Fault Diagnosis of Industrial Systems with Bayesian Networks and Neural Networks

In this work we propose a two-phases fault diagnosis framework for industrial processes and systems; which combines Bayesian networks with neural networks. The first phase, based just on discrete observed symptoms, generates a set; of suspicious faulty process components. The second phase analyzes continuous data coming from sensors attached to components of this set and identifies the fault mode of each one. In first phase we use a discrete Bayesian network model, where probabilistic relationships among system's components are stated. In second phase, we analyze sensor measurements of suspicious faulty components with a. probabilistic neural network; previously trained with the eigenvalues of collected data. We show promising results from simulations performed with a 24 nodes power network.