| Vigilance usually is defined as the ability to maintain focus of attention and to remain alertto stimuli for prolonged periods of time. In our daily lives, for many human-machine interac-tion systems, the operators should retain their vigilance above a constant level. Especially forcar drivers, losing vigilance may cause some serious traffic accidents. Therefore, research onvigilance monitoring algorithms and monitoring the driver’s vigilance level in real-time is an im-portant topic for safe driving in vehicle auxiliary, and has important practical significance to solvethe problem of traffic safety. Because EEG signal is the “gold standard” for vigilance detection,and the emergence of commercial portable EEG recording equipment makes EEG recording moreconvenient. In this thesis, we will focus on the key methods for EEG-based vigilance monitor-ing. The research topic include: on-line noise reduction and artifact removal algorithms for EEGpreprocessing, vigilance-related critical brain region location, efficient vigilance-related featureextraction algorithm, on-line EEG feature filtering algorithm, robust vigilance feature dimensionreduction algorithm, efficient on-line vigilance monitor algorithm, and vigilance automatic annota-tion algorithm. The main goal is, together with the portable EEG recording system, to provide keytechnologies and theories for future EEG and other physiology signals-based vigilance monitoringsystem.The main contributions and innovations of this thesis are listed as below:1. EEG on-line noise reduction and artifact removal. An adaptive on-line noise reduction andartifact removal algorithm is proposed, which can adaptively track the time-varying signalsource, automatically recognize the artifact source, and on-line reconstruct the EEG signalswithout artifact. This algorithm can solve the problem of noise or artifact interference inEEG recording. The artifact recognition precision is more than90%, which has reached thelevel of manual artifact annotation. Compared with other artifact removal algorithms, thisalgorithm can solve the problem of artifact source changing with time, and can support theon-line analysis mode.2. vigilance-related critical brain region location. The concept of EEG-based vigilance-relatedsynchronization and desynchronization is proposed. Based on this concept, three algo-rithms for vigilance-related critical brain region location is designed, which are correlation coefficient-based location algorithm, ICA-based location algorithm, and CSP-based locationalgorithm. And critical brain region is found, which is near the occipital lobe. RecordingEEG signals directly from the critical brain region can dramatically reduce the influenceof vigilance-unrelated EEG signals, simplify the EEG recording process, and enhance thepracticability of EEG-based vigilance analysis system.3. Vigilance-related feature extraction. For the logarithmic form of the EEG spectral feature,we have presented an explanation from the physical aspect, and defined it as the differentialentropy feature of EEG. The multiplicative noise problem of EEG feature is drawn out, andthe corresponding noise reduction method also is proposed. Based on the EEG-vigilanceexperiment data set, a systematic comparative study on performance and algorithm mech-anism is conducted for five kinds of EEG features, autoregressive coefficient, energy spec-trum, fractal dimension, sample entropy, and differential entropy. Experiment results showthat, differential entropy is the most accurate and stable EEG feature to reflect the vigilancechanges.4. On-line EEG feature filtering. A linear dynamical system (LDS)-based EEG feature filteringalgorithm is proposed, which is used to on-line remove the vigilance-unrelated EEG featurefrom the original EEG feature. Based on the EEG-vigilance experiment data set, a compar-ative study on performance and algorithm mechanism is conducted between the LDS-basedfiltering algorithm and the traditional moving average filter. Experiment results show that,LDS-based filtering algorithm can more fully use the observed EEG data, is better to elimi-nate the influence of the vigilance-unrelated EEG feature, and finally improve the vigilanceestimation accuracy.5. Robust vigilance feature dimension reduction. Usually the noise in EEG feature is muchstronger, if we directly use the ordinary PCA algorithm for feature dimension reduction,with the influence of noise, the extracted principal components will be skewed, then the vig-ilance estimation accuracy will be reduced. To solve the noise problem of feature dimensionreduction, noised unsensitive dimension reduction algorithm is introduced, which is RobustPCA. A comparative study is conducted based on experiment data. Experiment result showthat, compared with the ordinary PCA, Robust PCA can effectively improve the performanceof the vigilance estimation system when the EEG features containing lots of noise.6. Efficient on-line vigilance estimation algorithm. An on-line vigilance estimation algorithmnamed Larsen-ELM is proposed. Compared with SVM regression model, Larsen-ELM candramatically speed up the model training process while still achieve a similar vigilance es-timation accuracy. For large scale EEG-vigilance data analysis, Larsen-ELM has a highpractical value. 7. EEG-based automatic vigilance annotation algorithm. The traditional vigilance annotationmethods usually have a very high cost, very poor automation capability, and can not handlethe large scale data set. Therefore, two kinds of EEG-based vigilance annotation algorithms,discrete annotation algorithm and continuous annotation algorithm are proposed. Since theEEG feature is affected by the vigilance changes, the proposed algorithms can directly usedthe distribution information of the EEG feature combine with the clustering algorithm ormanifold learning algorithm for automatic vigilance annotation. Compared with the tradi-tional annotation algorithm, the proposed algorithms also can accurately annotate the trendof vigilance changes, but only have small errors in the detail vigilance annotation. As theproposed algorithms can automatically annotate the EEG data, and do not need training data,by using these algorithms, the efficiency of vigilance annotation on EEG-vigilance data setwill be dramatically improved.Finally, we have designed an EEG-vigilance experiment, and have collected lots of EEG-vigilance experiment data. With the support of adequate experiment data and from the aspectsof experiment results and algorithm mechanism, a systematic comparative study on system per-formance is conducted for different algorithms on every vigilance analysis sub process. Then therecommendation algorithms for every vigilance analysis sub process have been raised. Experimentresults show that, the proposed method for EEG-based vigilance estimation, can indeed achieveaccurate and reliable on-line vigilance estimation and off-line vigilance annotation. |
PDF Full Text Request