基于集成SDAE和EEG的跨被试认知工作负荷识别

doi:10.16180/j.cnki.issn1007-7820.2021.03.009

摘要/Abstract

摘要：

基于脑电信号评估人机系统中操作员认知工作负荷状态,可以有效阻止操作员工作性能下降。文中提出一种跨被试认知工作负荷分类器E-SDAE,以适应被试间脑电特征分布的变化。该算法包括高水平个性化特征抽象和决策融合两个模块。特征滤波器利用基学习器SDAE来抽象一组被试的脑电特征。监督分类器利用超限学习机的随机性来融合经Q-statistics处理后得到的滤波脑电抽象。任务1和任务2分别取得0.635 3和0.674 7的分类率,并且显著优于一些传统的认知工作负荷评估器。时间复杂度计算结果表明,E-SDAE的计算负荷对于高维脑电特征是可接受的。

关键词: 认知工作负荷, 堆叠去噪自动编码器, 超限学习机, 人机系统, 脑电图, 集成学习

Abstract:

Assessing the operator's cognitive workload status in a human-machine system based on EEG signals can effectively prevent the operator's performance from degrading. This study proposes a novel inter-subject CW classifier, E-SDAE, to adapt the variations of the EEG feature distributions across multiple subjects. The E-SDAE includes two cascade-connected modules: high level personalized feature abstractions and decision fusion. The feature filters employ the base learner SDAE to abstract the EEG features from a group of individuals. The supervised CW classifier exploites the random characteristics of ELM to fuse the filtered EEG abstractions under the implementation of Q-statistics. The classification rate achieves 0.635 3 and 0.674 7 for Task 1 and Task 2, respectively, and are significantly better than several conventional CW estimators. The time complexity calculation results show that the computational workload of the E-SDAE is also acceptable for high-dimensional EEG features.

Key words: cognitive workload, stacked denoising autoencoder, extreme learning machine, human-machine system, EEG, ensemble learning

中图分类号:

TP391

郑展鹏,尹钟. 基于集成SDAE和EEG的跨被试认知工作负荷识别[J]. 电子科技, 2021, 34(3): 48-53.

ZHENG Zhanpeng,YIN Zhong. Inter-Subject Recognition of Cognitive Workload Based on Ensemble SDAE and EEG[J]. Electronic Science and Technology, 2021, 34(3): 48-53.

图/表 8

图1

图2

图3

表1

表2

表3

表4

传统算法的计算复杂度"

算法	时间复杂度
NB	ο $m log m$
LR	∝L₁
KNN	ο $nD$
ANN	ο $L$
ELM	ο $L 3 + L 2 + LN m + n$
LSSVM	ο $n 3$
SAE	ο $∑ i = 1 m h m + n × c × l$

