Causal structure learning in process engineering using Bayes Nets and soft interventions
As modern industrial processes become more and more complex, machine learning is increasingly used to gain additional knowledge about the process from production data. Up to now these methods are usually based on correlations between process parameters and hence cannot describe cause-effect relationships. To overcome this problem we propose a hybrid (constraint and scoring-based) structure learning method based on a Bayesian Network to detect the causal structure out of its data. Starting with an empty graph, first the underlying undirected graph is learned and edges coming from root nodes are directed using a constraint-based approach. Next a scoring-based method is used in order to calculate the uncertainty for each possible directed edge and out of this construct the Bayesian Network. Finally soft interventions are applied to the process in order to learn the real causal structure. Our approach is tested in simulations on a chemical stirred tank reactor and on an experimental laboratory plant.