Estimating requirements from experience
Planners at BSH and other manufacturers traditionally decide on the quantities of spare parts to be stocked based on detailed knowledge of the items and the associated empirical values. In doing so, they draw on their experience from previously recorded product life cycles and consider previous sales of the respective spare part. However, this process is time-consuming, as all relevant information has to be gathered from different interfaces and systems. In addition, high safety stocks are traditionally calculated due to high uncertainties in the estimation. BSH has therefore commissioned the Supply Chain Services working group to find a data-driven solution that can provide automated and reliable support for the decision.
»Clean« master data as the basis
During the project, it quickly became clear that the task could not be solved trivially, even in a data-driven manner, for various reasons: Firstly, there were too few consistent data sets on which the analyses could be based, so master data had to be filled in using clever linking and heuristics. In addition, the classic methods for predicting time series were not applicable. These do not mathematically map the pattern of a demand curve but rather extend the trend of spare parts requirements linearly. The dispatchers at BSH therefore initially supported the AI experts from Fraunhofer with their specialist knowledge so that a “clean” database could be created together.
Reliable forecast using clustering
Based on this database, the researchers developed a new forecasting method for spare parts by deriving typical sales curves from various clusters of similar spare parts. If the need or demand for a new spare part is to be predicted, it is assigned to a specific life cycle cluster based on its master data. This clustering was repeated iteratively so that the data-driven forecast is now a valuable aid for dispatchers.
Development of a data-driven forecasting tool
To make the forecasting method developed here easy to use, the researchers developed a graphical interface that displays the most important decision-relevant data for a specific spare part in addition to the forecast. This forecasting tool can also be used by dispatchers to generate reports on all spare parts. For example, the system can also identify potential for scrapping by identifying items where future demand is unlikely. This tool has proven its worth over the last two years and everyone involved is working on expanding it - however, the final decision on the spare parts inventory always remains with the responsible planners.
Optimization of the storage quantity
Based on a precise demand forecast, a cost-optimized ordering and production strategy can be derived. The costs of an ordering strategy are made up of two conflicting components. On the one hand, a high stock level is associated with high storage costs or even scrapping costs if too many parts are stored. On the other hand, the unit costs decrease with the order quantity, so that large one-off orders are more cost-effective than many individual orders. Using mixed-integer programming techniques, the Fraunhofer Working Group for Supply Chain Services at Fraunhofer IIS calculates a cost-optimized ordering strategy that finds the best possible compromise between the two cost factors and at the same time always covers the forecast requirements. A particular challenge here is the erratic behavior of ordering costs with increasing quantities. However, thanks to clever modelling techniques, an optimized ordering strategy can be calculated in just a few seconds. This allows the optimization software to be used in real time. In a further research project, the optimization even considers uncertainties in the demand forecast to minimize the risk of uncovered demand.
Division Supply Chain Services at Fraunhofer IIS