Improvement of Fixation Elements Detection in Aircraft Manufacturing

Resumen

The requirements for accuracy and reliability in the manufacturing processes of large aircraft structures are among the most demanding in the industry due to the continuous development of advanced manufacturing processes with tight tolerances and high requirements for process integrity. The main technique for the operation of many of these processes is the detection and precise measurement of fasteners by artificial vision systems in real time, however these systems require adjustment of multiple parameters and do not work correctly in uncontrolled scenarios, which require intervention operations such as manual supervision of the measurement process, leading into a reduction in the autonomy of automated systems. In this study, a new Deep Learning algorithm based on a Single Shot Detector neural network is proposed for the detection and measurement drills and other fixation elements (such as rivets and temporary fasteners) in an uncontrolled industrial manufacturing environment. The convergence of the new network has been optimized for the detection of elements of a circular nature, instead of the generic anchor boxes usually used. In addition, a fine-tuning algorithm based on a new characterization parameter of the circular geometry is applied to the results obtained from the network. This new metric has made it possible to define a quality parameter with respect to the measurement made.