基于单模态生理信号无监督特征学习的驾驶压力识别

江润强; 陈兰岚; 谌鈫

doi:10.14135/j.cnki.1006-3080.20200728001

基于单模态生理信号无监督特征学习的驾驶压力识别

doi: 10.14135/j.cnki.1006-3080.20200728001

华东理工大学化工过程先进控制和优化技术教育部重点实验室，上海 200237

基金项目: 国家自然科学基金（61976091）；中央高校基本科研业务费专项资金

详细信息

作者简介:
江润强（1994—），男，山东淄博人，硕士生，主要研究方向为机器学习及其应用。E-mail：rqjiang_ecust@163.com

通讯作者:
陈兰岚，E-mail：llchen@ecust.edu.cn

中图分类号: TP391
计量
- 文章访问数: 605
- HTML全文浏览量: 407
- PDF下载量: 12
- 被引次数: 0
出版历程
- 收稿日期: 2020-07-28
- 网络出版日期: 2020-12-16

Driving Stress Detection Based on Unsupervised Feature Learning of Single Module Physiological Signal

Key Laboratory of Advanced Control and Optimization for Chemical Processes, Ministry of Education, East China University of Science and Technology,Shanghai 200237, China

摘要

摘要: 针对基于多模态生理信号分析的驾驶压力识别会影响驾驶员的行车舒适性，且传统的生理特征的提取需要依赖先验知识的问题，构建了基于单模态生理信号无监督特征学习的驾驶压力识别模型。首先采用单模态生理信号，通过构造卷积自编码器进行无监督的特征学习来提取抽象特征；然后将特征依次送入支持向量机、随机森林、K最近邻、梯度提升决策树4种不同的基分类器进行建模；最后引入集成学习的思想对不同基分类器的输出进行投票集成来提高驾驶压力识别的稳定性与准确性。实验结果表明，该模型在MIT-drivedb数据集的驾驶压力三分类任务中的准确率可达92.8%。
- 驾驶压力识别 /
- 单模态生理信号 /
- 卷积自编码器 /
- 抽象特征 /
- 集成学习
Abstract: The traffic accidents are mainly related to unfavorable driving status, e. g, fatigue, stress, distraction and sleepiness, resulting in a considerable amount of vehicle collisions and casualties every year. Commonly, stress can be viewed as a normal response of human body to dangerous or difficult events. With the advance of wearable sensor and wireless technique, researchers are paying increasing amount of attention to physiological measures, which have been proved to be highly correlated to driver’s mental states. In-vehicle intelligent systems could utilize these physiological measures automatically in various manners to help drivers better manage their negative driving status. So we focus on the research of physiological measures in this study. Commonly, driving stress detection based on multimodal physiological signals will affect the driving comfort of drivers and traditional physiological feature extraction techniques largely rely on prior knowledge. In this paper, single module physiological signal, i.e., GSR signal on the foot (FGSR) is used and abstract features are generated by unsupervised feature learning using 1D-convolutional auto-encoder (CAE). Then, the features are sent to four different base classifiers, namely k-nearest neighbor (KNN), gradient boosting decision tree (GBDT),support vector machine (SVM) and random forest (RF).Finally, the outputs of different base classifiers are integrated with voting method to improve the stability and accuracy of driving stress detection. The proposed model is validated on the MIT-drivedb data set, which shows that the feature extracted from GSR using convolutional auto-encoder has good representational ability for the driving stress and the ensemble of different base classifiers can improve the accuracy of final recognition results.
- driving stress detection /
- single module physiological signal /
- convolutional auto-encoder /
- abstract feature /
- ensemble learning

