A dynamic surface roughness prediction system based on machine learning for the 3D-printed carbon-fiber-reinforced-polymer (CFRP) turning

This study proposes a novel surface roughness prediction system that uses machine learning and dynamic inputs for additively-manufactured, Carbon-Fiber-Reinforced-Polymer tubular workpieces. First, an investigation of the effects of standard machining conditions on the generated surface roughness was carried out, to assess the machinability of the 3D-printed, composite workpieces during turning. Two sets of specimens were fabricated, each with different wall layer thickness (WT) and a set of experiments was designed with respect to the selected range of cutting-speed (Vc), feed (f) and depth-of-cut (ap). As expected, it was found that all process parameters affected the generated roughness with cutting-speed and feed contributing the most to the results. The research hypothesis was that an Artificial Neural Network (ANN) that includes vibration signals together with the cutting conditions would provide better surface roughness predictions. Two shallow, three-layered ANN models were used. The first model utilized the machining parameters and the second model was based on the first one, with the addition that the acquired acceleration signals, to provide meaningful representations of vibrations with the aid of the Principal Component Analysis. The first model yielded a Mean Absolute Percentage Error (MAPE) equal to 2.59%. The second model provided more accurate surface roughness predictions, with MAPE being reduced to 1.51%. Finally, a Generic Algorithm (GA) was employed to identify the optimal process parameters for minimizing the response. The best combination was determined to be: WT = 0.50 mm, Vc = 173.2 m/min, f = 0.04 mm/rev and ap = 0.50 mm.