Minimizing false-rejection rates in gas leak testing using an ensemble multiclass classifier for unbalanced data

Abstract

Leak testing is a fundamental stage in the manufacturing process of leak-tight products providing a non-destructive quality measurement. However, leak tests are often influenced by environmental and external factors that increase the risk of test inaccuracy. These phenomena are difficult to properly account for through mathematical models, as they are particular to each individual testing setup, nonetheless, they signify great expense to manufacturers due to substantial bottlenecking. In this paper, we introduce a real-world use-case at Bosch Thermotechnology facilities where over 23.98% of testing instances result in false-rejections. We then propose a data-driven, equipment agnostic, procedure for leak testing fault classification. We first identify seven relevant classes for fault diagnosis, and, due to the highly unbalanced nature of these classes (minority class represents only 0.27% of all data), we apply a novel unbalanced multiclass classification pipeline based on an ensemble of heterogeneous classifiers. Through this method, we are able to obtain F1-macro of 90%, representing a performance improvement of over 120% in comparison with conventional Auto-ML approaches.