Abstract: As one of the primary types of fires in China, electrical fires have long posed a serious threat to public safety and property. Traditional fixed-threshold alarm systems often suffer from poor adaptability in complex and variable electrical environments, leading to high rates of false alarms and missed detections, thereby limiting their early warning effectiveness. Against this problem, this study proposes a comprehensive early warning method for electrical fires that integrates a probabilistic alarm model. A dual-layer dynamic risk assessment framework is constructed by integrating Bayesian Networks with Long Short-Term Memory (LSTM) neural networks. This framework enables joint modeling and temporal analysis of key risk features such as line overheating, luminous connection, fault arc, etc. A set of early warning indicators is established along with corresponding data acquisition and processing strategies. On this basis, a comprehensive early warning platform is developed, incorporating data analysis, risk assessment, and warning notifications. Preliminary tests on small-sample datasets demonstrate the method’s ability to effectively distinguish between normal and fault conditions, showing high accuracy and promising applicability. This research offers a new pathway for developing intelligent electrical fire monitoring and early warning systems.

Key words: electrical fire, probability alarm, integrated early warning platform, cloud-edge-device collaboration, fire prevention