Machine learning based layer roughness modeling in robotic additive manufacturing

Abstract

Wire Arc Additive Manufacturing (WAAM) is a manufacturing technique that deposits metal layer upon layer to manufacture 3D parts based on welding processes. Most researchers considered weld bead width, height, and penetration as the characteristic performance in WAAM. However, layer roughness is also important because it affects the machining cost and mechanical properties of fabricated parts. If the roughness of a deposited layer can be reduced, less machining will be required, and material wastage will be reduced. Reduced layer roughness will also enable better bonding between adjacent layers. Hence, the deposition of weld beads with minimized roughness demands great attention. A few researchers who tried to investigate roughness in WAAM used straight paths for material deposition, but the investigation of the weaving path, which has a great potential to reduce layer roughness, has not been investigated well. The main contribution of this paper is about successfully implementing two machine learning methods to accurately model surface roughness in WAAM using a weaving path: Random Forest and Multilayer Perceptron (MLP) which is also known as Artificial Neural Network (ANN). Both methods are effective for modeling and predicting the layer roughness for a given set of robotic WAAM parameters, but Random Forest gave better results than MLP in terms of accuracy and computational time.