An accurate roughness prediction in milling processes through analytical evaluation and KNN regression approach

In milling processes, an accurate prediction of surface roughness depending on the cutting parameters of the workpiece and the geometric profiles of the tool cutting edges can be decisive for high-quality and functional products. Roughness real value obtained at the end of the machining cannot be calculated analytically with extreme accuracy because in addition to depending on “basic” cutting parameters (feed rate, infeed, Z-pitch, spindle speed), it is influenced by numerous “complementary” parameters that significantly alter the theoretical value. Among these complementary parameters that show a greater influence were identified as follows: the degree of lubrication (lc), the mechanical and grain characteristics of the tool and the material (gc), tool wear (tw), the presence of vibrations (v), material inhomogeneity (mi), intrinsic material micro-imperfections (mmi), and unwanted interactions with the environment (Ei). In this research, through an analytical evaluation based on the basic cutting parameters and a machine learning (ML) regression approach according to the complementary parameters, an accurate prediction model of real surface roughness was developed. A classic K-Nearest Neighbor (kNN) algorithm, operating on seven dimensionalities of complementary parameters with a squared Euclidean metric, was trained from key performance indicators (KPIs) with a fairly populated dataset of experimental roughness measurements. Specifically, a “full factorial” dataset was selected, which evaluated all combinations of the first four complementary parameters (lc, gc, tw, v) for different processes and a “reduced factorial” for the remaining three parameters (mi, mmi, Ei). A pre-processing phase with adequate cleanup and integration of the training data has shown that the proposed methodology can provide a prediction accuracy of up to 99.06%.