Research on Sate Recognition of a Human Pilot and Decision Making Model

本研究旨在通过分析人类飞行员在着陆操作中的行为来提高飞行安全性。在1998财年，我们测量了飞行员操作飞行模拟器的视点，并分析了飞行员在着陆飞行中注意的地方。这项测量表明，飞行员无法在大约100英尺高的情况下监控仪表盘。在1999财年，我们建立了人类飞行员的数值模型。该模型包含卡尔曼滤波，它可以从输入数据和输出数据估计状态变量，从而使估计误差最小化。在这种分析中，飞行员注意力的分配比率被处理为未知参数，并被优化以最小化卡尔曼滤波中的估计误差指标。研究发现，在最后进近阶段的状态估计中，跑道侧边缘的视觉线索比其他视觉线索更重要，应将注意力分散到多个视觉线索上，以使估计误差的总和最小。利用飞行模拟器获得的实验数据证实了这些结果。在本研究的最后一年，根据模拟器测试的结果，利用神经网络自动构建飞行员模型。该模型的主要输入是地平线和跑道等观测线索，输出是电梯角度。两个案例(有风和无风)的数据被用于学习。接下来，对所构建的飞行员模型在不同情况下进行了仿真，验证了模型的有效性和泛化能力。最后，运用因子分析方法对模型进行了评价。

This research is aiming for improving flight safety through the analysis of human pilot behavior at landing operation. In the fiscal year of 1998, we measured the viewpoint of a pilot operating a flight simulator and analyzed where the pilot put attention through a landing flight. This measurement revealed that the pilot cannot monitor the instrument panel under about 100 ft height. In the fiscal year of 1999, we made a numerical model of the human pilot. This model contains the Kalman Filter which can estimate the state variables from the input data and the output data to minimize the estimation error. In this analysis, the distribution ratio of a pilot attentiveness are dealt with unknown parameters and are optimized to minimize the estimation error index in the Kalman Filter. It is found that the visual cue of the runway side edges is more important than the others in the state estimation in the final approach phase, and that attention should be distributed to several visual cues to minimize the total sum of the estimation errors. These results are confirmed from experimental data obtained by using a flight simulator. In the final year of this research, the pilot model is automatically constructed using neural network from the results of the simulator tests. The main inputs to this model are observation cues such as the horizon and the runway, and the output is the elevator angle. The data on two cases (with wind, or with no wind) are used for learning. Next, the constructed pilot models are simulated in various cases and we check their validity and generalization. Finally, the models are evaluated using factor analysis.