Abstract
Causal discovery for data-driven quality improvements has been attracting increasing attention in the manufacturing domain due to the more diverse data accumulated in the wake of digital transformation. However, manufacturing data often exhibit heteroscedastic noise, which hinders the estimation performance of many existing functional causal models that assume the independence of noise terms. This study introduces a continuous optimization-based estimation method that can handle heteroscedastic noise under multivariate non-linear data with no latent confounders.Numerical experiments on synthetic data show that our estimation method improves the estimation
of the causal structure under the heteroscedastic noise. We also report the result of the estimation method to real-world data collected from a ceramic substrate manufacturing process, and the results also prove the effectiveness of our approach.
