Author: Wei Wu (National Cheng Kung University) - This study aims to combine deep learning algorithms with CFD simulation to establish a predictive maintenance strategy for shell-and-tube heat exchangers. First, Ansys Fluent was used to establish a deposition simulation model, and the discrete phase model (DPM) was used to simulate scaling areas and deposition distribution of deposition particles inside the heat exchanger. Moreover, the simulation results were combined with historical operating data from industry to build a deep learning prediction model, and time series algorithms such as long short-term memory networks (LSTM) and gated recurrent units (GRU) were introduced to predict future heat transfer performance and key operating indicators. In order to improve the performance of this physics-based AI model, this study designed multiple sets of variable combinations to explore the impact of combining historical data and simulation data on prediction accuracy. The results showed that the proposed physics-informed deep Learning strategy can effectively reduce noise interference and measurement errors, and significantly improve the model convergence speed and prediction ability. Among them, the GRU model performed best, with an RMSE of 0.0035 and a MAE of 0.0027. The coefficient of determination R² reached 0.9979, indicating that its structure is clean and its learning efficiency is better than LSTM. Finally, the warning system constructed by the completed model can accurately predict the heat exchange performance in the next 12 hours, and issue instant alarms for each level of warning values. This study can be successfully applied to real-plant heat exchange equipment to help operators grasp the best maintenance time, avoid unnecessary repairs and sudden damage, and improve the overall equipment operation efficiency and reliability.
