Flare Soot Monitoring Based on Thermal Infrared Image Processing and Attention-based Meta-learning

Industrial enterprises, such as thermal power plants and petrochemical factories, use the flare stack to eliminate volatile organic compounds (VOCs) emerged in the production process. Through complete burning of the flare gas, the flare stack manages to prevent air pollution and protect the safety of the plants. The monitoring of the flare stack is of great importance for the reason that uncomplete combustion will endanger the environment and even people’s lives. In this study, an effective and intelligent flare soot monitoring system has been designed to deal with this situation. Unlike some existing work, the proposed monitoring system tries to identify combustion status of the flare gas through both RGB image and thermal infrared image (see Fig. 1). The new designed system can monitor the flare soot without concerning the light and weather conditions. Further more, we also constructed a deep neural model based on attention mechanisms and meta-learning to cooperate the problem of lacking sufficient black smoke samples. Experimental results showed that the devised monitoring system can recognize black smoke through videos of flare combustion timely and accurately regardless of the light and weather.