Combination of Machine-Learning Algorithms for Fault Prediction in High-Precision Foundries

Foundry is one of the activities that has contributed to evolve the society, however, the manufacturing process is carried out in the same manner as it was many years ago. Therefore, several defects may appear in castings when the production process is already finished. One of the most difficult defect to detect is the microshrinkage: tiny porosities that appear inside the casting. Another important aspect that foundries have to control are the attributes that measure the faculty of the casting to withstand several loads and tensions, also called mechanical properties. Both cases need specialised staff and expensive machines to test the castings and, in the second one, also, destructive inspections that render the casting invalid. The solution is to model the foundry process to apply machine learning techniques to foresee what is the state of the casting before its production. In this paper we extend our previous research and we propose a general method to foresee all the defects via building a meta-classifier combining different methods and without the need for selecting the best algorithm for each defect or available data. Finally, we compare the obtained results showing that the new approach allows us to obtain better results, in terms of accuracy and error rates, for foretelling microshrinkages and the value of mechanical properties.