Modeling of the Visual Approach to Landing Using Neural Networks and Fuzzy Supervisory Control

Abstract

During the visual approach to landing of a x ed wing aircraft, a human pilot bases control and timing of subsequent maneuvers mainly on the out-the-window view, as there is not su cient time to read all instruments. The skill of making smooth and soft landings is acquired mainly through experience. Research has been done to identify the most important features in the visual scene (cues) for two phases of the visual approach to landing: glide slope tracking and the are maneuver. Using simulator and real ight data, neural networks have been trained for both phases to mimic the pilot’s control based on the visual cues available. By using the operator in neuron transfer functions, a transparent model is obtained. Fuzzy supervisory control is proposed to couple the networks and thus provide insight in the pilot’s decision making process with respect to timing the are initiation. The visual approach to landing is generally considered one of the most demanding phases in human pilot control [1]. The combination of high workload, having to interpret the visual scene, timing the initiation of subsequent maneuvers and executing those maneuvers, all with the risks inherent to lowaltitude ight, makes this process di cult to learn for new pilots. Real and/or simulated experience is indispensable to obtain and maintain landing skills, and performance feedback is thought to greatly improve learning e ciency [1, 2]. However, most pilots cannot explain what they look at or how they make their decisions and even training methods are not consistent. The research presented in this paper focuses on nding the visual cues a pilot uses, through analysis of scene and ight control data. A method is presented to construct a model of a human pilot which takes visual cues and generates longitudinal control actions during the visual approach to landing. This model is based on numerical data from real or simulated landings by human pilots. The model itself however is merely used to verify correspondence between the real pilot and the model. Of main interest are the structure and parameters of the resulting model, i.e., the driving inputs, internal relations and thresholds, as these give insight in the pilot’s (subconscious) behavior. The knowledge gained from this ireconstruction of the pilot’s mindi would be useful in training or evaluation of pilots: if we know how experienced pilots use the available visual cues to make smooth and soft landings, these insights can be taught