Abstract: In the assembly pull production mode, the discrete manufacturing workshop, which is characterized by decentralized production tasks, flexible equipment layout, and complex processes, serves as the core carrier of the multi-factory(shop) production mode. However, the dynamic drift of bottleneck units in both temporal and spatial dimensions has become a critical challenge that limits the improvement of production efficiency and resource utilization. Thus, investigating the bottleneck prediction problem in discrete workshops is crucial for enhancing overall production efficiency in multi-factory settings. To address this, a spatiotemporal network prediction model integrating dual attention mechanisms, CNN-LSTM-DAM (Convolutional neural network-long short term memory-dual attention mechanism), is proposed. First, a composite definition-based bottleneck identification model is constructed, considering the multi-attribute coupling of bottleneck units. Second, the identified historical quasi-bottleneck data are input into the spatial feature perceiver integrated with CNN and spatial attention mechanism, as well as the temporal feature perceiver integrated with LSTM and state attention mechanism, as auxiliary data to further enhance the model's ability to capture spatial and temporal information in production sequence data. Finally, through ablation experiments, the proposed model is compared with other LSTM variants, such as gated recurrent unit (GRU) and bidirectional long short term memory (BiLSTM). The results verify the model's accuracy and effectiveness in predicting bottleneck units and their drift trends within a given time delay.

Key words: bottleneck prediction, multi-attribute bottleneck, dual attention mechanism, CNN-LSTM, multi-step time series forecasting