2. Model
We first developed a discrete-event simulation model using SIMIO. While most simulation models’ focus is on products flow, our need was to model sequences of tasks: some performed by workers, some by workers and machines, others simply by machines. A conceptual representation of the model is given in Figure 1. The plant has 21 CNC machines (c), 31 workers with different expertise, (d) four measurement machines (f) including two coordinates measurement machines (CMM), one optic comparator and one manual optic comparator. From the APN production schedule (a) we can extrapolate a list of tasks: parts machining (blue), measurement (red) and others (yellow). Parts are machined on CNCs. At the same time, other tasks are accomplished by the workers on a worktable (b), although some may require the CNC to stop. After leaving the work center (e), some parts need to be measured for quality control. This is a complex process that involves many machines and workers. If a measure is out of specification, a corrective action task (e.g. tool change) is created and inserted first place into the waiting queue.
The different aspects of the system model were validated by interviews with experts working at APN in order to confirm that the model is representative of field reality. Expected duration distribution for the production tasks were extrapolated from the manufacturing execution system (MES) historical database. However, measurement and quality control come from “educated guess” by the process engineers.
Figure 1: Conceptual representation of the simulation model.
3. Experiments and Results
Different policies were compared regarding how they affect productivity (total time needed to carry on a production schedule) and the workers total traveled distance. The base case scenario is called OneWorker. One worker is allocated to one CNC. Once his shift is over, another worker takes over. Under the ListWorkers scenario, we are aware each worker is trained for which types of machines and tasks. For each task, we dynamically allocate it the closest free and compatible worker. As an upper bound, we also simulated the utopic AllWorkers scenario, where each worker would be trained for all machines and tasks. The following are the average results from 100 replications (95 % confidence intervals) of a single given production schedule (data extraction was very time consuming).
Figure 2: (a) Production time (in Hours) and (b) Workers total distance traveled (km).
On average, dynamic allocation (ListWorkers) reduces total production time by 40.4 % (see Figure 2). What is even more surprising (and explains the previous) is by how much the confidence interval shrink too. Indeed, under the ListWorkers scenario we are much less affected by the stochastic nature of the processing times. The AllWorkers scenario leads to additional improvements, but it supposes important training costs. As for the walking distance (Figure 3), it is very similar for OneWorker and ListWorkers. AllWorkers greatly impacts walking distance.
