Prediction of surface roughness in CNC turning by model-assisted response surface method

Various statistical approaches such as classical regression and modern machine learning methods have been applied to measurement data for estimating the status of manufacturing processes, which is now boosted by the movement of Internet of Things (IoT). In this study, we attempt to integrate an analytical tool model of surface roughness and measurement data of CNC turning to develop a modeling approach which does not depend too much on data, but also effectively uses existing analytical models. As in previous researches, we use cutting speed, feed rate, depth of cut and three acceleration components from an accelerometer to predict surface roughness. Co-Kriging method is employed to integrate the above measurements and a well-known model of surface roughness in turning. It was confirmed that the approach improved the prediction accuracy when only small amount of data is available for model construction. Meanwhile, the accuracy of ordinary Kriging method, which only depends on data, is suitable when measurement data sufficiently spans the parameter space, being expected that it may be rare in actual operations. We also attempted to detect outlier of measurements using the Co-Kriging method, which might be a non-trivial task when there is no additional information to evaluate the validity of the measurement data.