Bayesian classifiers applied to the Tennessee Eastman process

Area of Science:

• Artificial Intelligence
• Machine Learning
• Process Engineering

Background:

• Fault diagnosis is primarily a classification task.
• Bayesian networks (BNs) offer advantages for classification.
• Classifiers are often components within broader decision-making systems.

Purpose of the Study:

• To propose a Bayesian method, the dynamic Markov blanket classifier, for industrial process diagnosis.
• To induce accurate Bayesian classifiers with reliable probability estimates.
• To reveal relationships among the most relevant variables in industrial processes.

Main Methods:

• Introduction of the dynamic Markov blanket classifier for industrial diagnosis.
• Presentation of a new method, variable ordering multiple offspring sampling, for inducing Bayesian networks as classifiers.
• Assessment of performance using data from the Tennessee Eastman process benchmark.

Main Results:

• The proposed algorithms achieve good classification accuracies.
• The methods provide valuable insights into relevant variables for fault diagnosis.
• Performance was compared against naive Bayes and tree augmented network classifiers.

Conclusions:

• The dynamic Markov blanket classifier and variable ordering multiple offspring sampling are effective for industrial fault diagnosis.
• These Bayesian methods enhance classification performance and variable relevance identification.
• The study confirms the utility of these approaches for complex industrial systems.