Enhancing deep learning based multi-parameter fire detection by training on simulated and combustion experimental dataset

To address the critical reliability concern of high false alarm and missed detection rates in single-parameter fire sensors, this study proposes a reliability-oriented multi-parameter fire detection method by training on both simulated and combustion experimental dataset. The proposed method incorporates domain adaptation and gate recurrent unit (DA-GRU) architecture, integrating two critical parameters—temperature and carbon monoxide (CO) concentration—to enhance detection reliability and reduce uncertainty in fire identification. To overcome the problem of insufficient training data for deep learning models, a high-quality, cross-domain, and multi-source dataset is constructed by combining numerical simulations with combustion experiments. Domain adaptation (DA) is employed to align the deep feature between simulated and experimental data, thereby enhancing the generalization ability and robustness of the model. Architecturally, the gate recurrent unit (GRU) is adopted to effectively capture the temporal dynamics of fire sequences, enabling accurate and efficient classification of complex fire scenarios. Results show that the proposed DA-GRU model achieves superior recognition performance with an accuracy of 96.32%, significantly outperforming baseline GRU models trained on mixed datasets without adaptation. Multiple ablation experiments and reliability analyses further validates the robustness, generalizability, and application potential of the proposed method in fire detection.