基于注意力机制的飞行学员疲劳分类网络研究

Research on Fatigue Classification Network of Student Pilot Based on Attention Mechanism

近年的通用航空事故中,出现在飞行训练期间由人因引起的事故约1/3。疲劳是人因中重要的一种,准确地识别飞行学员的疲劳状态对飞行安全有积极意义。提出了一种融合注意力机制的飞行学员精神疲劳分类网络,使用深度可分离卷积对网络参数进行精简,引入注意力机制充分利用通道与空间特征对脑电无用信息进行抑制的同时增大了关键信息权重。使用公开疲劳数据集SEED-VIG对网络进行预训练,然后通过迁移学习将其应用于飞行学员精神疲劳分类,网络在对4种精神疲劳状态分类时达到了84.12%的准确率,并在NVIDIA Jetson Nano上达到了27samples/s的运行速度。该网络可有效识别飞行学员的疲劳类型,并及时向飞行教员反馈学员精神状态情况,为判断飞行学员是否适合训练提供生理依据,同时该网络满足移动设备和嵌入式设备的计算要求。

In general aviation accidents in recent years, accidents caused by humans during flight training are about1/3. Fatigue is an important one of human factors. Accurately identifying the fatigue state of student pilot is of positive significance to flight safety. This study proposes a mental fatigue classification network for pilots that integrates attention mechanism, uses deep separable convolution to streamline network parameters, and introduces attention mechanism to make full use of channel and spatial features to suppress useless EEG information while increasing the weight of key information. Use the public fatigue data set SEED-VIG to pre-train the network, and then apply it to the mental fatigue classification of flight students through migration learning. The network achieves an accuracy of 84.12% when classifying the four mental fatigue states, and is tested in NVIDIA Jetson. The Nano has reached a running speed of 27 samples/s. The network can effectively identify the type of student pilot fatigue, and timely feedback the student’s mental state to the flight instructor, which provides a physiological basis for judging whether the student pilot is suitable for training, at the same time, the network can meet the computing requirements of mobile devices and embedded devices.