Improving the manufacturing system of anti-scatter grids at Philips
Ever heard of the DRC-FJSSP-SSO? This ‘Dual Resource Constraint Flexible Job Scheduling Problem with Scarce Setup Operators’ is at the heart of Bas Giesen’s thesis about improving the manufacturing system of anti-scatter grids at Philips.
At the 1D grids factory of Philips in Best, anti-scatter grids are produced that can be used in Xray scanners. With these grids, less radiation is required, while image quality is improved simultaneously. However, production comes with certain challenges, such as low yield, limited machine capacity and the need for skilled operators.
Where it started
At the start of the thesis project, 90 percent of all grids were completed within 29.6 days. Nonetheless, in combination with a large backlog, on time delivery was not at the desired level. By researching the planning and scheduling of the anti-scatter grids, Bas Giesen’s aim was to increase delivery performance to 95 percent. To do so, the production line was divided into three stages:
- Pre-processing. An inventory policy is implemented to guarantee a 99% fill rate of the dry fiber, as the dry fiber is an essential raw material that is used at the start of production. The inventory policy was calculated by means of an event simulation.
- Processing. This stage is seen as the aforementioned DRC-FJSSP-SSO, meaning that it contains a scheduling problem in which a job is produced on a machine, and set up by an operator. Both the machine and the operator must be suitable and available for the job. A Mixed-Integer Linear Program (MILP) and Genetic Algorithm (GA) were used to solve the problem and come to the most efficient solution. The MILP creates optimal solutions, but is computationally expensive, which is why GA is commonly used.
- Post– The capacity per processing step is calculated by determining the amount of time an operator is usually processing a single product. By doing this, the required time per product can be verified and therefore outflow of stage 2 (processing) and inflow of stage 3 (post-processing) can be matched.
As mentioned before, 29.6 days were needed to complete 90% of all grids with the previous scheduling method. However, the improved method decreased this to 22.3 days. The improvement is purely based on having the right number of operators at the right processing step and releasing a new production order once a grid is scrapped. The computation time of the GA was too high, but learnings of the scheduling mechanisms could still be derived. For example, the algorithm tried to schedule the same height and line type on the machines. As changing heights/line type requires a new set up by one of the operators, this is very rational.
With the recommendations, a delivery period of six weeks can be guaranteed. However, due to a big backlog on previous orders, this can not be said with certainty for the coming weeks. Full capacity of operators is needed to reduce the backlog, after which on-time delivery can meet the targets.
In summary, Philips should take reorder levels of dry fiber into account to prevent a production stop. Further, three operators are required in the last step of production, to match outflow of stage 2. And lastly, setups have to be considered. While these advices are very specific for this production line, the methodology is generally relevant. Focusing and improving specific parts of the chain, can lead to significant increases.