Features selection approaches for an objective control of cosmetic quality of coated surfaces

Abstract

The cosmetic aspect is one of the main functions of industrial surfaces in numerous applications. Even the smallest surface defects may have a critical effect on the cosmetic tolerability of such industrial surfaces. Thus, surfaces are generally coated at the last manufacturing process stage to cover existing defects and to certify their cosmetic quality. The surface quality is however constantly controlled after coating that induces an increase of lead-time increase and production costs. This is due to a various flaw patterns and a lack of uncoated surfaces specifications. Hence, the identification of relevant surface morphological parameters underlies an objective and automatic cosmetic control performance. In fact, this relevant parameter selection allows tracking surface flaws during the coating finishing operation. This paper presents a comprehensive overview of various feature selection tools for data analysis (Neighbourhood Component Analysis (NCA), ReliefF, Sequential wrapper method, Decision tree) to extract relevant information out of physical data. A design of experiment based on scratches test on amorphous polymers to generate typical controlled defects has been performed. Then, several cosmetic defects characteristics were extracted from experimental measurements. Feature selection approaches were applied and compared to determine the most relevant parameters. The advantages and limitations of each method for data analysis have been highlighted in the case of real engineering surface quality control.