Abstract:
Over two million people die annually due to work-related accidents. A significant factor
contributing to this large number of fatal accidents is the passive nature of modernday
jobs. A person on a sedentary job tends to lose attention, and his mind starts
wandering off, although he physically looks attentive. There are no physical cues of
an inattentive mental state. Mental states with no secondary information, such as
physical cues, are categorized as pure mental attention states. In this thesis, we use a
multi-layer perceptron classifier for improving the detection of pure mental states using
electroencephalography (EEG) signals. These signals are very noisy, so we use STFT
with the Blackman window for pre-processing of data. We restrict the frequencies
between 0-18 Hz and generate generic features vectors for 7 EEG channels. 80 % of
the data is used for the neural network training while 10 % for validation and the
rest for the testing. The neural network can classify the three mental states: Focused,
Unfocused, and Drowsy with an overall accuracy of 66.7 %. The data collected for
the FZ channel has a testing accuracy of 87.5 %. The highest specificity for each
mental attention state is as follows: 89.3 % for the "Focused Mental State" from the
CZ channel, 92.9 % for the "Unfocused Mental State" on the data collected by the PZ
electrode, and for the "Drowsy Mental State" all channel outputs had 100 % specificity
except CZ.
v
