Sensor-Based Indoor Fire Forecasting Using Transformer Encoder.

Area of Science:

• Computer Science
• Artificial Intelligence
• Sensor Technology

Background:

• Indoor fires pose significant risks to life and property.
• Existing fire prediction systems often rely on traditional machine learning or recurrent neural networks.
• Advanced predictive modeling is crucial for mitigating fire-related damages.

Purpose of the Study:

• To propose a novel deep learning architecture for early fire detection.
• To evaluate the efficacy of Transformer encoders in analyzing multi-sensor time-series data for fire prediction.
• To compare the proposed model's performance against established methods.

Main Methods:

• Utilized a stack of Transformer encoders to process time-series sensor data.
• Input data consisted of sequential sensor values capturing environmental parameters.
• Model trained and validated on two distinct datasets representing varying complexity.

Main Results:

• Traditional machine learning models outperformed the Transformer model on a simple dataset.
• The proposed Transformer encoder model demonstrated superior performance on a complex dataset.
• This indicates a higher potential for the model in real-world applications with intricate patterns.

Conclusions:

• Transformer encoders offer a viable and potentially superior approach for complex fire prediction tasks.
• The model's effectiveness in complex scenarios suggests its applicability in advanced fire safety systems.
• Further research can explore optimizing the architecture for diverse environmental conditions.