Using deep learning to improve the efficiency of wireless sensor networks

Abstract

Early fire detection is crucial for reducing human and material losses. One of the current methods for a real-time fire detection is using wireless sensor networks based on deep learning. The wireless sensor network can also broadcast real-time video of the fire site to the operation center. This enables security providers and firefighters to take quick and efficient measures to control fire and prevent its spreading out before it causes severe damage. Yet one of the problems of data transmission in wireless sensor networks is due to communication channels with insufficient capacity. This article is aimed at developing a method that uses deep learning to transmit video data through a wireless sensor network meant for the fire detection and data transfer to the fire-fighting operation center via an unreliable communication channel.  The proposed method is based on sending only important data from the whole video to the receiving side when fires are detected.  Each video frame is generated on the receiving side using reference data previously stored along with the transmitted information. Fire detection average probability was 90% when YOLOv5 deep learning model was used to detect fires and send information. Data transfer sometimes amounts to 3% of all information enclosed in the video frame further to simulation results. There were calculated values of the signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR) to check the proposed method efficiency in a low-capacity communication channel. The proposed method efficiently transmits only the important image-enclosed data and passably reconstructs the image in unreliable communication networks and it showed better results in data transmission than conventional methods.