An investigation of neural networks for F-16 fault diagnosis. I. System description

摘要

The authors report results of ongoing research exploring the use of artificial neural networks (ANNs) for F-16 flight line diagnostics. ANNs hold the promise of solving difficult logistics problems such as multiple fault diagnosis, prognostication, changing configurations and environments, and inaccurate diagnosis attributable to incomplete and/or flawed rules. The authors tested three representative ANNs to see which type worked best for the problem considered. The authors chose back propagation (BPN) and counterpropagation (CPN) because they are considered to be two of the more promising pattern matching paradigms. The binary adaptive resonance theory I (ART1) was also chosen because it learns faster than CPN or BPN and has online adaptation (i.e. does not have to be totally retrained every time a new pattern is discovered). Online adaptation is a powerful attribute, allowing new associations to be immediately incorporated into the knowledge base on the flight line or wherever needed. The authors explain the advantages and drawbacks to each network tested and describe how they were trained using historical flight line data. Of the three ANNs examined, ART1 proved to be the most appropriate and was able to produce multiple-symptom-to-multiple-fault diagnoses.