Early fire detection is essential for preventing fatalities and large-scale economic and environmental damage.
Traditional systems like smoke alarms suffer from delayed detection, especially in complex environments.
This work proposes an attention-based architecture for early fire detection using multivariate time-series sensor data from distributed sensor nodes.
It leverages multi-headed attention as a way of capturing long-ranging dependencies in data, enabling more effective discrimination between fire and nuisance scenarios.
Experiments are conducted dataset containing multiple types of fire events carried out in an industrial hall.
After preprocessing the data, the attention model is trained and compared against a feed-forward neural network baseline.
Results show that the attention-based approach achieves superior performance across all evaluated metrics, with 99.6 % accuracy, higher precision (0.89), improved F1-score (0.937), and a significantly reduced false positive rate.
These findings demonstrate that attention mechanisms are highly effective for modeling multivariate sensor dynamics and provide a promising foundation for early fire detection systems of the next generation.

Early fire detection,Multivariate sensor data,Attention mechanisms,Time-series analysis,Deep learning,Residual networks,Recurrent Neural Network,Gradient Descent,Sensor fusion,anomaly detection,Distributed sensor network,Industrial safety systems,False al