Provision of Controlled and Consistent Light Distribution Over an Uneven Topography to Maximize Efficacy of Machine-Vision based Defect Identification

Abstract

Current manual inspection has risen as a bottleneck in aircraft component manufacture. Not only are conventional inspections slow and tiring, but the outcomes are also subjective and prone to inaccuracies, leading to limited streamlining of production, and in turn, a negative impact on competitive pricing. As a result, robotization and automation have become essential and have made significant strides in recent gas turbine production. At the root of the problem, challenges manufacturers face include the large variety of materials, variety of shapes and topographies, and automation systems' reliability. Automating the visual inspection process requires the employment of an optomechanical system. Such a system must capture images of a work-piece and then analyze the resulting pixel-domain data. Under properly controlled illumination, the captured images would be handled by the machine vision system easily and with a high degree of reliability and repeatability than manual inspection. This work provides controlled and consistent lighting methodologies to facilitate quality image data collection over complex-shaped, highly reflective surfaces. Two quantities were controlled: the luminous intensity of the light source and the illuminance of the surface under consideration. The central approach entails using arrays of independently controlled light sources to generate different lighting patterns, structures, and colors. Such results consider the geometry and 3D pose of parts in the environment and the surface topography of the work-piece to be inspected, hence amplifying a classical image capturing system (patent application serial no. 63/180,631). We discuss the mathematical problem formulation, the analytic solution, the optimality of the proposed shape-adaptive lighting, and experimental results in the imaging of curved parts common in aerospace manufacturing. Furthermore, we provide four solution configurations and compare the resulting defect identification using a transfer-learning deep neural network. The results assisted in creating an identification accuracy of more than 95 percent using only sparse data. In particular, compared with a non-shape-adaptive configuration, the proposed methods increased training and validation accuracy by 1.8 % and 1.6 %, respectively, and reduced false alarms and miss rates by 1.0 % and 2.25 %.