Unveiling Operational Dynamics: Causally Interpretable Clustering for Understanding the Usage of Plastic Injection Molding Machines

With the increasing adoption of Machine Learning in industrial applications and state-of-the-art implementations predominantly based on effective black-box models such as deep learning, there is a rising interest in utilizing such models to gain operational insight and facilitate knowledge discovery. Gaining operational insight is challenging in complex industrial processes such as plastic injection molding due to many context parameters (e.g. machine types or product requirements) and domain knowledge such as the intricacies of plastic’s thermo-viscoelastic properties and the nature of its operation. As a result, adjusting the process is mainly based on the engineers’ previous knowledge and experience. To reduce these high barriers of learning, this paper presents a novel approach for interpretable clustering of plastic injection molding operational dynamics to gain insight into their operation and configuration characteristics. Going beyond associational methods and statistical properties of usage data, this approach utilizes Causal Discovery to understand the underlying process mechanisms. Based on the causal graphs, similarly, set-up processes are clustered and characterized through statistical, graph-based, and explainable AI methods. Utilizing real-world usage data, this approach resulted in valuable knowledge of machine-, business-, and process-related differences in the operational dynamics of the plastic injection molding process. This novel approach shows potential for applications in a broader context across other fields of science and engineering, where it can be used to address the ambiguity of existing clustering approaches by clustering on generative causal mechanisms and providing causally interpretable cluster characteristics.