Analysis of Pilot Landing Control Using Neural Network

Abstract

A neural network (NN) modeling approach has been developed to analyze human pilots’ control by utilizing the recorded time history of visual cues and pilot control inputs. The contribution ratios of each visual cue are estimated by analyzing the obtained NN models. This paper compares the importance of visual cues such as the horizon, a runway shape, and a runway marker. NN models are obtained in all cases wherein a pilot intentionally changes the attentiveness to visual cues by using a flight simulator. The obtained contribution ratio and flight simulation results that are controlled by the estimated NN model clarify the role of each visual cue. Additionally, an experimental method is developed to record the time histories of visual cues and pilot control inputs by analyzing recorded video data during a real flight. The developed method makes it possible to obtain the NN model of a pilot for a real landing case. I. Introduction HE most difficult maneuver for airline pilots is perhaps a landing approach. However, it has yet not been automated. In particular, an airline pilot cannot afford to read instruments during the final approach. Therefore, they have to estimate airspeed, descent rate, altitude, and pitch angle of the aircraft by using visual cues from the cockpit. It is considered that this estimation skill plays an important role in the smoothness of the landing. 1 Since it is difficult to analyze the estimation skill directly, the author’s team has developed a new analysis tool using neural network (NN) modeling techniques. 2,3,4 This method utilizes the recorded time histories of visual cues such as the horizon, runway shape, and runway marker and the corresponding time history of the pilot control inputs. The contribution ratio and sensitivity of each visual cue to the pilot control can be estimated. Artificial NNs are mathematical models that emulate biological nervous systems and consist of a large number of highly interconnected processing elements such as neurons. 5 The NN model is adjusted by a learning process. Note that the NNs can relate input/output data that have high nonlinearity. The obtained NNs can be applied to automatic recognition systems and automatic control systems of complicated problems including human control. We have applied the NNs to analyze airplane pilots’ maneuvers during the landing phase. The input data for the NNs are visual cues, e.g., runway geometries and the horizon, and control column input. The output data from the NNs are control column and throttle lever deflections. The contribution ratios of each NN input to the pilot control column are computed to analyze the pilot maneuvers. It has been recognized that the most difficult problem in creating the NN models is deciding whether or not the obtained model has generalization capability. Generalization refers to the NN producing reasonable outputs for inputs not encountered during the training. In our analysis, Genetic Algorithms (GAs) 6 are applied to determine some parameters in the NN models, as described in Ref. 4, where this approach can increase the generalization capability. 4