Knowledge-guided hierarchical deep learning for large-space fire temperature-field reconstruction

This paper proposes a hierarchical collaborative deep learning framework to address physical inconsistency and the reliability bottleneck induced by parameter uncertainty in large-space fire temperature field reconstruction. The framework integrates state identification, physical-parameter inference, and spatial reconstruction. First, a cascaded temporal feature extraction module using LSTM-FS identifies fire combustion stages. Then, a specialized LSTM-BiEC model autonomously predicts fire source coordinates and physical constants from discrete sensor time series, effectively mitigating epistemic uncertainty associated with manual parameter presetting. Finally, a physics-informed neural network based on the bi-exponential radial diffusion law (RCDBiE-PINN) reconstructs high-fidelity steady-state temperature fields. Compared to existing single-exponential decay or data-only models, this framework more accurately captures near-source extreme values and far-field decay characteristics. This study’s main contribution lies in its ability to automatically represent, infer, and utilize fire dynamics knowledge from sparse sensor data. By converting raw sensor streams into explicit fire-state and physical-parameter knowledge and using this knowledge to instantiate a physics-constrained reconstruction model, the framework reduces repeated manual state identification and parameter presetting during scene-level deployment and provides a machine-readable temperature-field interface for downstream safety assessment.