Abstract: About 80% of water traffic accidents involve human factors, and deck officer fatigue is one of the critical causes of ship traffic accidents. On road, the development of driver fatigue detection technology based on EEG（Electroencephalogram） has been helpful in quick and accurately identifying the degree of driver fatigue. However, the accuracy of EEG-based driver fatigue detection is affected by the sensitivity of EEG signal and individual differences. For research, EEG signals of several examinees is collected during a ship simulation training. The EEG signals in three channels of the prefrontal lobe related to fatigue are processed. Various EEG signal features such as MAV（Mean Absolute Value）, SD（Standard Deviation）, RMS（Root Mean Square）, and SE（Shannon Entropy） are extracted. Based on the KSS（Karolinska Sleepiness Scale）, deck officer fatigue is classified into three levels: “alert”, “middle”, and “fatigue”. The fatigue levels of the examinees are discriminated with different classification algorithms and the results are compared. The comparison shows that Bi-LSTM（Bi-Long Short Term Memory） has the best performance with the classification accuracy of 88.63%. The current study demonstrates the power of this method in studying the individual-insensitive three-class classification problem.