Using neural networks to assess flight deck human-automation interaction

The increased complexity and interconnectivity of flight deck automation has made the prediction of human-automation interaction (HAI) difficult and has resulted in a number of accidents and incidents. There is a need to develop objective and robust methods by which the changes in HAI brought about by the introduction of new automation into the flight deck could be predicted and assessed prior to implementation and without use of extensive simulation. This paper presents a method to model a parametrization of flight deck automation known as HART and link it to HAI consequences using a backpropagation neural network approach. The transformation of the HART into a computational model suitable for modeling as a neural network is described. To test and train the network data were collected from 40 airline pilots for six HAI consequences based on one scenario family consisting of a baseline and four variants. For a binary classification of HAI consequences, the neural network successfully classified 62-78.5% depending on the consequence. The results were verified using a decision tree analysis.