Life cycle prediction for spare parts at IBM spare parts operations

Introduction: The research is executed at one of the departments of IBM SPO, being Life Cycle Planning. This department is, amongst other things, responsible for calculating the need to cover the usage of a Field Replaceable Unit (FRU) over the remaining service period (RSP) in case a supplier stops producing and IBM has a final opportunity to acquire parts, known as a Last Time Buy (LTB). Due to the time horizons, that can be up to 16 years, the amount of parts to acquire is difficult to forecast. Motivation: Based on a pilot project executed for the division Lenovo, focused on Commodity Based Lifecycle Forecasting (CBLF), the idea existed that there should be a commonality between the FRUs with a similar usage pattern, which can be used for clustering. Having the opportunity to assign an FRU to a cluster would make long term forecasting easier, since the usage pattern is known. Therefore the aim of this research is the following: “Investigate which characteristics of an FRU are related to a specific usage pattern and how this information can be used to cluster FRUs into groups with a similar usage pattern, in order to improve the forecast accuracy.” Research methodology: The start of the research was the determination of the current forecast accuracy and the different types of usage patterns the FRUs follow. After the FRUs were assigned to a specific usage pattern, an investigation of the relation between the usage pattern and a set of characteristics was executed, to determine whether characteristics can be related to a specific usage pattern and can be used to cluster the FRUs. Therefore we used a combination of statistical testing, data analysis and factorial analysis. As a second step, we assessed the performance of the pilot that triggered the research, and we discussed and tested options to improve this method, based on a simulation of historical LTBs. Results: Based on the analysis of the usage patterns, 12 different partial usage patterns were identified over a period of 5 year historical usage, because that is the amount of years for which historical data is stored. It appeared to be impossible to combine these partial usage patterns into a limited set of usage patterns, due to the large amount of possible combinations. With respect to the forecast accuracy, the performance is determined based on the bias, Mean Absolute Deviation (MAD) and the Mean Absolute Percentage Error (MAPE). The results indicated that the standard decline approach, in which a fixed decline factor is used for every year the need has to be forecasted, has an aggregated MAPE value of 235% against 314% for CBLF. As a result, standard decline leads to more accurate forecasts on an aggregated level. CBLF has a more accurate result when it is actually the best approach, with an average MAPE of 203% for CBLF compared to an average MAPE of 241% for the standard decline. However, the results of both approaches indicate room for improvement, because the average aggregated difference between the forecasted and actual need is more than 50%, which indicates an inaccurate forecast. With respect to identification and clustering we investigated the possibilities for 8 characteristics, being Age FRU, Brand, Commodity, Division, Forecasted Reliability, LTB Month forecast, RSP and TM144 status (indicates if the LTB is executed before or after production stops). On an aggregated level, statistical tests indicate that all characteristics could have a relationship with a partial usage pattern. On commodity level, with a commodity being a group of FRUs with similar purpose (like batteries), some of the characteristics can be excluded based on the statistical tests. We focused on clustering approaches for the commodity HDD, based on a combination of ranges, sub ranges and historical usage data. The amount of FRUs selected using the clustering approaches ranges in most cases between 1 and 5 FRUs, resembling less than 5% of the entire selection. The percentage of FRUs successfully identified can be high, but is not considered to be representative, since 1 FRU correctly identified from a selection of 1 can also be coincidence. In the second part of the research, we focused on the performance of CBLF. The analysis pointed out three main causes for an inaccurate forecast with CBLF, being a difference between the observed usage pattern and the pattern used to forecast future usage with, a difference between the expected and actual time on the market and a difference between the point in time the observed usage starts and the point in time in which the usage starts based on the curve used to forecast future usage. Different improvement options are tested to minimize the effects of the main causes and all options have realized accurate forecasts for at least 1 FRU in the selection, but a straightforward method to determine the most appropriate forecast method for all FRUs could not be developed. Of the different forecasting options considered, the aggregated MAPE values for a selection of 13 fast moving FRUs are 63% for Weibull, 71% for Gamma and 133% for standard decline. For slow movers, the aggregated MAPE values for a selection of 6 FRUs are 78% for scaled CBLF, 114% for Croston’s method and 139% for standard decline. These methods are combined in an Excel tool, that can be used to visualize the possible usage patterns an FRU might follow and indicate what ranges of usage might be realized. This could help the SPO team in making a decision about the amount to acquire, but can also clarify the advantages that range forecasting can offer.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.