表4

表5

参考文献 17

[1]	Paas F, Tuovinen J E. Cognitive load measurement as means to advance cognitive load theory[J]. Educational Psychologist, 2003,38(1):63-71.
[2]	陈善广, 姜国华, 王春慧, 等. 航天人因工程研究进展[J]. 载人航天, 2015,21(2):95-105.
	Chen Guangshan, Jiang Guohua, Wang Chunhui, et al. Advancement in space human factors engineering[J]. Manned Spaceflight, 2015,21(2):95-105.
[3]	莫秋云, 李荣敬. 山区公路条件下驾驶员生理特性研究[J]. 电子科技, 2013,26(3):83-89.
	Mo Qiuyun, Li Rongjing. Driver’s physiological characteristics under mountain road conditions[J]. Electronic Science and Technology, 2013,26(3):83-89.
[4]	赵云飞, 张立国, 张勤, 等. BP神经网络在AP1000核电站事故诊断应用中的初步研究[J]. 原子能科学技术, 2014,48(S1):480-484.
	Zhao Yunfei, Zhang Liguo, Zhang Qin, et al. Preliminary study on application of BP neural network in AP1000 nuclear power plant accident diagnosis[J]. Atomic Energy Science and Technology, 2014,48(S1):480-484.
[5]	Colligan L, Potts H W, Finn C T, et al. Cognitive workload changes for nurses transitioning from a legacy system with paper documentation to a commercial electronic health record[J]. International Journal of Medical Informatics, 2015,84(7):469-476. pmid: 25868807
[6]	刘辉, 杜玉晓, 彭杰, 等. 脑-机接口技术发展[J]. 电子科技, 2011,24(5):116-119.
	Liu Hui, Du Yuxiao, Peng Jie, et al. A review of brain-computer interface development[J]. Electronic Science and Technology, 2011,24(5):116-119.
[7]	Antonenko P, Paas F, Grabner R, et al. Using electroencephalography to measure cognitive load[J]. Educational Psychology Review, 2010,22(4):425-438.
[8]	Gevins A, Mc Evoy L K, Smith M E, et al. Long-term and within-day variability of working memory performance and EEG in individuals[J]. Clinical Neurophysiology, 2012,123(7):1291-1299. pmid: 22154302
[9]	Model D, Zibulevsky M. Learning subject-specific spatial and temporal filters for single-trial EEG classification[J]. Neuro Image, 2006,32(4):1631-1641. pmid: 16828316
[10]	Yin Z, Zhang J. Operator functional state classification using least square support vector machine based recursive feature elimination technique[J]. Computer Methods and Programs in Biomedicine, 2014,113(1):101-115. doi: 10.1016/j.cmpb.2013.09.007 pmid: 24138846
[11]	Yin Z, Zhang J. Cross-session classification of mental workload levels using EEG and an adaptive deep learning model[J]. Biomedical Signal Processing and Control, 2017,33(1):30-47.
[12]	Sauer J, Nickel P, Wastell D. Designing automation for complex work environments under different levels of stress[J]. Applied Ergonomics, 2013,44(1):119-127.
[13]	Zhang X Z, Zheng W L, Lu B L. EEG-based sleep quality evaluation with deep transfer learning[C]. Shanghai:International Conference on Neural Information Processing, 2017.
[14]	Baldwin C, Penaranda B. Adaptive training using an artificial neural network and EEG metrics for within and cross-task workload classification[J]. Neuro Image, 2012,59(1):48-56. pmid: 21835243
[15]	Fan J, Wade J W, Key A P, et al. EEG-based affect and workload recognition in a virtual driving environment for ASD intervention[J]. IEEE Transactions on Biomedical Engineering, 2018,65(1):43-51. pmid: 28422647
[16]	Ho T K K, Gwak J, Park C M, et al. Discrimination of mental workload levels from multi-channel fNIR Susing deep leaning-based approaches[J]. IEEE Access, 2019,7(1):24392-24403.
[17]	Yin Z, Zhang J. Cross-subject recognition of operator functional states via EEG and switching deep belief networks with adaptive weights[J]. Neurocomputing, 2017,260(1):349-366.

分类器	被试	精度	灵敏度	特异度	查准率	阴性值
	A	0.416 1	0.729 2	0.450 6	0.726 6	0.454 3
	B	0.441 5	0.744 4	0.476 1	0.739 7	0.482 4
	C	0.400 7	0.730 6	0.414 4	0.714 1	0.434 0
	D	0.380 2	0.701 4	0.418 3	0.707 0	0.412 0
SAE	E	0.395 4	0.725 6	0.445 6	0.723 4	0.448 8
	F	0.415 9	0.719 4	0.385 0	0.700 8	0.406 1
	G	0.432 0	0.764 7	0.400 0	0.718 3	0.459 0
	H	0.351 1	0.706 9	0.378 9	0.694 1	0.395 9
	Mean	0.404 1	0.727 8	0.421 1	0.715 5	0.436 6
	A	0.453 0	0.703 9	0.553 3	0.759 3	0.482 7
	B	0.467 6	0.759 4	0.681 1	0.826 2	0.587 0
	C	0.427 8	0.638 6	0.554 4	0.741 2	0.434 5
	D	0.352 8	0.605 0	0.445 6	0.684 1	0.364 4
ELM	E	0.519 4	0.767 8	0.752 2	0.861 9	0.616 8
	F	0.415 7	0.672 2	0.551 7	0.750 1	0.456 6
	G	0.425 6	0.712 5	0.624 4	0.793 0	0.518 0
	H	0.426 1	0.770 6	0.612 2	0.800 5	0.566 8
	Mean	0.436 0	0.703 8	0.596 9	0.777 0	0.5033
	A	0.631 7	0.785 3	0.741 7	0.861 7	0.644 4
	B	0.694 4	0.853 1	0.780 0	0.885 5	0.727 9
	C	0.629 3	0.845 3	0.402 2	0.736 9	0.593 5
	D	0.529 1	0.909 4	0.407 8	0.753 0	0.719 9
E-SDAE	E	0.754 6	0.820 6	0.793 9	0.896 1	0.711 9
	F	0.552 0	0.770 8	0.487 8	0.750 1	0.519 6
	G	0.626 1	0.710 0	0.588 3	0.787 7	0.527 4
	H	0.665 4	0.891 7	0.760 6	0.881 4	0.806 5
	Mean	0.635 3	0.823 3	0.620 3	0.819 0	0.656 4

