Hybrid particle swarm optimization and physics informed neural network algorithm for temperature field reconstruction in industrial factory fires

作者：Yan Li, Bin Sun · 发表于：Journal of Building Engineering · 年份：2025 · DOI：10.1016/j.jobe.2025.114203 · 被引用次数：4 · 研究领域：Fire dynamics and safety research、Fire Detection and Safety Systems、Nuclear Engineering Thermal-Hydraulics

Industrial factory fires pose significant threats to structural safety, necessitating accurate temperature field reconstruction for emergency response. However, traditional methods are constrained by difficulties in obtaining key parameters in real time, reliance on dense sensor arrays, and limitations in generalization capability in complex fire scenarios. To address these challenges, this study proposes a hybrid algorithm combining particle swarm optimization with physics informed neural networks (PSO-PINN). The method embeds the heat balance equation as physical constraints into the neural network architecture, enabling real-time reconstruction of the full temperature field using only sparse sensor data. Furthermore, the PSO algorithm is incorporated to adaptively optimize hyperparameters—including batch size, learning rate, and activation function-thereby overcoming the sensitivity of PINN to manual parameter tuning. Validation based on two sets of model-scale fire experiments and one set of full-scale numerical simulations demonstrate that the PSO-PINN algorithm achieves significantly higher accuracy than comparative methods, with mean absolute errors as low as 0.426 °C, 0.429 °C, and 2.073 °C in the three tests. Compared to the traditional PINN and the fixed physical model-based PSO algorithm, the maximum reductions in prediction error reach 48.11 % and 39.49 %, respectively. This study provides a novel and practical technical framework for structural integrity assessme...