Experimental validation of a machine learning algorithm for roughness quantification in laser cutting
Nikita Levichev, Alberto Tomás García, Masoud Kardan, Dirk Cattrysse, Joost R. Duflou
Abstract
Automatic quality assessment of laser cut parts is an essential building block in the deployment of data-driven strategies for process parameter optimization, predictive maintenance and continuous improvement in sheet metal workshops. From an industrial perspective, the major quality features of a laser cut edge are roughness, perpendicularity deviation and dross occurrence on the lower edge. While conventional contact-based profilometry can be used for highly accurate roughness measurements, its broad adoption in the daily operations of sheet metal workshops cannot be realized without significant deterioration of the workshop performance. As a faster and less demanding alternative to contact measurements, image-based roughness quantification by means of artificial neural networks has been evaluated in recent publications for a single material-thickness combination. Although promising results have been reported, extensive testing of the system performance is required before the system can be deployed in an industrial setting. Therefore, as the next step towards industrial implementation, this contribution validates the performance of a multilayer perceptron algorithm for roughness prediction of materials with different reflectivity behavior, such as stainless steel, mild steel and aluminium. Additionally, the robustness of the method depending on the cut edge surface conditions and type of lighting is investigated.