Machine Learning models to forecast defects occurrence on foundry products

Manufacturing defects negatively affect production cost and environmental impact. This impact is even heavier for production processes that are particularly energy intensive, such as in the metallurgical industry. In the cast iron foundry sector, components manufacturing can be affected by various defects that depend on quality of raw materials fed to the melting process, process parameters, cast iron final temperature, ferroalloys additions and final composition. Defects formation is a multifactorial phenomenon, for which relevant factors are not easy to identify, as it is also a rare event. This paper presents a set of models based on machine learning methodologies for predicting and classifying defects on foundry production lines to assist process operators in managing the process and, eventually, stop production when it is not possible to adjust process parameters. The modelling phase exploits decision trees methodologies enhanced with algorithms for augmenting unbalanced datasets related to defects occurrences. The combination of these methodologies produces efficient models showing very encouraging results.