Application of Machine Learning-Based EEG Decoding Algorithms in State Recognition for Railway Participants
基於機器學習的腦電解碼算法在鐵路人員狀態識別中的應用研究
Student thesis: Doctoral Thesis
Author(s)
Related Research Unit(s)
Detail(s)
Awarding Institution | |
---|---|
Supervisors/Advisors |
|
Award date | 25 Jun 2024 |
Link(s)
Permanent Link | https://scholars.cityu.edu.hk/en/theses/theses(2a5734c9-da29-4024-9ceb-4fc863d8c8a0).html |
---|---|
Other link(s) | Links |
Abstract
In the ongoing advancement of railway systems, ensuring safety and improving passenger experiences are of great importance. In addition to sophisticated equipment, research related to human factors plays a crucial role, as precise identification of human states can significantly improve the safety and comfort of railway systems. This thesis delves into the application of machine learning algorithms for the decoding of electroencephalogram (EEG) signals to recognize various states of individuals within railway settings. The study introduces an ensemble learning EEG decoding algorithm for few-channel EEG, a multitask learning EEG decoding algorithm for the extraction of intrinsic task information, an end-to-end deep learning EEG decoding algorithm to analyze the original EEG, and a spatial-temporal deep learning EEG decoding algorithm to decode the source signals of EEG. The research encompasses the detection of driver distractions and fatigue, the evaluation of passenger comfort, the estimation of the mental workload of workers, and the recognition of emotions among all participants.
In the first part, we studied the issue of detecting fatigue and distraction of train drivers based on frontal EEG features. Driver fatigue and distraction were the main factors in railway accidents. We conducted experiments, established a novel EEG dataset, and extracted features such as energy, entropy, and frontal asymmetry ratios. An ensemble learning method named RLX was proposed which demonstrated superior performance compared to conventional machine learning algorithms like SVM and KNN. Furthermore, we proposed a temporal voting algorithm capable of enhancing performance by aggregating classification probabilities over time without the need to retrain the base model. This algorithm named RLX-TV successfully mined frontal EEG information, completing the driver state detection task and marking a notable advancement in real-time, nonintrusive monitoring methods for railway safety.
In the second part, we study the issue of evaluating the comfort of railway passengers based on EEG features. Passenger comfort was influenced by various factors, such as emotions, vibrations, and noise. The ability to mine the intrinsic information of multiple factors can improve the evaluation of passenger comfort. We proposed EEG-DEMTL, a multitask learning network based on evolutionary computation principles, to evaluate comfort through EEG signals. The network structure allowed the exploration of correlations between multiple factors and overall comfort, with a differential evolution algorithm improving performance. We designed field experiments and collected data that showed that, compared to traditional methods, the multitask learning structure significantly enhanced performance.
In the third part, we study the issue of evaluating the mental workload of railway workers based on the original EEG signal. The methods mentioned earlier are based on manually extracted EEG features, which have limited applicability. In this section, we proposed an end-to-end brain-computer interface algorithm called EEG-TNet. This system combined automatic pre-processing with a specialized neural network, achieving a highly accurate solution with no manual intervention required, capable of accurately estimating mental workload. Both cross-subject and within-subject experiments validated the effectiveness of this model.
In the fourth part, we study the issue of recognizing the emotions of railway participants based on EEG source signals. Due to the volume conduction effect, the channel locations of the EEG signals could not accurately correspond to the cortical locations of neural discharges. In this section, we used the WMNE method to reconstruct EEG sources and constructed CT-EEGNet, an end-to-end network that avoids traditional feature extraction. This network combined convolution and the Transformer encoder, extracting temporal and spatial information from EEG source signals, and outperformed existing models in accuracy. The results in the HR-EEG4EMO data set demonstrated the robustness of this method.
In summary, this thesis makes significant contributions to the application of machine learning in railway systems, providing innovative solutions for state recognition through EEG decoding. The proposed methods are expected to improve the safety and comfort of railway systems, paving the way for future developments in this field.
In the first part, we studied the issue of detecting fatigue and distraction of train drivers based on frontal EEG features. Driver fatigue and distraction were the main factors in railway accidents. We conducted experiments, established a novel EEG dataset, and extracted features such as energy, entropy, and frontal asymmetry ratios. An ensemble learning method named RLX was proposed which demonstrated superior performance compared to conventional machine learning algorithms like SVM and KNN. Furthermore, we proposed a temporal voting algorithm capable of enhancing performance by aggregating classification probabilities over time without the need to retrain the base model. This algorithm named RLX-TV successfully mined frontal EEG information, completing the driver state detection task and marking a notable advancement in real-time, nonintrusive monitoring methods for railway safety.
In the second part, we study the issue of evaluating the comfort of railway passengers based on EEG features. Passenger comfort was influenced by various factors, such as emotions, vibrations, and noise. The ability to mine the intrinsic information of multiple factors can improve the evaluation of passenger comfort. We proposed EEG-DEMTL, a multitask learning network based on evolutionary computation principles, to evaluate comfort through EEG signals. The network structure allowed the exploration of correlations between multiple factors and overall comfort, with a differential evolution algorithm improving performance. We designed field experiments and collected data that showed that, compared to traditional methods, the multitask learning structure significantly enhanced performance.
In the third part, we study the issue of evaluating the mental workload of railway workers based on the original EEG signal. The methods mentioned earlier are based on manually extracted EEG features, which have limited applicability. In this section, we proposed an end-to-end brain-computer interface algorithm called EEG-TNet. This system combined automatic pre-processing with a specialized neural network, achieving a highly accurate solution with no manual intervention required, capable of accurately estimating mental workload. Both cross-subject and within-subject experiments validated the effectiveness of this model.
In the fourth part, we study the issue of recognizing the emotions of railway participants based on EEG source signals. Due to the volume conduction effect, the channel locations of the EEG signals could not accurately correspond to the cortical locations of neural discharges. In this section, we used the WMNE method to reconstruct EEG sources and constructed CT-EEGNet, an end-to-end network that avoids traditional feature extraction. This network combined convolution and the Transformer encoder, extracting temporal and spatial information from EEG source signals, and outperformed existing models in accuracy. The results in the HR-EEG4EMO data set demonstrated the robustness of this method.
In summary, this thesis makes significant contributions to the application of machine learning in railway systems, providing innovative solutions for state recognition through EEG decoding. The proposed methods are expected to improve the safety and comfort of railway systems, paving the way for future developments in this field.