Abstract: Discrete manufacturing commonly faces bottleneck issues that affect production efficiency. Traditional static bottleneck identification methods cannot effectively resolve the dynamic bottleneck shifts in complex remanufacturing environments. To address this, it proposes a dynamic bottleneck prediction model termed LMD-QPSO-LSTM, which integrates Local Mean Decomposition (LMD) method with Long Short-Term Memory (LSTM) networks and utilizes an enhanced Quantum Particle Swarm Optimization (QPSO) algorithm to fine-tune LSTM. Initially, machine energy consumption attributes are used to define the dynamic bottleneck index, and LMD method is employed to decompose bottleneck sequences, thereby reducing data volatility. Subsequently, an Attention Mechanism (AM) is integrated to improve the LSTM network′s learning capability, alongside the use of an enhanced QPSO algorithm to optimize LSTM parameters for optimal selection. Finally, the components of the bottleneck index are predicted, and the prediction results are reconstructed. Simulation experiments demonstrate that the dynamic bottleneck prediction method based on LMD-QPSO-LSTM effectively enhances prediction accuracy and accurately tracks changes in bottleneck positions. Compared to other models, this approach reduces the Mean Absolute Error (MAE) by at least 52.63 %, the Mean Absolute Percentage Error (MAPE) by 25.14 %, and the Root Mean Square Error (RMSE) by 45.78 %.

Key words: local mean decomposition, long short-term memory network, improved quantum particle swarm optimization algorithm, dynamic bottleneck prediction, bottleneck drift