Bayesian Fault Diagnosis Using Principal Component Analysis Approach with Continuous Evidence

For fault diagnosis problems where the historical data is from a number of monitors, conventional likelihood estimation approaches for Bayesian diagnosis are typically independent or lumped approach. However, for most chemical processes the monitor outputs are often not independent, but exhibit correlations to some extent; as for the lumped approach, it is infeasible due to the curse of dimensionality and the limited size of historical dataset. Also there is another limitation to the accuracy of the diagnosis that the continuous monitor readings are commonly discretized to discrete values, therefore information of the continuous data cannot be fully utilized. In this paper principal component analysis (PCA) approach is proposed to transform the evidence into independent pieces, and kernel density estimation is used to improve the diagnosis performance. The application to the Tennessee Eastman Challenge process using the benchmark data demonstrates the effectiveness of the proposed approach.