Throughput maximization by dynamic worksharing in unbalanced and multistage production lines

We consider a serial production system with manual or low capital cost operations (tasks). We assume the set of required tasks have been ordered into a production sequence and tasks have been assigned to workstations. As a result of the required cycle time and discrete operation times, the line may be unbalanced leading to poor resource utilization. Variability in operation times from unit to unit may exacerbate this problem. To improve productivity, we consider partially cross-training workers, allowing small inter-stage buffers, and using dynamic line balancing. Each work/workstation is assumed to have a set of tasks that must be performed at that workstation. However, workers are cross-trained such that one or more tasks can be performed at either of a pair of adjacent workstations. Thus each worker at an interior workstation has a set of a, b, and c tasks for which they are trained. When a unit arrives at their workstation, tasks of type a must be done if they were left undone by the upstream worker. Type b tasks are always done at the workstation. After completing the required a and b tasks, the worker decides whether to complete one or more of the c tasks or pass the unit downstream. Rules are proposed and evaluated for guiding real-time worker decisions concerning whether to continue on the next task or to pass the unit downstream. We refer to this as dynamic line balancing (DLB). Two types of DLB models are discussed. In the first case we assume only the number of units in the downstream buffer is known to the worker. A threshold based heuristic that attempts to avoid starvation is found effective. In the second case we assume the worker knows the total work content in the downstream buffer and in process at the downstream station. The Half-Full Buffer rule from the literature is adopted to this situation and shown to be effective. We examine both unbalanced and balanced task divisions in two-stage models. The proposed DLB rules are shown to be effective in improving line productivity while requiring only low to moderate levels of work-in-process inventory. DLB is able to achieve line efficiency that is very close to the theoretical limit. For task divisions that are moderately unbalanced, DLB is able to recover line throughput to that of the balanced case. We also examine the effect of the shared task size and the number of decision points (granularity) for the worker. For the non-preemptive case with a single decision point, the existence of an optimal shared task size, usually half the size of the fixed tasks, is witnessed. However, this is not the case for granular shared tasks, where larger total time for the shared task implies higher productivity. The two-stage heuristics are extended for use in general multi-stage lines. Experimental results are given for models of three and five stages. Results are positive and parallel the two-stage results.