A self-supervised leak detection method for natural gas gathering pipelines considering unlabeled multi-class non-leak data

Detecting leaks in natural gas gathering pipelines is paramount for the safe and reliable operation of the gas and oil industry. Due to the lack of leak data and the changes in leak features, semi-supervised leak detection methods that use normal data for health model learning have attracted much attention. However, these approaches usually consider one-class normal samples as health data, which may fail to fit the reality of unlabeled multi-class non-leak data under variable operating conditions. In addition, existing semi-supervised methods often suffer from insufficient representation learning as they employ step-by-step training or rely on the low-level reconstruction of autoencoders. To address the above two key challenges, this paper proposes a novel end-to-end self-supervised leak detection method, self-supervised multi-sphere support vector data description. Specifically, it utilizes the presented multi-sphere support vector data description to model unlabeled multi-class non-leak data and the introduced self-supervised learning strategy to boost the representation learning of the end-to-end semi-supervised model. Moreover, the categories of unlabeled multi-class non-leak data are learned in an unsupervised way through alternating feature clustering and pseudo-label-based classification. A robust leak score calculation method is also designed to improve the performance of the proposed method. Finally, the experimental results on the field data collected from pipelines demonstrate the effectiveness of the proposed method.