Real-time indoor early fire detection in industrial environments using optimized multisensor edge deep learning

Abstract Early detection of fire hazards in indoor industrial environments is essential for ensuring personnel safety, reducing material losses, and maintaining operational continuity. Traditional fire detection systems primarily rely on post-ignition indicators such as smoke, temperature spikes, or visible flames, which often results in delayed alarms and limited preventive capability. To overcome these limitations, this paper proposes an end-to-end edge-based framework for proactive fire risk prediction. The system integrates multi-sensor environmental monitoring with a lightweight deep learning (DL) model deployed on a resource-constrained microcontroller. The proposed framework utilizes a public dataset containing environmental sensor measurements, which is systematically preprocessed to improve data quality and consistency. Feature selection is performed using Pearson correlation analysis and analysis of variance to identify the most informative sensor variables. To address class imbalance and improve model robustness, the synthetic minority over-sampling technique is applied during the training phase. A custom deep neural network is then trained and optimized using TensorFlow. To enable efficient deployment on embedded hardware, knowledge distillation is employed to transfer knowledge from a high-capacity teacher model to a lightweight student model suitable for edge execution. The resulting model achieves an accuracy of 93%, with a mean average precision of 0.96 and a recall of 0.83. Despite its strong predictive performance, the model maintains a very small memory footprint of 332.9 KB of Flash and 68.5 KB of RAM, enabling real-time inference on an ESP32 microcontroller. Prediction outputs are transmitted to a mobile web interface, providing continuous monitoring and real-time situational awareness for industrial safety. These findings highlight the potential of embedded edge intelligence for deploying proactive and resource-efficient fire risk prediction systems in industrial environments.