Data mining for fault diagnostics

In manufacturing processes the automated identification of faulty operating conditions that might lead to insufficient product quality and reduced availability of the equipment is an important and challenging task. This paper proposes a data mining approach to the identification of complex faults, i.e. unplanned machine stops in plastic injection molding. Several data mining methods are considered, with a focus on the abilities to reveal patterns of faulty operating conditions and on the interpretation of the induced models with the objective to find the data mining method that best corresponds to the nature of the plastic-injection-molding process and the related data. Well-known data mining methods, i.e. J48, random forests, JRip rules, naïve Bayes, and k-nearest neighbors are applied to real industrial data. The results show that tested data mining methods can be effectively used to reveal patterns related to faulty operating conditions. The interpretation capacity of the tested methods, their ability to describe the operating conditions, and to reveal patterns related to faulty operating conditions, are demonstrated and discussed