Efficient manufacturing processes and performance qualification via active learning: Application to a cylindrical plunge grinding platform

Abstract

The industry invests significant resources towards qualification of its individual processes and machines to assure quality and productivity of the process chain. Process qualification traditionally involves employing elaborate experimental methods to find the response surface mapping the response to various process parameters and measurements. Most of the existing methods are passive experimental designs which take into account the limits of the parameter space and a design method (CCD, Taguchi, orthogonal etc.) to identify the points in the parameter space. More often than not, these methods need a lot of experiments to be conducted and do not take into account how the response changes with each experiment. Also, the number of experiments increase combinatorically to get a desired estimate of the response surface. The formulation of mathematical models for complex, high dimensional, inherently nonstationary, and stochastic systems like abrasive machining process and also catering to the process-machine interactions is challenging. In this work, to address the other alternative for cost-effective experimentation: we adapt a Query by Committee (QBC) based active learning approach where we sequentially find the next best experimental point to reduce the uncertainty of prediction of surface roughness over the sample space. The method uses a carefully curated list of committee members, (i.e., models) which predict the response surface at each instant and selects the next experimental point based on a metric called prediction deviation. We used a real-world dataset from a cylindrical plunge grinding platform to test if the QBC approach performs better than a passive CCD design. The machine tool used is the next generation precision grinder (NGPG) from IIT Madras which is capable to finishing components to an IT3 tolerance grade. We compared the QBC based active learning model to a previous random forest model built on a dataset which gave a test accuracy (R2) of 85% using 178 experimental points. It is demonstrated that similar prediction accuracies can be achieved by reducing the number of experiments by about 65%. The merits of the model in the choice of the members of the committee and the advantage of the current experimental design compared to random experimentation were presented.