Classification in Early Fire Detection Using Multi-Sensor Nodes-A Transfer Learning Approach

Area of Science:

• Engineering
• Computer Science
• Fire Safety

Background:

• Early fire detection is critical for protecting lives and property, especially in vulnerable historic buildings.
• Sparse real-world data for training fire detection models is a significant challenge due to infrequent fire events.

Purpose of the Study:

• To investigate the transferability of early fire detection models trained on small-scale data to a full-scale environment.
• To evaluate feature representation transfer and instance transfer techniques for multi-sensor fire detection.

Main Methods:

• Linear Discriminant Analysis (LDA) was used for feature space transformation on source domain data.
• TrAdaBoost algorithm was applied for instance transfer, adapting models with sparse target domain data.
• Classification performance was evaluated for four fire types across different sensor node positions.

Main Results:

• LDA achieved up to 69% classification rate and Cohen's Kappa of 0.58 in the full-scale room.
• TrAdaBoost improved average classification to 73% and Cohen's Kappa to 0.63 with targeted data boosting.
• Sensor nodes near walls showed lower classification performance; excessive boosting led to overfitting.

Conclusions:

• Feature and instance transfer learning are viable for early fire detection using multi-sensor data.
• Transfer learning can bridge the gap between limited training data and real-world application, enhancing fire safety systems.
• Careful application of instance transfer is necessary to avoid overfitting and maintain generalizability.