Remaining time prediction in manufacturing systems: an approach based on ML and process mining

Abstract

The remaining time prediction of production orders in the manufacturing domain is of major concern among production, planning, and control (PPC) managers. PPC managers must deal with significant uncertainty regarding the promise of delivering products to customers. Many techniques use data to predict the remaining time of production orders, such as neural networks, time series analysis, and non-parametric statistical models, among others. A powerful way to deal with these new machine-based data records is through process mining techniques, which can summarize and collect information about the underlying process based on event logs. This paper proposes a hybrid predictive model based on annotated transition-systems and machine learning models tailored to better predict ongoing production orders in industrial manufacturing environments. The linear combination of models is performed by optimizing a linear programming (LP) model that minimizes the combined absolute errors of predictions. We tested our new approach on artificially created logs. Results showed that our approach provides better accuracy measures than all the other tested methods for the test instances.