Using Bayesian Networks to forecast spares demand from equipment failures in a changing service logistics context

A problem faced by some Logistic Support Organisations (LSOs) is that of forecasting the demand for spare parts, corresponding to equipment failures within the system. Here we are particularly concerned with a final phase of operations and the opportunity to place only a single order to cover demand during this phase. The problem is further complicated when the service logistics context can change during this final phase, e.g. as the number of systems supported or the LSO's resources change. Such a problem is typical of the final phase of many military operations.

The LSO operates the recovery and repair loop for the equipment in question. By developing a simulation of the LSO, we can generate synthetic operational data regarding equipment breakdowns, etc. We then split that data into a training set and a test set in order to compare several approaches to forecasting demand in the final operational phase. We are particularly interested in the application of Bayesian network models for this type of forecasting since these offer a way of combining hard observational data with subjective expert opinion.

Different LSO configurations were simulated to create a test dataset and the simulation results were compared with the various forecasts. The BN that learned from training data performed best, followed by a hybrid BN design combining expert elicitation and machine learning, and then a logistic regression model. An expert-adjusted exponential smoothing model was the poorest performer and these differences were statistically significant. The paper concludes with a discussion of the results, some implications for practice and suggestions for future work.