Towards automated gas leak detection through cluster analysis of mass spectrometer data.

• The data from the mass spectrometer was clustered by Euclidean distance to evaluate the presence or absence of vacuum leaks. • The leaked data and the no-leak data were classified into different clusters in the dendrogram, and this clustering was effective. • In this clustering, it is necessary to use data that is not affected by plasma discharge or fluctuation due to vacuum pumping over a long period. In order to generate high-performance plasma for future fusion power generation, it is desirable to keep high quality vacuum during experiment. Mass spectrometer is commonly used to monitor the vacuum quality and to record the amount of atoms and molecules in the vacuum vessel. Leak is the most serious accident to avoid that can nullify an experiment and even harm researchers. Detecting leaks are ever more important since it can be easily overlooked, e.g., when the deterioration in the vacuum degree is modest. This forces the researcher to carefully observe the vacuum and mass spectrometer data. This article presents a way to suggest potential leaks in the vacuum vessel by analyzing mass spectrometer data. This is done by utilizing the Euclidean distance between composition ratios at different times for the clustering using the daily composition ratio. We show that our cluster analysis is an effective way of separating these two cases, which results in a semi-automatic determination of leaks is more efficient than the current norm, which is to check many measures to find a small abnormality in the data manually. We plan further model improvements for long-term evaluation. [ABSTRACT FROM AUTHOR]