A Parallel Gated Recurrent Units (P-GRUs) network for the shifting lateness bottleneck prediction in make-to-order production system

Abstract In the make-to-order production system, the lateness bottleneck is the constraint of just-in-time management and orders on-time delivery. Since the dynamic nature of the manufacturing system, the bottleneck frequently shifts and influences the stability during the production runs. Therefore, predicting the bottleneck allows operators to foresee the future production status and to make proactive decision towards a balanced-line. Based on the large volumes of manufacturing data collected by Internet of Things (IoT), a novel Parallel gated recurrent units (P-GRUs) network with main inputs and auxiliary inputs are particularly developed for shifting bottleneck prediction. The designed P-GRUs can capture the temporal correlations of shifting bottlenecks and depict the production status simultaneously to make accurate bottleneck prediction. The P-GRUs model is applied in a large-scale production system to validate the performance and demonstrate the practical impacts. Finally, the experiment results from both real-world production as well as simulation environment show that the P-GRUs model yields better performance than benchmark models, including Autoregressive integrated moving average model (ARIMA), vanilla Recurrent nueral network (RNN), Deep neural network (DNN), and regular GRUs network.