Applying Neural Networks to the F-35 Seam Validation Process

Abstract

The F-35 Joint Strike Fighter (JSF) costs 79–115 million dollars to produce and is currently the most sophisticated military aircraft in history. In order to ensure that the vehicles meet expected performance standards, detailed and often expensive inspections of the aircraft structure must be performed at each production step. This paper describes work performed to integrate automation into one of the costliest manufacturing processes in the JSF program, known as Seam Validation, via the development and staged deployment of machine learning (ML)-based classifiers. Presented are two high performing artificial neural network (ANN) models capable of achieving classification accuracies of 94.3 percent and 90.8 percent respectively on troubleshooting/rework evaluations of the gap and mismatch components of seams on the aircrafts' structure. Once fully integrated into the production infrastructure, these models can potentially save tens to hundreds of thousands of labor hours over the life of the JSF program. Work presented in this paper discusses ML method development and describes rigorous model testing. The first step in model development required the identification of an algorithm that could cope with the relatively limited and unbalanced data available. To accomplish this, multiple technologies were investigated, including a support vector machine (SVM), a Bayesian network, and an artificial neural network (ANN). Of these model types, the ANN exhibited the best performance and was optimized through various described methods. Utilizing this model optimization approach, this work demonstrates an effective procedure for the development and successful deployment of high performing and robust machine learning technologies, effectively reducing costs and automating mass production processes like the JSF.