Probability Density Estimation and Bayesian Causal Analysis Based Fault Detection and Root Identification

Fault detection and diagnosis, as an important means to ensure industrial safety and profitability, has been given much attention. The traditional Bayesian network (BN) that is a typical graphical model has many applications in this area, but it has great limitations in the processing of continuous variables. Based on the BN model of system causal structure, this paper proposes the kernel density estimation (KDE) method to estimate the probability density function instead of the parameter learning of the traditional Bayesian network. In addition, the evaluation index of estimation quality as a test standard is strictly deduced for ensuring the accuracy of the model. Anomalous process behavior can be detected and diagnosed by examining the changes of the probability density. The improved method is more convenient than the traditional BN when dealing with the process data, since there is no need to do discretization or the Gaussian assumption. Industrial simulation experiments show that the proposed method can accurately detect system faults and trace back to the source of faults.