A validation methodology for neural network based flight controlsystems

摘要

A significant problem associated the inclusion of a neural networkin an avionics system is validating that system to the requiredreliability level. To accomplish this, it is necessary to associate a“probability of failure” with the neural network and,ultimately, with the operational flight program. It would be morecorrect to say that the probability of excitation of the network in anunvalidated portion of its input space is required. In this sense,estimation errors associated with neural networks are like latenthardware faults, and techniques that were previously used to measure theprobability of failure of hardware due to fault latency can be used tomeasure the probability of failure of the network. A methodology wasdeveloped and applied to a flight controller designed to operate in awell defined environment. The controller incorporated a network toestimate nonlinear portions of plant performance. The results of thestudy indicates that the technique could be used to provide a finalvalidation of the network and the controller to a specified reliabilitylevel and to evaluate the role of flight test in networkvalidation