使用三維/四維卷積神經網路的腦電波系統用於預測駕駛表現

Abstract

汽車事故一直是全球死亡的主要原因。大多數事故是由於誤操作和行為失誤造成的，主要是由於嗜睡和疲勞造成。換句話說，預測駕駛者的認知狀態甚至監控駕駛時的反應時間是具有淺力的研究。為了預測人腦的狀態，腦電波圖已經被證明是監測人類大腦和人類行為狀態的指標。在EEG信號分析中，因為自動特徵擷取，機器學習演算法成為最常使用的方法。另一方面，深度學習演算法在許多研究中，已經能提高了許多領域機器學習的準確性。有些研究也已經在腦電信號分析中採用了深度學習算法。然而，EEG信號的分析是相當富有難度的，因為腦電信號相較於其他信號處理是相對複雜。在傳統的深度學習演算法中，EEG信號的淺在信息不能被有效地提取分析。例如，EEG信號分析的時間關聯性和EEG通道間空間關聯性是不可或缺的，但在當前的演算法中卻不能有效地提取。為了解決這個問題，本研究應用了一種基於卷積神經網路(CNN)的算法，三維卷積神經網路(3D CNN)來解決時間關聯性的問題。換句話說，透過應用3D CNN在腦電波信號分析，該模塊可以有效地提取EGG信號中的時間關聯性訊息以及頻率訊息。此外，該研究還提出了一種最新的深度學習演算法四維卷積網路(4D CNN)，4D CNN也是一種基於CNN的算法。在EEG信號的特徵提取方面，4D CNN能夠透過在特徵提取中執行四個卷積來將不同的屬性，如：頻率，時間信息和EEG通道的空間訊息結合在一起。通過這樣做，此模組可以在認知狀態監測系統中精確地映射EEG-RT關係。這項工作的貢獻是根據大腦研究知識，對腦電信號分析中的深度學習方法進行改進。

Motor vehicle accidents have been the leading cause of fatalities worldwide. Most of those accidents are due to mistaken operation and behavioral lapses and majorly are caused by drowsiness and fatigue. In other words, predicting the cognitive state of drivers or even monitoring the reaction time (RT) while driving is a potential research. To predict activity of human brain, electroencephalogram (EEG) has been proved be an indicator of monitoring the state of human brain and human behavior. In EEG signal analysis, the most adopted method is machine learning since the strength of auto feature extraction. On the other hand, deep leaning algorithm has been shown improving the accuracy of machine learning in many fields. Some works have adopted deep learning algorithm in EEG signals analysis already. However, the analysis of EEG signals can be extremely challenge, since the signals is relatively complicated in signal processing. In traditional deep learning method, the essential information of EEG signals cannot be considered into analysis efficiently. For example, temporal information and the spatial information of EEG channels are indispensable in EEG signal analysis but cannot be extracted well in current methods. In order to solve this problem, this study applied a proposed convolutional neural network based (CNN-based) algorithm, 3D convolutional neural network (3D CNN) to solve the temporal information. In other words, by applying 3D CNN on EEG signal analysis, the module can extract temporal information, frequency information in EGG channels. Furthermore, the study also proposed a novel deep learning method 4D convolutional neural network (4D CNN), which is also a CNN-based algorithm. In terms of feature extraction of EEG signals, 4D CNN is able to combined different attributes, frequency, temporal information and spatial information of EEG channels together by performing four convolutions in feature extraction. By doing so the model can map EEG-RT relationship precisely in the cognitive state monitoring system. The contribution of this work is improving deep learning method in the analysis of EEG signals according to the knowledge of brain researches.