Evaluating mixed-model assembly line sequencing heuristics for just-in-time production systems

• Published in: Journal of operations management 1990-08, Vol.9 (3), p.371-390
• Main Authors: Sumichrast, Robert T., Russell, Roberta S.
• Format: Article
• Language: English
• Subjects: Assembly lines; Effectiveness; Integer programming; Just in time; Mathematical models; Methods; Theory
• ISSN: 0272-6963; 1873-1317
• DOI: 10.1016/0272-6963(90)90161-6

When mixed-model assembly lines use components fabricated in-house, the demand for these components is not uniform over time and is affected by the sequence of models on the assembly line. Thus, without proper mixed-model sequencing and the subsequent smoothing of component demand, the effectiveness of a just-in-time (JIT) production system is limited. This paper focuses on making component usage uniform. Five sequencing methods are reviewed, two suggested by Monden (1983) and used at Toyota (GC1 and GC2), and three proposed by Miltenburg (1989) (M-A1, MA3H1, and M-A3H2). Their performance is evaluated for the special case when all models use the same components. For all of the sequencing methods tested the mean absolute deviation of model usage varies directly with the number of models produced. There is no clear relationship between the mean absolute deviation of model usage and demand pattern or length of production sequence. Method M-A3H2 produces the highest quality feasible solutions under all conditions tested. The relative performance of the methods does not appear to be related to the number of models, demand type, or length of production sequence. To compare methods for the more general case of different models requiring different components, a mixed integer programming (MIP) model is presented as a way to find an optimal sequence. The M1P creates optimal solutions but is too slow to be used in practice. The two goal chasing heuristics used by Toyota were considered on the basis of their ability to schedule production to use components linearly over time. It is shown that these methods differ widely in their ability to generate good sequences. The performance of both is best when the products assembled have simple product structures. When models require more than one of a given component or when models require many different components, the performance worsens. The difference in performance is very small for GC1, but significant for GC2.