(523c) Bayesian Clustering Based Unsupervised Fault Detection and Diagnosis of Industrial Processes With Shifting Operating Conditions

For multimode processes, the conventional multivariate statistical processes monitoring (MSPM) techniques such as multiple principal component analysis (MPCA), multiple partial least squares (MPLS) and Gaussian mixture model (GMM) methods require fault-free data from all different operating conditions in order to train the data-driven models of normal process operations. However, the historical data collected from multimode industrial processes are often contaminated with faulty samples so that the above approaches may not well-suited.

In this study, a novel Bayesian clustering based fault detection and diagnosis approach is proposed to identify different operating conditions as well as detect and diagnose faults in complex multimode processes. An infinite Gaussian mixture model constructed upon Dirichlet process is constructed to characterize different normal and faulty operating conditions simultaneously. In contrast to the conventional clustering techniques that need the pre-identified number of clusters, the presented approach automatically determines the number of clusters based on process data patterns corresponding to different normal and faulty operating conditions.  In order to handle the intractable posterior density functions of various clusters, variational Bayesian method is employed to approximate the posterior distribution by minimizing the Kullback–Leibler divergence of a tractable proxy distribution to the true posterior distribution through an iterative optimization procedure. Then the monitored samples are attributed into different clusters based on the approximated posterior probabilities. Consequently, different operating conditions are identified and the faulty samples are also isolated. The detected faults are further diagnosed through a probability density based contribution index termed as free energy index so as to quantify the abnormality of each process variable.

The proposed approach is demonstrated using the Tennessee Eastman Chemical process along with its comparison to several traditional methods. The operating mode identification and fault detection performance of the presented method is evaluated against K-means, K-nearest neighbor (KNN) and fuzzy C-means (FCM) techniques while the fault diagnosis capability is examined by comparing with PLS based contribution decomposition and reconstruction methods. The operating mode identification, fault detection and diagnosis capability of the proposed Bayesian clustering approach is proven with accurate classification of normal and faulty samples as well as reliable diagnosis of faulty variables.

References

[1] Yu, J. and Qin, S.J. (2008). Multimode process monitoring with Bayesian inference-based finite Gaussian mixture models. AIChE J., 54, 1811-1829.

[2] Yu, J. (2012). A nonlinear kernel Gaussian mixture model based inferential monitoring approach for fault detection and diagnosis of chemical processes. Chem. Eng. Sci., 68, 506-519.