分类器	被试	精度	灵敏度	特异度	查准率	阴性值
	I	0.455 9	0.631 0	0.619 8	0.711 6	0.534 1
	J	0.395 9	0.570 1	0.511 2	0.636 6	0.442 4
	K	0.459 7	0.685 1	0.676 7	0.760 7	0.588 9
SAE	L	0.499 7	0.658 0	0.762 1	0.806 0	0.597 6
	M	0.397 2	0.612 1	0.594 0	0.693 4	0.505 1
	N	0.414 8	0.546 6	0.597 4	0.670 3	0.468 5
	Mean	0.437 2	0.617 1	0.626 9	0.713 1	0.522 8
	I	0.441 0	0.594 8	0.685 3	0.739 3	0.530 0
	J	0.423 8	0.424 7	0.662 9	0.635 0	0.443 3
	K	0.509 3	0.671 3	0.757 8	0.806 9	0.605 1
ELM	L	0.505 2	0.647 1	0.762 9	0.802 1	0.592 3
	M	0.475 9	0.525 3	0.694 0	0.720 0	0.493 9
	N	0.397 6	0.496 6	0.600 0	0.651 1	0.442 4
	Mean	0.458 8	0.560 0	0.693 8	0.725 7	0.517 8
	I	0.643 1	0.624 7	0.798 3	0.836 8	0.605 2
	J	0.558 6	0.635 1	0.741 4	0.784 2	0.642 1
	K	0.801 7	0.868 4	0.922 4	0.947 3	0.828 0
E-SDAE	L	0.782 8	0.880 5	0.889 7	0.923 1	0.832 0
	M	0.659 0	0.944 8	0.765 5	0.858 2	0.904 8
	N	0.603 1	0.687 9	0.687 9	0.813 1	0.796 8
	Mean	0.674 7	0.807 1	0.800 9	0.860 4	0.768 2

模型	任务1		任务2
模型	均值	标准偏差	均值	标准偏差
ELM	0.436 0	0	0.458 8	0
LSSVM	0.442 9	0	0.476 3	0
KNN	0.397 7	2.593 5×10^-16	0.443 4	2.009 3×10^-16
ANN	0.447 7	2.102 8×10^-16	0.485 9	3.457 9×10^-16
LR	0.432 4	2.628 5×10^-16	0.427 8	2.850 5×10^-16
NB	0.396 6	2.278 0×10^-16	0.368 2	1.635 5×10^-16
SAE	0.404 1	0	0.437 2	0
E-SDAE	0.635 3	0.072 6	0.674 7	0.097 8

文献	特征	分类器	人机任务	分类器类型	精度
[13]	EEG	三分类, TLDA	睡眠	独立被试	0.821 6
					Task 1: 0.590 0
[14]	EEG	二分类, ANN	阅读	特定被试	Task 2: 0.600 0
					Task 3: 0.590 0
[15]	EEG	二分类, KNN	驾驶	独立被试	0.860 0
[16]	fNIR	二分类, SVM AdaBoost DBN CNN	Stroop任务	混合被试	模型1: 0.647 4 模型2: 0.711 3 模型3: 0.842 6 模型4: 0.727 7
		三分类			脑力负荷: 0.683 8
[17]	EEG	五分类	AutoCAMS	跨被试	精神疲劳: 0.769 0
		SDBN			双负荷: 0.544 4
本文	EEG	三分类 E-SDAE	AutoCAMS	跨被试	Task 1: 0.635 3 Task 2: 0.674 7