Bayesian approach to fault diagnosis with ambiguous historical modes

Abstract

Fault diagnosis plays an important role in diverse engineering areas including industrial control loop systems, mechanical systems, etc. Bayesian methods are a class of data-driven fault diagnosis methods developed in recent years. However, one difficulty with Bayesian methods is that they do not deal with the case that there is uncertainty about the underlying mode in the historical data. For this problem, a new approach is proposed in this paper, through which the ambiguous modes are softly classified by combing historical data, current evidence and the prior knowledge under a Bayesian framework. In addition, weighted kernel density estimation instead of classic histogram method is used for likelihood estimation to enhance diagnosis. The proposed Bayesian approach is tested on the fault diagnosis of Tennessee Eastman (TE) process using benchmark data and the proposed approach performs better in comparison with typical previous methods.