after a series of modern improvements in vehicles,driving systems and equipment,the functions and reliability of "machine" and "environment" become stronger and stronger,which makes "man" become the main constraint factor of train running safety.Among all the participants in railway transportation,train drivers play the most direct and prominent role.In addition to monitoring the information of train operation,train drivers are supposed to deal with various emergencies at any time.Therefore,it is of great significance to evaluate the fatigue condition of high-speed railway drivers timely and accurately,so as to ensure the safety of train operation.Considering that high-speed rail drivers need to use a lot of standards call-andanswer expressions in their daily work,and the voice features can reflect the fatigue state information of human body well,this paper studies and constructs the fatigue detection technology of high-speed rail drivers based on the voice features and deep neural network learning methods neural network learning method based on depth,which provides technical and theoretical support for higher level of human-computer cooperation.The main contents of this paper are as follows:(1)A high-speed rail driver fatigue detection voice database(High Speed rail driver Fatigue Corpus,HSRD-FC)was established.Based on the standard terms of call-andanswer for high-speed rail drivers,the phoneme fatigue sensitivity is studied.The Mel Frequency Cepstrum Coefficient features of different phonemes are used as input,and SVM as fatigue detection classifier to identify the fatigue sensitivity of each phoneme,which can determine the text design of high-speed rail driver fatigue detection voice database.Combined with the working characteristics of high-speed rail drivers,AX-CPT task is used to design a scheme to induce mental fatigue in laboratory environment.Through the multi-modal data analysis of subjective score,performance data and heart rate variability,the effectiveness of fatigue induction is verified,and the collected voice data is marked,so as to establish the voice database of fatigue detection for high-speed rail drivers.(2)A fatigue detection model based on deep learning was established,and lightweight improvements are made to the model.In this paper,a feature extraction scheme of spectrogram is constructed,which includes preprocessing processes such as speech denoising and endpoint detection.The basic network structure is preliminarily established by comparing the performance of different network structures.A series of super-parameters of the basic network structure are optimized to improve the performance of the basic network structure.In order to accelerate the training speed and reduce the model volume,the convolution layer is changed into Ghost Module layer.Experimental results indicate that the accuracy of the lightweight CNN model constructed in this part is 89.81% on the HSRD-FC data set.(3)Research on scaling and data augmentation of fatigue features were carried out.To further improve the detection performance,aiming at the input features,the influence mechanism of human fatigue on spectrogram was analyzed,which are optimized via using different acoustic scales.The over-fitting problem is solved effectively by augmenting the fatigued speech data.Experimental simulation results show that the accuracy of the fatigue detection scheme using Mayer spectrum is 91.36% on the test set,and the accuracy of the fatigue detection scheme using "Mel spectrum + data enhancement" is 91.98%,which can meet the needs of actual fatigue detection.