(375i) Detection of Faults in Multi-Modal Plant Operation Via Latent Space Models Based on First-Principles Plant Models

Equipment malfunction or instrument faults cause degradation of the performance of industrial plants. They may cause abnormal operating conditions, potentially threatening process safety and profitability. In order to prevent faults from becoming failures or malfunctions, process engineers have systematized procedures to monitor plant operation and correct the impact of faults as much as possible. Closed-loop control strategies can compensate for some process faults by ensuring that the plant produces on-spec products via adjustment of manipulated variables. Unfortunately, not all deviations from normal operating conditions can be corrected via control actions, leading to an unacceptable plant operation.

Closed-loop control actions mask the effects of process faults, making their detection more difficult. Effective fault detection methods must be able to detect the faults shortly after their occurrence and should produce a very low number of false positive alarms. In addition, transitions from one mode to another and operations under different operating modes should not be flagged as abnormal behavior.

Fault detection methods can be divided into methods based on plant data and those based on first-principles models. First-principles-based methods, such as the parity equations method [1], compare variables in the plant with those computed by the first principles model to detect faults.

Methods based on plant data became popular after MacGregor et al.'s pioneering work [2] on using principal component analysis (PCA) to detect process faults, leading to numerous variants of reduced dimensionality representations of the plant operation, such as dynamic PCA [3], kernel PCA [4], etc.

Previous methods rely on a model of normal operation to measure deviations from the normal operating conditions to determine the occurrence of a fault. Braatz et al. [5] adopted a different approach: they represented a plant through a Bayesian Recurrent Neural Network, which can be used to decide whether a (set of) measurement(s) represents a likely fault. If the process's faulty operation is suspected, they proceed to identify the fault.

Methods based on a model of a normal plant operation derived from plant data are effective over the range of available plant data. The advent of advanced control strategies, such as model predictive control (Dynamic Matrix Control [6], etc.), has created a situation where normal operating data cover a narrow region around each of the operating modes. This requires having as many data-driven models of normal operation as there are operating modes, as a method trained on data from a specific mode will generate erroneous signals when applied in a different operating mode. Additionally, these methods can produce false positive alarms during transitions between modes.

This work introduces a method that combines the strengths of first-principles-based and data-driven methods. Since first-principles models of many plants are widely available via current design or operation analysis practices (widespread use of Aspen Plus, Aspen HYSYS, Pro/II), these models can be used to generate data representing normal operation. Such data can be used to train reduced dimensionality models of the normal operation (e.g., PCA, autoencoders, etc.) and to detect deviations from normal operation via analysis of variables in the latent space. This data can then train reduced dimensionality models (such as PCA, autoencoders, etc.) to detect deviations from normal operating conditions by analyzing variables in the latent space. The sensitivity of the method with respect to mismatches between the plant and the model is analyzed, and model correction methods are presented. The behavior of PCA or autoencoders trained on one mode of operation during the transition to another mode and during the operation in that mode is examined.

The proposed framework is tested on the Tennessee Eastman Process on the different operating modes of the plant. Contrary to other approaches to this problem, a separate reduced space model was created for each main equipment of the plant (reactor, separator, stripper). The results suggest that incorporating the process topology into the fault detection framework is advantageous as it narrows down the search space during the fault root cause analysis (the equipment affected by the fault is revealed during the detection task). The novelty of the framework lies in (i) the parallel arrangement of the reduced dimensionality model to generate features in a reduced space that enables a clustering-based score to provide a reliable measure of the dissimilarity of the normal and abnormal operation and (ii) in a modular approach to building reduced dimensionality models. The framework is independent of the latent space modeling technique so that it can be used with a variety of reduced dimensionality models (PCA, KPCA, DPCA, autoencoders, etc.)

References:

[1] T. Höfling and R. Isermann, “Fault detection based on adaptive parity equations and single-parameter tracking,” Control Eng Pract, vol. 4, no. 10, 1996, doi: 10.1016/0967-0661(96)00146-3.

[2] J. V. Kresta, J. F. Macgregor, and T. E. Marlin, “Multivariate statistical monitoring of process operating performance,” Can J Chem Eng, vol. 69, no. 1, 1991, doi: 10.1002/cjce.5450690105.

[3] W. Ku, R. H. Storer, and C. Georgakis, “Disturbance detection and isolation by dynamic principal component analysis,” Chemometrics and Intelligent Laboratory Systems, vol. 30, no. 1, 1995, doi: 10.1016/0169-7439(95)00076-3.

[4] B. Schölkopf, A. Smola, and K. R. Müller, “Nonlinear Component Analysis as a Kernel Eigenvalue Problem,” Neural Comput, vol. 10, no. 5, 1998, doi: 10.1162/089976698300017467.

[5] W. Sun, A. R. C. Paiva, P. Xu, A. Sundaram, and R. D. Braatz, “Fault detection and identification using Bayesian recurrent neural networks,” Comput Chem Eng, vol. 141, 2020, doi: 10.1016/j.compchemeng.2020.106991.

[6] C. R. Cutler and B. L. Ramaker, “DYNAMIC MATRIX CONTROL - A COMPUTER CONTROL ALGORITHM.,” J Environ Sci Health B, vol. 1, 1980.