Abstract
Root cause analysis is used to find the specific fault location and cause of a fault during system fault diagnosis. It is an important step in fault diagnosis. The root cause analysis method based on causality starts from the origin of the causal connection between transactions and infers the location and cause of the mechanism failure by analyzing the causal impact of variables between systems, which has methodological advantages. Causal analysis methods based on transfer entropy are proven to have biases in calculation results, so there is a phenomenon of calculating false causal relationships, which leads to the problem of insufficient accuracy in root cause analysis. Liang-Kleeman information flow (LKIF) is a kind of information entropy that can effectively carry out causal inference, which can avoid obtaining wrong causal relationships. We propose a root cause analysis method that combines graphical lasso and information flow. In view of the large amount of redundant information in industrial data due to the coupling effect of industrial systems, graphical lasso (Glasso) is a high-precision dimensionality reduction method suitable for large-scale and high-dimensional datasets. To ensure the timeliness of root cause analysis, graphical lasso uses dimensionality reduction of the data. Then, LKIF is used to calculate the information flow intensity of each relevant variable, infer the causal relationship between the variable pairs, and trace the root cause of the fault. On the Tennessee Eastman simulation platform, root cause analysis was performed on all faults, and two root cause analysis solutions, transfer entropy and information flow, were compared. Experimental results show that the LKIF-Glasso method can effectively detect the root cause of faults and display the propagation of faults throughout the process. It further shows that information flow has a better effect in root cause analysis than transfer entropy. And through the root cause analysis of the step failure of the stripper, the reason why information flow is superior to transfer entropy is explained in detail.