Gas Leak Source Localization in Complex Chemical Plants using Deep Learning and Synthetic Data

Timely localization of the source point of a gas leak is crucial in preventing secondary and tertiary disasters caused by chemical gas leaks. Passive infrared imaging technology is widely utilized in the detection of hazardous gas leaks. However, previous deep learning-based detection algorithms have solely focused on recognizing the occurrence of a gas leak, neglecting the identification of the source point. Additionally, due to the challenge of acquiring hazardous gas leakage data in complex scenarios, previous models are only applicable to simple and undisturbed situations, failing to meet the requirements of daily safety maintenance in real chemical plants.
\nThe objective of this study is to develop a systematic approach for detecting gas leaks in complex chemical plant scenarios. Firstly, we propose a novel target detection model called VidFRCNN, which leverages the characteristics of the leaking gas by fully integrating motion and texture information. Our model outperforms other classical target detection models in the field of gas leakage detection. Building upon the original VidFRCNN model, we further enhance its performance by creating a lighter and faster classification model called VidFRCNN_class. Experimental results demonstrate the superiority of our classification model compared to other existing models.
\nFurthermore, we employ an adversarial generative network-based method to generate a virtual dataset with a complex background, referred to as ComplexGasVid. Comparative experiments demonstrate that the generated virtual dataset significantly improves the detection performance of our model in complex scenes.
\nThe code for our approach will be made publicly available on GitHub.