(584c) Efficient Process Monitoring and Causality Analysis of Processes Via the Integrated Use of the Graphical Lasso and Markov Random Fields Modeling

Process monitoring is an important aspect of safe operation of process plants. Various methods exist which monitor the process using data-driven methods, but they all have certain limitations. Conventional methods, such as PCA and PLS, are effective in monitoring large systems by analyzing the reduced system space, but assumes the variables are linearly correlated, resulting in inaccurate results for faults occurring in small regions. Also, since the individual characteristics of the variables are mitigated, it is difficult to analyze and isolate the source of the fault. Recent methods have tried resolve these problems by developing new methods for process monitoring. An example of these methods are the use of variant auto-encoders and SVMs, since these methods can deal with nonlinear and non-Gaussian variables, and show more accurate monitoring performances. However, they are still limited in terms of fault isolation and propagation path analysis, due to the difficulty in causality analysis. The use of Bayesian networks is another exemplary effort to enable fault detection and isolation, and a few studies evaluate causality analysis performances of Bayesian network monitoring. But a critical limitation in using Bayesian networks is that the user should be readily aware of the causal relationship between variables, and that cyclic relationships cannot be modeled.

In this study, a novel monitoring method for accurately detecting the faults and analyzing the cause of the faults is proposed. Named the Glasso-MRF monitoring framework, this method integrates the use of the graphical lasso algorithm (G-lasso) and the Markov random field (MRF) modeling framework to divide the monitored variables into relevant groups and then detect the faults separately. Graphical lasso uses the lasso constraint on the coefficient of variables, making the inverse covariance matrix of variables to be of sparse form. The use of G-Lasso downsizes the process into groups that are highly correlated, enabling the fault propagation path to be identified rather simply, and relieving the computational complexity of the MRF-based monitoring. MRF modeling can extensively model the variable relationships including cyclic ones. Also, since the probability of MRF can be calculated using factor graphs, the individual contribution of each variable to the fault can be calculated, enabling effective fault isolation. The inference of MRFs are usually complex due to the existence of partition functions, but by down-sizing the system using G-lasso, this problem is resolved as well. The proposed method was applied to the well-known Tennessee Eastman process to test out its performance. Surprisingly, the detection rate was higher than any other state-of-the-art monitoring methods, including auto-encoders and Bayesian networks, showing more than 96% detection rates for all of the 21 faults encoded in the Tennessee Eastman process. Also, the fault propagation path and causality analysis results proved to be reliable, with respect to the related results in previous studies. These results prove that, the proposed methodology can effectively detect the fault as well as identify their causes, and show its propagation throughout the process, without any a priori knowledge of the process variables.