Roughness prediction of laser cut edges by image processing and artificial neural networks

Abstract

Quality assessment of parts produced by laser cutting is a time-consuming task which typically requires expert knowledge. One of the main characteristics determining the quality of a cut edge is the roughness of the striations profile generated during the cutting process. Usual practice is to measure roughness by means of contact-based profilometry. However, such measurement requires significant effort to be performed and is difficult to automate in industrial practice. In this paper, an algorithm based on the multilayer perceptron architecture is described for efficient image-based roughness prediction of laser cut samples. The simple architecture of the selected neural network allows performing the training of the network with a low number of data points. This offers a clear advantage for possible industrial implementations. Additionally, the influence of the lighting conditions on the roughness prediction quality is evaluated.