Investigation of failures in rotational moulding using historical production dataset and machine learning

Abstract Rotational moulding (RM) is a versatile manufacturing process widely used for producing lightweight, seamless plastic components, but its potential is often constrained by challenges in optimizing production parameters for diverse product geometries and simultaneous batch production. This study addresses the pressing need for a data-driven approach to enhance RM efficiency and reduce defects under non-optimal process conditions. Leveraging historical production data from a medium-sized RM enterprise, an Ensemble Learning-based machine learning (ML) model was developed to predict failure probabilities across 390 product-process combinations. Input parameters are heating temperature, speed, mould volume, product mass. The model achieved an accuracy of 97.17%, identifying optimal parameter ranges for minimizing defects. The results revealed that deviations between machine and product-specific conditions, particularly in heating temperature and rotational speed, significantly increased failure probabilities. Products with intermediate sizes and masses were most susceptible to failures, while extreme values of mould volume occupancy showed a lower likelihood of failures. Notably, the study highlighted the critical importance of maintaining minimal delta heating temperature and speed ratio disparities to ensure product quality. This approach offers a robust framework for optimizing RM processes without costly sensorization, making it especially beneficial for small- and medium-sized enterprises.