BiLSTM-LN-SA: A Novel Integrated Model with Self-Attention for Multi-Sensor Fire Detection.

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Multi-sensor fire detection is crucial but challenged by high false alarm rates.
• Existing methods struggle with complex environments and deep time-series data analysis.

Purpose of the Study:

• To develop a novel fire detection model, BiLSTM-LN-SA, for enhanced robustness and accuracy.
• To improve adaptability and generalization capabilities in diverse environmental scenarios.

Main Methods:

• Integrated Bidirectional Long Short-Term Memory (BiLSTM) for time-series feature extraction.
• Incorporated Layer Normalization (LN) to mitigate feature distribution shifts.
• Utilized Self-Attention (SA) mechanism for dynamic feature recalibration and fusion.

Main Results:

• Achieved 98.38% test accuracy, 0.98 F1-score, and 0.99 AUC.
• Demonstrated significantly lower false positive (1.50%) and false negative (1.85%) rates.
• Outperformed existing methods like EIF-LSTM, rTPNN, and MLP.

Conclusions:

• The BiLSTM-LN-SA model offers superior performance and reliability in multi-sensor fire detection.
• Layer Normalization and Self-Attention are key components for adaptability and feature fusion.
• The model shows strong generalization capability for practical applications in varied environments.