Manufacturing Lead Time Estimation with the Combination of Simulation and Statistical Learning Methods

Abstract

In the paper, a novel method is introduced for selecting tuning parameters improving accuracy and robustness for multi-model based prediction of manufacturing lead times. Prediction is made by setting up models using statistical learning methods (multivariate regression); trained, validated and tested on log data gathered by manufacturing execution systems (MES). Relevant features, i.e., the predictors most contributing to the response, are selected from a wider range of system parameters. The proposed method is tested on data provided by a discrete event simulation model (as a part of a simulation-based prediction framework) of a small-sized flow-shop system. Accordingly, log data are generated by simulation experiments, substituting the function of a MES system, while considering several different system settings (e.g., job arrival rate, test rejection rate). By inserting the prediction models into a simulation-based decision support system, prospective simulations anticipating near-future deviations and/or disturbances, could be supported. Consequently, simulation could be applied for reactive, disturbance-handling purposes, and, moreover, for training the prediction models.