Embedded Spatial-Temporal Convolutional Neural Network Based on Scattered Light Signals for Fire and Interferential...

Area of Science:

• Fire safety engineering
• Artificial intelligence in sensor technology
• Aerosol science

Background:

• Photoelectric smoke detectors offer early fire warnings but suffer high false alarm rates due to varying smoke types and interferential aerosols.
• This limitation hinders their widespread adoption in residential settings, particularly in China.

Purpose of the Study:

• To develop an advanced model for accurate identification and classification of fire smoke and interferential aerosols.
• To enhance the reliability of photoelectric smoke detectors by minimizing false alarms.

Main Methods:

• An embedded spatial-temporal convolutional neural network (EST-CNN) was designed, featuring information fusion, scattering feature extraction, and aerosol classification modules.
• A novel two-dimensional spatial-temporal scattering (2D-TS) matrix was developed for fusing scattered light intensity data.
• The EST-CNN was trained and validated using experimental data from a custom-built fire test platform and a dual-wavelength, dual-angle photoelectric smoke detector.

Main Results:

• The EST-CNN achieved an average classification accuracy of 98.96% for various aerosols.
• The model boasts a compact size with only 67 kB of network parameters.
• Extensive experiments confirmed the optimal network parameters for high performance.

Conclusions:

• The developed EST-CNN model demonstrates high efficacy in distinguishing fire smoke from interferential aerosols.
• The model's efficiency and accuracy make it suitable for direct integration into existing commercial photoelectric smoke detectors.
• Implementing the EST-CNN can significantly improve the reliability and reduce false alarms in residential fire detection systems.