Selecting the best clustering variables for grouping mass-customized products involving workers' learning

Clustering of product models is an important technique in highly customized production environments, where variety is a key competitive dimension. When the goal is to create product platforms, models are usually grouped based on parts similarity in terms of morphology and demand, and integer programming has been typically used for that. However, product groupings that yield efficient platforms not necessarily optimize the operation of manual assembly lines used to produce them. In this paper, the goal is to cluster product models with similar processing needs into families, such that an efficiency of production programming and resources allocation are maximized when products are obtained though manual assembly operations. It is known that clustering results are highly dependent on the proper choice of clustering variables. To address that problem, we propose a method to select the best clustering variables aimed at grouping customized product models into families. Two groups of clustering variables are considered: those generated by an expert assessment on product features that may impact on productivity, and those representing workers' learning rate, obtained through the learning curve modeling. The method integrates the "leave one variable out at a time" elimination procedure with a k-means clustering technique. When applied to a shoe manufacturing process, the proposed method uses only 2 out of 12 candidate variables and increases the grouping quality, measured by the Silhouette Index, to 0.89 from 0.40.