An improved Bayesian network method for fault diagnosis

Abstract

In modern industrial processes, the complexity and continuity of production plants usually carry many potential risks with complex relationship among process facilities. Once these potential risks occur, catastrophic disasters will bring serious harm on both human and environment. Identifying the root failure cause in advance can efficiently prevent the catastrophic disaster. The complexity and uncertain relationship among units, subsystems and operate parameters can cause the failure of many diagnosis methods. In this paper, an improved Bayesian Network (BN) is proposed for fault diagnosis with its ability to describe the uncertain knowledge and causal reasoning. The proposed method is divided into three steps: 1) Determine the network of BN by hybrid technique with process knowledge and data-driven correlation analysis; 2) Update BN parameters with Expectation Maximization (EM) algorithm; 3) Analyze the root failure cause based the occurrence probability of variables. The effectiveness of the proposed method is validated on the Tennessee Eastman Process (TEP).