A Probabilistic Digital Twin for Improving Production Scheduling

In the manufacturing sector, production plans heavily depend on knowing how long each work order will take. Typically, a Manufacturing Execution System (MES) estimates the working time of every order once, by giving it a cycle time (in terms of seconds per piece) before production starts, and never updates it, although the machines’ own data (the so-called SCADA records) show how fast each order is really running. This thesis focuses on this problem, designing a method to predict, while an order is running, how many working hours it still needs, and how certain that prediction is. The prediction is made at ten checkpoints during each order, defined as shares of the ordered quantity. Three models based on gradient-boosted decision trees are trained for this purpose. All of them learn a correction to the MES plan, using the plan itself, the progress of the order so far, and the history of the article and the machine, and they differ only in what they aim at. The first is trained with the absolute-error loss and returns the most likely value, the median. The other two are trained with the quantile loss and return a lower and an upper bound, the 10th and the 90th percentile. Each prediction therefore arrives as a range with a most likely value inside it. The evaluation uses only information available up until the moment of the prediction, and tests the model on orders that started after all the orders it learned from. The per- formance evaluation results show that, on 680 test predictions from seven achines, the proposed framework reduces the error of the MES plan by 48% and 37% in terms of hours and absolute-error loss respectively, and is closer to the ground truth than the MES plan in 73.5% of the cases. Much of the gain comes from correcting the plan’s systematic optimism. Based on such results, we claim that this work provides the predictive module of a digital twin of a manufacturing plant, that is, an estimate of the remaining working time, with its uncertainty, that a planner or a scheduler can use.