Abstract
This work focuses on developing a smoke detection model that can be employed in an automated, end-to-end flare smoke detection, alerting and distributed control solution that leverages existing flare Closed Circuit Television (CCTV) cameras used at manufacturing facilities. At the core of this solution is a deep-learning computer vision model that is developed leveraging an extensive and diverse data set and the model has demonstrated fit-for-use performance in wide-ranging conditions. The smoke detection model uses a novel approach for real time detection of hydrocarbon emission leveraging semantic segmentation, through compact vision transformers. This innovative approach enables real-time identification of smoke in flare images, which is a key control signal for an efficient closed-loop system.
The training dataset includes a rich diversity of flare images (collected from global locations and environment conditions), including steam images with blank smoke masks to mitigate false positives. Additionally, images with minimal smoke were included to enhance the model's sensitivity to light smoke detection. External testing was conducted on previously unseen datasets, ensuring the model's robustness and generalizability. The developed model performs robust smoke detection with high precision alerting and minimizing false positives from steam images. This application has been deployed in the plant within an end-to-end solution, creating a positive impact on operations.
