An Efficient IoT-based Fire Detection System Using Quantized Deep Learning Model on Resource-Constrained Devices

Fire hazards have been a source of huge losses in households, industries, and public places. The conventional fire safety system integrates embedded systems such as sensors and microcontrollers to alert users. Recent studies present image processing-based fire detection with high power-consuming microprocessors. As the system needs to operate continuously to monitor any possible fire hazard, this leads to excessive energy consumption and creates heating issues in system hardware. In this paper, we present an energy-efficient fire detection system combining embedded systems, IoT, and deep learning-based algorithms. Additionally, an enhanced deep learning algorithm utilizing CNN architecture is proposed, reducing model size and complexity for faster operation in resource-constrained devices. Our proposed model achieved a testing accuracy of 96.48%, a pruned model accuracy of 88% with a quantized model size of 1.8 MB. Additionally, the system demonstrated a model inference time of 1.9 mS while consuming an average power of 5.3 W.