A. Portable EEG acquisition module
Fig. 3-23(a ~ c) are the front-end analog circuit and digital control circuit in our portable EEG acquisition module, and the whole EEG acquisition module respectively, and the size of each circuit compared with a coin of one NTD was shown in Fig. 3-23.
There are three leads in our portable EEG acquisition module, includes EEG input, reference, and virtual ground of the front-end analog circuit. The electrodes connected
and behind right ear respectively. The specification of portable EEG acquisition module was listed in Table 3-3.
(a)
(b)
(c)
Fig. 3-23: (a) The front-end analog circuit, (b) the digital control circuit, and (c) the whole portable EEG acquisition module with single channel.
Table 3-3: The spec of portable EEG acquisition module
Type Portable EEG Acquisition Module
Channel Number 1~8
System Output Voltage Range 0~3V
Gain 5000 Bandwidth 0.1~100Hz
ADC Resolution 12bits
Output Current 29.5mA
Battery Lithium 3.7V 450mAh 15~33hr
Full Scale Input Range 577μV
Sampling 512Hz
Input Impedance greater than 10MΩ
Common Mode Rejection Ratio 77dB
Power Supply Rejection Ratio 88dB
Size 18mm x 20mm and 25mm x 40mm
B. DSP Module and SD Card Circuit
The expanded SD card circuit was shown in Fig. 3-24(a). It looked like a SD/MMC card, which can easily be plugged into the SD/MMC socket in DSP module.
The size of expanded SD card circuit is 24mm x 32mm. Fig. 3-24(b) is the illustration for application of expanded SD card circuit.
(a)
(b)
Fig. 3-24: (a) The expanded SD card circuit and (b) illustration for application of expanded SD card circuit
Chapter4 Unsupervised Approach
Based on the unsupervised analysis flowchart in Fig. 2-8, we will further discuss the details of every analysis diagrams in the following sessions. In order to find out the real driving behavior information, first we calculate the driver’s driving performance by using the record in simulation experiment. Moreover, we use the unsupervised analysis method to analyze the corresponding EEG information, including the preprocessing, alert model construction, and computation of the deviation using Mahalanobis distance method.
4.1 Driving Performance
The VR-based four-lane straight highway scene was applied in the experiment.
In this scene, the four lanes from left to right are separated by a median stripe and the distance from the left side to the right side of the road was equally divided into 256 points indicating the position of the vehicle as the digital output signal of the VR scene at each time instant. The width of each lane and the car is 60 units and 32 units, respectively. Fig. 2-4 shows an example of the driving performance represented by the vehicle deviation trajectories. We have defined an indirect index of the subject’s alertness level (driving performance) as the deviation between the center of the vehicle and the center of the cruising lane. VR driving simulation environment will randomly start a deviation event to move the car to right or left side in the car driving experiments. Subjects needs to sense those sudden movements and trying to make a reversely turn to back to the third lane. At one time, the VR environment also outputs
deviation event.
Fig. 4-1: The example of deviation event and car trajectories
In Fig. 4-2, the driving trajectories that we recorded followed below steps to show the driving performance. For restoring trajectories data, event trigger removal is the first process that we do. After deviation response offset, the positions of every experiment trial aren’t consistent, so that we need to remove the baseline every trial.
The results of the second step will leave right or left turn trajectories. And then absolute trials to collocate total right / left turn data. Typically the drowsiness level fluctuates with cycle lengths longer than 4 minutes [64], [65], and hence we smooth the indirect alertness level index using a causal 90-sec moving window advancing.
This helps us to eliminate variance with cycle lengths shorter than 1-2 minutes. We emphasize that this index is used only to validate our approach, and it is not as an input to develop the model for the alert state of the subject.
Fig. 4-2: The processing steps of driving performance
Following the above 4 steps, an example of driving performance are shown as
Fig. 4-3. Fig. 4-3(a) shows the original driving data which including event triggers, and Fig. 4-3(b ~ d) shows the results of 4 steps respectively. The final driving performance is in Fig. 4-3(e). Thus, we use this result to compare with MD*(MDT, MDA, and MTC) and implement in correlation analysis with the driver’s performance.
Fig. 4-3: Example of driving performance analysis. (a ~ d) are the fragment of information which marked by two lines. (a) is the original driving trajectories data which including deviation event triggers. (b) is the result which has passed through event trigger removal. (c) is the absolute result. (d) is the result which has smoothed by 90-sec moving average. (e) shows the total driving performance data.
4.2 Smoothing of the Power Spectra
Before extracting the power spectra of alpha and theta rhythms, raw EEG data would be preprocessed to remove power line noise and increase the resolution in the low frequency spectra. In this smoothing method, we used a moving average, as a low-pass filter to cut-off at 32 Hz in and filter noise over 32 Hz. A moving average filter was used to minimize the presence of artifacts in the EEG records of all sub-windows. Next, we down sample 8 times to 64Hz, so that every sub-window only left 64 points in one second. Those two preprocessing methods can decrease the unnecessary noise and increase the low frequency band information in theta and alpha band spectra. Go on, building up an 8 second moving window to save sub-windows, and displace a sub-window in every 1 second. The first FFT result will be produced at 8th seconds; moreover other FFT results will be in every following 1 second. The smoothing method of moving window can reserve the low frequency information of EEG power spectra longer to further analysis. Thus, for each session EEG log power time series at alpha band as well as at theta band with 1 sec time intervals were generated. Fig. 4-4 showed the processes of spectra analysis as precedence.
Fig. 4-4: Processes of spectra analysis as precedence