Abstract
A video based flare image monitoring system is developed for real-time estimation of the flare gas flow rate at the edge. Depending on the desired trade-off between speed and accuracy, either an object detection (EfficientDet Dx) or instance segmentation (Mask R-CNN) model is used for real-time detection of flare and smoke instances in the input video stream. Organic and synthetic data is used to achieve high precision and recall (greater than 0.98) for both flare and smoke. The detected rectangular bounding boxes or polygon masks are used to estimate the flame size, and predict the flare gas exit velocity or equivalently flow rate. The estimated flow rate is within +/- 10% of a reference flow meter. The Deep Learning models are "edgified" in order to shrink the size and improve the inference speed by similar to 3X on small footprint edge devices. The deep learning models of the flame size are combined with first principles knowledge to estimate the flare volumetric flow.