Reliable machine learning models for manufacturing processes

Abstract Data-driven condition monitoring in manufacturing using machine learning (ML) offers great potential for operating processes reliably and cost-effectively. Particularly in blanking, a manufacturing process characterized by high production rates combined with narrow tolerances, data-driven process monitoring techniques are necessary to ensure a reliable operation. However, an inherent weakness of these data-driven techniques is their sensitivity to changes in the system configuration. Such changes directly affect the data validity and lead to shifts in the underlying data distribution, resulting in a dramatic deterioration of the model’s ability to predict a target variable. Therefore, the goal of this work is to introduce the deep hybrid modelling (DHM), which enables reliable ML-based condition monitoring even when the boundary conditions of the system fluctuate and lead to random process deviations. The potential of the introduced strategy is demonstrated by predicting the abrasive wear state during a blanking operation with the simultaneous occurrence of a data shift due to a random fluctuation of the semi-finished product parameters. In addition, a similarity measure is introduced that allows a user-friendly and computationally efficient quantification of the data shift and the associated magnification of the occurring fluctuations in the process boundary conditions. To ensure a reliable ML-based prediction depending on the magnitude of the data shift, different modelling frameworks are introduced. For this purpose, a resilient modelling strategy based on DHM and a flexible modelling strategy based on domain adaptation technique called adaptive DeepCORAL are proposed.