Nonsteady Gas Leakage Detection Method with a Dataset Self-Feeding Segmentation Network

Effective gas leak detection holds significant economic, environmental, and safety importance. To solve the problems of low generalization ability of alkane gas leakage monitoring algorithm and the danger of on-site sampling of datasets, an off-site learning medium wave infrared imaging instrument was designed by analyzing the optical characteristics of alkane gas. The plume model is used for modeling, combined with the environmental background images taken in the field deployment, the gas leakage images in various scenes are fused and generated, and the dataset is created for fixed scene training to deal with various forms of unstable alkane gas, reduce the need for on-site gas leakage data collection and improve safety. Experimental and on-site results show that the F1 index of the gas segmentation algorithm is 0.78, the Intersection over Union (IoU) is 84.50%. The instrument has good environmental adaptability and stability, can operate stably in complex and changeable industrial environment, and effectively ensures the safety in production.