Fatigue Monitoring System Modeling And Hardware Achievement Based On EEG Signal Analysis

Recent development of society economy have made basic transport infrastructure more and more powerful, meanwhile the traffic accident caused by it is also increasing that make great loss of people life and belongs. Among these reason that make traffic accidents, fatigue driving take a lead. Although many car manufacturers have released a great number of fatigue monitoring systems, most of them is limited by its car model and accuracy. what’s more, these existed systems make judgements based on human being’s physical movements like eyes and head that is easy to make mistake and is not unique. For that reason, it is necessary to design a new fatigue monitoring system using EEG signal analysis methods. In this scheme, the only thing to do is to differentiate states of conscious and fatigue. This method is unique theoretically.Taking above reasons into consideration, I have researched and designed a new fatigue monitoring system using EEG signal. In my paper, whole contents is divided into two parts, they are fatigue monitoring software system modeling and hardware system achievement separately. In first stage, I have extracted three features:C0complexity, Kc complexity and ApEn of EEG raw data in the MIT EEG data warehouse included two states of conscious and fatigue. And more, I have used Mallat algorithm of wavelet packet transform to decompose EEG signal to obtain alpha, beta,theta, delta rhythms, then calculated the alpha rhythm power spectral density percentage as the auxiliary method on the ARM9 platform and the fourth feature on the Android platform. After making T test for four feature, I was sure that they have obvious difference between states of conscious and fatigue and can be regarded as effective features. To tell two stages apart, I have adopted Hierarchical Mixture of Experts(HME)as a classifier which contain two layers. The bottom layer have four expert networks which is responsible for the input and training of raw data. From bottom to top layer, there are bottom gate networks and top gate networks which isused to soft-partition data zone. The whole HME is trained by EM algorithm and every sub-networks was designed as an BP network that have three layers and is converged using Momentum gradient algorithm. Finally, I began to train our HME classifier which have a stop mark of precision of 0.05 utilizing four feature values and chosen lengths of raw data as network input. When training process is stopped, I obtained final value of 60% ? 8.398% and 65.125% ? 29.375% as HME test precision on the ARM9 and Android platform separately.In the hardware system stage, I first made an EEG signal acquisition analog circuit that is ascertained to be effective to get EEG signal from people forehead and to be used by next step processor. I have transplanted the modeling system on the ARM9 and smart android phone platform separately. ARM9 platform was combined with self-designed circuit while smart phone was connected with NeuroSky incorporation EEG head wear. Both of them can be roughly response our design requirement, but the accuracy of smart phone platform is higher than ARM9platform’s which have some bearing on EEG signal’s noise. So I finally decided to use smart phone platform as my ultimate hardware system.