Deep-learning-based data loss reconstruction for spatiotemporal temperature in piloti structures: Enhancing applicability with limited datasets

A time-temperature curve, representing the fire characteristics of structures, can be obtained by real fire experiments. However, these experiments are inherently susceptible to data loss, which can compromise the accuracy of results. To address this challenge, this study proposed the framework utilizing a long-short-term memory (LSTM) with Bayesian optimization to reconstruct temperature histories by learning the spatiotemporal correlation of the data. The proposed framework is first validated using simulated datasets from computational fluid dynamics analyses. The field applicability of the model is further demonstrated through real fire test results, affirming its reliability in practical scenarios. The study also introduces a novel data processing technique to mitigate overfitting issues in LSTM applications with limited data, enhancing the robustness and reliability of temperature history reconstruction. Overall, the results highlight the potential of deep learning in accurately and practically reconstructing temperature histories in fire experiments.

Automated summary

This study proposes a framework using long-short-term memory (LSTM) with Bayesian optimization to reconstruct temperature histories in fire experiments by learning the spatiotemporal correlation of the data. The framework is validated using simulated datasets and real fire test results, demonstrating its reliability and practicality. A novel data processing technique is also introduced to mitigate overfitting issues, enhancing the robustness and reliability of temperature history reconstruction. Overall, the results highlight the potential of deep learning in accurately and practically reconstructing temperature histories in fire experiments.