HTML全文

图 1 驾驶压力识别模型框架

Figure 1. Framework of driving stress estimation model

下载: 全尺寸图片幻灯片

图 2 一维卷积自编码器示意图

Figure 2. Schematic diagram of one dimensional convolution auto-encoder

下载: 全尺寸图片幻灯片

图 3 数据集实验流程

Figure 3. Experiment flow of data set

下载: 全尺寸图片幻灯片

图 4 一维卷积自编码器的训练

Figure 4. Training of one dimensional convolutional auto-encoder

下载: 全尺寸图片幻灯片

图 5 抽象特征的可视化

Figure 5. Visualization of abstract features

下载: 全尺寸图片幻灯片

图 6 不同路况下的ROC曲线以及AUC

Figure 6. ROC curves and AUC under different road conditions

下载: 全尺寸图片幻灯片

图 7 不同维度下的特征学习算法比较

Figure 7. Comparison of feature learning algorithms in different dimensions

下载: 全尺寸图片幻灯片

表 1 卷积自编码器的超参数

Table 1. Hyperparameters of the convolutional auto-encoder

Layer	Type	Hyperparameters	Feature map
1	Input	−	(3 100,1)
2	Convolution	5×1×16	(3 100,16)
3	Pooling	2×1	(1 550,16)
4	Convolution	5×1×32	(1 550,32)
5	Pooling	2×1	(775,32)
6	Reshape	−	(24 800,1)
7	Dense	−	(dim,1)
8	Dense	−	(775×dim,1)
9	Reshape	−	(775,dim)
10	Upsampling	2×1	(1 550,dim)
11	Deconvolution	5×1×32	(1 550,32)
12	Upsampling	2×1	(3 100,32)
13	Deconvolution	5×1×16	(3 100,16)
14	Deconvolution	5×1×1	(3 100,1)

下载: 导出CSV

表 2 基分类器及参数设置

Table 2. Base classifier and parameter setting

Classifier	Parameter
KNN	n_neighbors:1,2,3,4,5
RF	max_feature:1,2,3,4,5
GBDT	max_feature:1,2,3,4,5
SVM	C、gamma：[0.000 1,0.001,…,100]

下载: 导出CSV

表 3 不同基分类器不同维度下的识别结果

Table 3. Recognition results under different dimensions of different base classifiers

Dimension	Accuracy/%
Dimension	KNN	GBDT	RF	SVM
2	86.729	86.579	84.135	86.917
4	90.602	89.248	90.038	90.150
8	92.519	89.925	90.789	92.030
16	92.331	88.346	89.699	92.105
32	87.782	85.827	87.632	88.684
64	88.571	86.617	88.195	88.008

下载: 导出CSV

表 4 基分类器的投票集成

Table 4. Voting integration of base classifiers

Method	Precision/%	Recall/%	F1/%	Accuracy/%
Mean(base classifiers)	91.291	91.244	90.721	91.316
Voting(soft)	92.805	92.771	92.340	92.820
Voting(hard)	92.384	92.569	91.864	92.444

下载: 导出CSV

表 5 同类研究的对比

Table 5. Comparison of similar studies

Method	Physiological signal	Accuracy/%
Reference[19]	ECG,EMG,GSR,RESP	97.4
Reference[24]	ECG	83
Reference[25]	ECG	98.6
Reference[6]	FGSR	88.91
Reference[26]	FGSR	81.82
This paper	FGSR	92.82

下载: 导出CSV

参考文献(26)

