Dynamic production bottleneck prediction using a data-driven method in discrete manufacturing system

In the complex discrete manufacturing system (DMS), the production bottleneck shifts in space as time goes on and constrains operational efficiency. Accurate proactive production bottleneck prediction provides a reliable basis for dynamic production decisions and helps to improve management timeliness and production efficiency. According to the production characteristics of DMS and the relationship between supply and demand, the production bottleneck is given a new quantification. A long and short-term memory network (LSTM) with dual attention mechanism and a dynamic updating method for the source model are proposed to predict production bottlenecks accurately. Firstly, feature and state attention mechanisms are designed to improve the feature extraction and prediction ability of LSTM. Secondly, as the applicability of the prediction model gradually declines over time, sliding time windows and fast Hoeffding concept detection are combined to trigger the update of model parameters. Then a competitive strategy is explored to choose the source model that is the most suitable for the current data distribution in the model library. Model-based transfer learning is adopted to update the source model parameters, making the prediction model highly adaptive. Subsequently, an elimination strategy is set to update the model library to ensure its timeliness. Finally, experiments demonstrate that the proposed method is effective in bottleneck prediction and superior to other methods.