An approach to learning from both good and poor factory performance in a Kanban-based just-in-time production system

In a JIT manufacturing environment it may be desirable to learn from an archived history of data that contains information that reflects less than optimal factory performance. The purpose of this paper is to use rule induction to predict JIT factory performance from past data that reflects both poor (saturated or starved) and good (efficient) factory performance. Inductive learning techniques have previously been applied to JIT production systems (Markham et al. , Computers and Industrial Engineering, 34 , 717-726, 1998; Markham et al. , International Journal of Manufacturing Technology Management, 11 (4), 239-246, 2000), but these techniques were only applied to data sets that reflected a well-performing factory. This paper presents an approach based on inductive learning in a JIT manufacturing environment that (1) accurately classifies and predicts factory performance based on shop factors, and (2) identifies the important relationships between the shop factors that determine factory performance. An example application is presented in which the classification and regression tree (CART) technique is used to predict saturated, starved or efficient factory performance based on dynamic shop floor data. This means that the relationship between the variables that cause poor factory performance can be discovered and measures to assure efficient performance can then be taken.

Keywords:

• Rule Induction
• Classification And Regression Trees (CART)
• Kanban
• Just-in-time