[1]	WANG J S, LIN C W, YANG Y T C. A k-nearest-neighbor classifier with heart rate variability feature-based transformation algorithm for driving stress recognition[J]. Neurocomputing, 2013, 116(10): 136-143.
[2]	CHUNG W Y, CHONG T W, LEE B G. Methods to detect and reduce driver stress: A review[J]. International Journal of Automotive Technology, 2019, 20(5): 1051-1063. doi: 10.1007/s12239-019-0099-3
[3]	ZHANG C, WANG H, FU R R. Automated detection of driver fatigue based on entropy and complexity measures[J]. IEEE Transactions on Intelligent Transportation Systems, 2014, 15(1): 168-177. doi: 10.1109/TITS.2013.2275192
[4]	WANG H, WU C, LI T, et al. Driving fatigue classification based on fusion entropy analysis combining EOG and EEG[J]. IEEE Access, 2019, 7: 61975-61986. doi: 10.1109/ACCESS.2019.2915533
[5]	ZHANG B, MORERE Y, LANGLET C, et al. Reaction time and physiological signals for stress recognition[J]. Biomedical Signal Processing and Control, 2017, 38(9): 100-107.
[6]	CHEN L L, ZHAO Y, YE P F, et al. Detecting driving stress in physiological signals based on multimodal feature analysis and kernel classifiers[J]. Expert Systems with Applications An International Journal, 2017, 85: 279-291. doi: 10.1016/j.eswa.2017.01.040
[7]	OMID D, VIKAS R, MOJTABA T. Wearable driver distraction identification on-the-road via continuous decomposition of galvanic skin responses[J]. Sensors, 2018, 18(2): 503. doi: 10.3390/s18020503
[8]	BENGIO Y, COURVILLE A, et al. Representation learning: a review and new perspectives[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2013, 35(8): 1798-1828.
[9]	ZHAO L M, LI R, ZHENG W L, et al. Classification of five emotions from eeg and eye movement signals: Complementary representation properties[C]//2019 9th International IEEE/EMBS Conference on Neural Engineering (NER), San Francisco, USA: IEEE, 2019: 611-614,
[10]	REN Y F, WU Y. Convolutional deep belief networks for feature extraction of EEG signal[C]//2014 International Joint Conference on Neural Networks (IJCNN). Beijing: IEEE, 2014: 2850-2853.
[11]	WU H F, H Q, WANG D Q. A CNN-SVM combined model for pattern recognition of knee motion using mechanomyography signals[J]. Journal of Electromyography and Kinesiology, 2018, 42: 136-142. doi: 10.1016/j.jelekin.2018.07.005
[12]	HINTON G E, ZEMEL R S. Autoencoders, minimum description length and helmholtz free energy[C]//6th International Conference on Neural Information Processing Systems. USA: ACM, 1993: 3-10.
[13]	DU B, XIONG W, WU J, et al. Stacked convolutional denoising auto-encoders for feature representation[J]. IEEE Transactions on Cybernetics, 2017, 47(4): 1017-1027. doi: 10.1109/TCYB.2016.2536638
[14]	田莉莉, 邹俊忠, 张见, 等. 基于改进的卷积神经网络脑电信号情感识别[J]. 计算机工程与应用, 2019, 55(22): 99-105. doi: 10.3778/j.issn.1002-8331.1807-0219
[15]	WEN T, ZHANG Z. Deep convolution neural network and autoencoders-based unsupervised feature learning of EEG signals[J]. IEEE Access, 2019, 6: 25399-25410.
[16]	MASCI J, MEIER U, CIREŞAN D. Stacked convolutional auto encoders for hierarchical feature extraction[C]//the 21st International Conference on Artificial Neural Networks and Machine Learning(ICANN). Berlin Heidelberg: Springer, 2011: 52-59.
[17]	魏琛, 陈兰岚, 张傲. 基于集成卷积神经网络的脑电情感识别[J]. 华东理工大学学报(自然科学版), 2019, 45(4): 614-622.
[18]	LIU Y, SHRIBERG E. Comparing evaluation metrics for sentence boundary detection[C]//2007 International Conference on Acoustics, Speech and Signal Processing. USA: IEEE, 2007: 185-188.
[19]	HEALEY J A, PICARD R W. Detecting stress during real-world driving tasks using physiological sensors[J]. IEEE Transactions on Intelligent Transportation Systems, 2005, 6(2): 156-166. doi: 10.1109/TITS.2005.848368
[20]	LAURENS V D M, HINTON G. Visualizing data using t-sne[J]. Journal of Machine Learning Research, 2008, 2605: 2579-2605.
[21]	COX, TREVOR F, COX MICHAEL. Multidimensional scaling[J]. Journal of the Royal Stata Society, 2001, 46(2): 1050-1057.
[22]	KUNCHEVA L I, FAITHFULL W J. PCA feature extraction for change detection in multidimensional unlabeled data[J]. IEEE Transactions on Neural Networks and Learning Systems, 2014, 25(1): 69-80. doi: 10.1109/TNNLS.2013.2248094
[23]	EFTEKHARI A, BABAIE-ZADEH M, MOGHADDAM H A. Two-dimensional random projection[J]. Signal Processing, 2011, 91(7): 1589-1603. doi: 10.1016/j.sigpro.2011.01.002
[24]	MUNLA N, KHALIL M, SHAHIN A, et al. Driver stress level detection using HRV analysis[C]//2015 International Conference on Advances in Biomedical Engineering. Beirut Lebanon: IEEE, 2015: 61-64.
[25]	RIZWAN M F, FARHAD F, MASHUK F, et al. Design of a biosignal based stress detection system using machine learning techniques[C]//2019 International Conference on Robotics, Electrical and Signal Processing Techniques. Dhaka, Bangladesh: IEEE, 2019: 364-368.
[26]	LIU YUN, DU S Q. Psychological stress level detection based on electrodermal activity[J]. Behavioural Brain Research, 2018, 341: 50-53. doi: 10.1016/j.bbr.2017.12.021