Development of a Fire Risk Prediction System Based on Deep Learning

Sensor-based safety monitoring systems play an important role in early detection and prevention of fire and explosion incidents at industrial pumping stations. Modern stations integrate multiple sensors such as temperature, humidity, gas, dust, air quality and fire sensors to assess operational status. In practice, continuous monitoring and coordinated analysis of multi-sensor data conducted by experienced professionals can significantly reduce the risk of fire and explosion. However, sustaining continuous expert-based monitoring is challenging due to high operational costs, manpower demands, and practical constraints. Consequently, automated monitoring and forecasting systems are required to deliver continuous and timely risk assessments while minimizing dependence on manual supervision. However, sensor signals often contain noise that is nonlinear and susceptible to environmental influences, making traditional threshold comparison methods unstable. This paper proposes a fire monitoring and forecasting system based on time series data and deep learning model with three status levels including safe, warning and dangerous. The models used and compared include decision trees, artificial neural networks, and long short-term memory networks with a twelve-step time window. Multi-sensor data are normalized and organized into time series windows to reduce noise and reflect fluctuations in real-world conditions. Experimental results on sensor data collected at the pumping station show that the long short-term memory network achieves higher accuracy and precision than the other two models. Contributing to improving the reliability of safety monitoring at the pumping station and creating a basis for practical implementation.