Data-Driven Inventory Policy Assignment in ETO Environments Using Fuzzy K-Prototypes Clustering

In engineer-to-order (ETO) manufacturing environments, the high variability of final product configurations makes it difficult to consistently estimate material consumption and, consequently, to define appropriate inventory control policies. This paper proposes a data-driven framework based on unsupervised learning to identify product typologies from historical manufacturing orders in a real industrial context. The approach employs a fuzzy k-prototypes algorithm to cluster mixed-type data, allowing the simultaneous treatment of numerical and categorical variables. In the case study, the proposed crisp-BOM-based scenario achieved a 28.67% reduction in line-side WIP and a 10.79% reduction in linear storage space, corresponding to the release of approximately two to three assembly stations. From the resulting fuzzy memberships, probabilistic bill of materials (BOM) structures are constructed, capturing the inherent variability of material consumption across different product configurations. A defuzzification procedure is then applied to obtain a crisp BOM representation suitable for operational decision-making. Additionally, a material versatility indicator based on entropy is introduced to quantify the dispersion of each material across product typologies. This indicator, together with the estimated consumption per cluster, is used as input for an analytical inventory model that supports the classification of materials into kanban or kitting policies. The methodology is validated using real data from a high- and medium-voltage switchgear manufacturing plant, comprising over 60,000 order–material observations. The results show that the proposed framework enables a more structured characterization of material behavior, reducing reliance on planner experience and improving the consistency of inventory policy decisions. From an industrial perspective, the approach provides a practical and scalable tool for aligning inventory strategies with the actual consumption patterns of ETO systems.